Decentralized federated learning (DFL) enables collaborative training through peer-to-peer (P2P) communication, removing reliance on a central coordinator and thereby improving fault tolerance and scalability. In practical DFL deployments, clients differ substantially in data distributions, computational resources, and network bandwidth. Consequently, enforcing a single model architecture across all clients is often infeasible or inefficient, motivating model-heterogeneous DFL. Prior heterogeneous approaches typically exchange distilled knowledge (e.g., logits, soft labels, or class prototypes) rather than model weights; however, under label-distribution skew they often improve performance primarily on locally observed (seen) classes and generalize poorly to locally missing (unseen) classes. We propose ProtoGenesis, a model-heterogeneous DFL framework that improves clients' recognition of unseen classes by combining (i) a semantic-preserving autoencoder trained on public data to support privacy-oriented sample reconstruction from compact latent embeddings and (ii) prototype-based regularization to stabilize representation learning. Clients proactively request lowdimensional embeddings and prototype statistics for selected classes from neighboring peers, reconstruct class-consistent samples locally, and augment training without exchanging raw data or full model parameters. We evaluate ProtoGenesis on image and text classification tasks using CIFAR-10, CIFAR-100, and DBpedia under heterogeneous model-mixing and non-IID settings. ProtoGenesis achieves comparable accuracy on seen classes while improving accuracy on unseen classes form near 0% to up to 70%. Privacy analysis shows that sensitive information cannot be reconstructed from shared embeddings. Moreover, ProtoGenesis reduces communication overhead by about 1500 × and demonstrates low inference and communication latency on a real-world Jetson-based DFL system.

Large Language Models (LLMs) have become foundational components in a wide range of applications, including natural language understanding and generation, embodied intelligence, and scientific discovery. As their computational requirements continue to grow, these models are increasingly deployed as cloud-based services, allowing users to access powerful LLMs via the Internet. However, this deployment model introduces a new class of threat: denial-of-service (DoS) attacks via unbounded reasoning, where adversaries craft specially designed inputs that cause the model to enter excessively long or infinite generation loops. These attacks can exhaust backend compute resources, degrading or denying service to legitimate users. To mitigate such risks, many LLM providers adopt a closed-source, black box setting to obscure model internals. In this paper, we propose ThinkTrap, a novel input-space optimization framework for DoS attacks against LLM services even in black-box environments. The core idea of ThinkTrap is to first map discrete tokens into a continuous embedding space, then undertake efficient black-box optimization in a low-dimensional subspace exploiting input sparsity. The goal of this optimization is to identify adversarial prompts that induce extended or non-terminating generation across several state-of-the-art LLMs, achieving DoS with minimal token overhead. We evaluate ThinkTrap across multiple commercial, closed-source LLM services and observe that it can consistently induce abnormally long outputs and noticeable response-side degradation under black-box access. To further quantify the system-level impact of the attack, we conduct controlled experiments on private LLM deployments, where ThinkTrap reduces service throughput to as low as 1% of its original capacity and, in extreme cases, induces complete service failure due to resource exhaustion.

With the increasing popularity of smart wearable devices, such as earbuds and smart watches, presents new challenges for seamless and secure user authentication due to their limited user interfaces. Conventional biometric methods, including voice, fingerprints, and facial recognition often face issues such as usability limitations, interference from noise, or vulnerability to spoofing attacks. This paper introduces a novel authentication system called BaroAuth, which utilizes the stable and unique Speech-aware Pressure Sequences (SPSs) patterns captured by a miniaturized MEMS barometer embedded in earbuds. The design of BaroAuth hinges on two important observations. First, the production of speech relies on the coordinated movements of articulatory organs, including the tongue, jaw, and soft palate. These organs, through the activity of the temporomandibular joint (TMJ), alter the shape of the ear canal, thereby causing subtle pressure changes that encode the speaker's unique physiological characteristics and articulatory dynamics. Second, SPSs show significant intra-individual consistency and considerable inter-individual variability, which barometers can effectively measure. Meanwhile, we develop the BaroAuth prototype and carry out comprehensive experiments based on it. The experimental findings reveal that BaroAuth demonstrates a mean false-acceptance rate (FAR) of 0.41% and a false-rejection rate (FRR) of 1.23%, respectively, even under complex attack scenarios.

Ensuring stable and high-quality real-time video analytics for computationally constrained mobile agents is essential. However, limited computing resources and network bandwidth present significant challenges in meeting the objective of low response time and high inference accuracy. In this paper, we present DiVE, an edge-assisted video analytics system that utilizes motion vectors calculated by video codec to extract foregrounds and differentially encode frames. DiVE removes rotational components from motion vectors by solving over-determined linear equations and filters noisy motion vectors based on the observation that motion vectors of static objects point to the same point when the ego agent purely translates. To distinguish foregrounds from backgrounds, DiVE estimates the ground based on observations that all foregrounds stand on the ground and motion vectors on static objects at the same height have the same normalized magnitude. DiVE then uses region-growing-based clustering to identify foreground objects. An adaptive bitrate allocation method is applied to optimize accuracy under estimated bandwidth. We implement a prototype and conduct extensive experiments to evaluate the performance of DiVE. The results demonstrate that DiVE can improve detection accuracy by up to 19.0% and reduce response time by up to 56.0% compared with other video analytics schemes in real-world traces.

Computer-vision-based assessment on waste sorting is desired to replace manpower supervision in Shanghai city. Due to the hardness of labeling a multitude of waste images, it is infeasible to train a semantic segmentation model for this purpose directly. In this work, we construct a new dataset consisting of 12, 208 waste images, upon which seed regions (i.e., patches) are annotated and classified into 21 categories in a crowdsourcing fashion. To obtain pixel-level labels to train an effective segmentation model, we propose a weakly-supervised waste image pseudo label generation scheme, called WISNet. Specifically, we train a cohesive feature extractor with contrastive prototype learning, incorporating an unsupervised classification pretext task to help the extractor focus on more discriminative regions even with the same category. Furthermore, we propose an effective iterative patch expansion method to generate accurate pixel-level pseudo labels. Given these generated pseudo labels, a few-shot segmentation model can be trained to segment waste images. We implement and deploy WISNet in real-world scenarios and conduct intensive experiments. Results show that WISNet can achieve a state-of-the-art 40.2% final segmentation mIoU on our waste benchmark, outperforming all other baselines and demonstrating its efficacy. The dataset and code will be publicly available at: https://github.com/shifan-Z/WISNet

Inertial measurement units (IMUs), have been prevalently used in a wide range of mobile perception applications such as activity recognition and user authentication, where a large amount of labelled data are normally required to train a satisfactory model. However, it is difficult to label micro-activities in massive IMU data due to the hardness of understanding raw IMU data and the lack of ground truth. In this paper, we propose a novel fine-grained user perception approach, called Saga, which only needs a small amount of labelled IMU data to achieve stunning user perception accuracy. The core idea of Saga is to first pre-train a backbone feature extraction model, utilizing the rich semantic information of different levels embedded in the massive unlabelled IMU data. Meanwhile, for a specific downstream user perception application, Bayesian Optimization is employed to determine the optimal weights for pre-training tasks involving different semantic levels. We implement Saga on five typical mobile phones and evaluate Saga on three typical tasks on three IMU datasets. Results show that when only using about 100 training samples per class, Saga can achieve over 90% accuracy of the full-fledged model trained on over ten thousands training samples with no additional system overhead.

Decentralized Federated Learning (DFL) removes the central server in traditional Centralized FL, eliminating performance and security bottlenecks while improving scalability for large-scale cross-device federated scenarios. In these scenarios, devices are heterogeneous in computation, storage, and data distribution, requiring personalized models. A common strategy for personalized DFL is for each device to collaborate with others having similar data distributions to train a federated model and then perform local pruning. However, in the absence of a central server, along with privacy concerns and the limited inference capabilities of low-end devices, challenges emerge in terms of communication, computation, and model convergence. This paper presents an efficient personalized DFL method, named PRISAM, based on BN (Batch Normalization) masks. Each device adaptively adjusts its BN mask, enabling effective structured pruning with minimal performance loss. By exchanging BN masks, communication overhead for similarity comparison is reduced by a factor of 100,000, comparing to exchanging an entire model, while computational costs are also significantly lowered. Extensive experiments show that PRISAM significantly improves personalized model performance on three datasets across two models, outperforming state-of-the-art methods, while greatly reducing computational and communication overhead.

To extend the receptive field of a mobile agent in complex scenarios, it is essential for multiple agents to cooperate with each other. However, it is challenging to achieve comprehensive 3D perception at the minimal computational and communication costs. In this paper, we propose CoPe, a lightweight and efficient collaborative 3D perception scheme for mobile agents. The main idea of CoPe is for an ego agent to query the most helpful information from its neighboring agents through a lightweight network attention mechanism. To this end, at each agent, we first leverage Singular Value Decomposition (SVD) to decompose a full-size point cloud feature into components. Meanwhile, with the novel self-attention and cross-attention algorithms, we respectively select the key component of an ego agent that best represent the point cloud of the ego agent as a query, and valuable components of each helper agent that are most relevant to the query as the answer. After feature reconstruction and aggregation, an ego agent can have a comprehensive understanding about the scene and make accurate predictions on downstream tasks. CoPe is lightweight and can be easily implemented on mobile devices. Results of extensive experiments conducted on both real-world and simulation datasets demonstrate that CoPe can achieve superior 3D object detection accuracy while significantly reducing the incurred computational and communication costs.

The growing adoption of smart wearable devices (e.g., earbuds and smart watches) poses new challenges for secure and seamless user authentication due to their limited interaction interfaces. Conventional biometric methods, including fingerprints, voice, and facial recognition, often suffer from usability constraints, noise interference, or susceptibility to spoofing attacks. In this paper, we propose BaroAuth, a novel authentication system that utilizes the stable and distinctive patterns of Speech-aware Pressure Sequences (SPSs) captured by miniature MEMS barometers embedded in earbuds. The design of BaroAuth is based on two key observations. First, speech production involves coordinated movements of articulatory organs, such as the jaw, tongue, and soft palate, which reshape the ear canal geometry via the temporomandibular joint (TMJ), generating subtle pressure variations that encode the speaker’s unique articulatory dynamics and physiological traits. Second, SPSs demonstrate strong intra-individual consistency and notable inter-individual variability, which can be effectively captured by barometers. We implement the prototype of BaroAuth and conduct comprehensive experiments on it. Experimental results demonstrate that BaroAuth achieves a mean false-acceptance rate (FAR) and false-rejection rate (FRR) of 1.62% and 1.74%, respectively, even under sophisticated attack scenarios.

The city perception problem which means to estimate the future information from historical data is of great importance for smart city applications. In this paper, by analyzing two metropolitan scale sensory datasets, we find that city perception data demonstrate strong spatial-temporal correlation. Based on this observation, we propose an innovative deep learning method, called DeepSeer, which can accurately estimate and predict signals at a metropolitan scale with short-term noisy sensory data. To this end, DeepSeer first reconstructs the original short-term data using multi-channel singular spectrum analysis (MSSA) to obtain sufficient training data. Second, because the metropolitan scale data would introduce an unacceptable computational complexity, DeepSeer only selects the sites that are spatially most relevant for prediction and uses a extended Treestructure Long Short-Term Memory (Tree-LSTM) to extract and fuse the features of different sites. We conduct extensive trace-driven simulations on both datasets. The average prediction accuracy of DeepSeer for both datasets can reach 90.75% and$\mathbf{9 1. 1 9 \%}$, respectively. The experiment results demonstrate the efficacy of the DeepSeer design.

Speech Emotion Recognition (SER), as a core technology in intelligent interaction and affective computing, faces critical privacy security challenges. Speech not only contains emotional information, but also encompasses sensitive data such as conversational content and speaker identity markers. The sensitive information could be potentially reconstructed through Automatic Speech Recognition (ASR), posing substantial privacy leakage risks. Existing privacy protection methods focus mainly on suppressing the leakage of demographic characteristics such as speaker identity, while there is a lack of systematic research on the speech content. The traditional SER and ASR joint training paradigm relies on the sequence alignment mechanism of ASR, which leads to deep coupling of emotion features with the corresponding text, causing a fundamental contradiction between “performance improvement” and “privacy protection”. To address this problem, this paper proposes an attentionbased cascaded learning enabling privacy-preserving (Aclipse) strategy, which is divided into two stages. In the first stage, the emotion-retaining (ER) module and semantics-obfuscating (SO) module in cascaded manner, which consist of the channel-level attention mechanism, guides the targeted injection of adversarial perturbation by identifying and strengthening high-frequency feature channels that are critical to ASR performance, thereby significantly improving the semantic obfuscation to achieve speech privacy protection while minimizing the negative impact on SER performance. In the second stage, we freeze the SO module parameters to maintain its interference effect on ASR, and fine-tune the ER to improve the performance of SER by adjusting the channel-level attention mechanism to focus on emotion-related information. The experimental results on the IEMOCAP dataset show that Aclipse achieves an effective balance between performance and privacy protection, providing a new idea to build a trustworthy SER system.

With the rapid advancement of generative forgery technologies, the detection of multi-modal deepfake audio-video content has become an urgent demand in cyber security and forensic analysis. However, detecting audio-video deepfakes remains challenging: forgery traces are often subtle, dispersed, and highly resolution-dependent; existing multimodal methods rely on simple concatenation or shallow interactions, leading to insufficient exploitation of cross-modal consistency. To address these issues, we propose a Cross-level Multi-modal Fusion (CLMF) framework that progressively integrates audio cues into visual representations through cross-level attention, adaptively enhancing complementary information while suppressing redundancy. In addition, we design an Adaptive Audio Feature Enhancement module (AAFE) to highlight subtle frequencydomain artifacts often masked by noise, and a Multi-scale Visual Feature Enhancement module (MVFE) to capture both local texture inconsistencies and global structural distortions. These components jointly achieve robust and consistent cross-modal alignment of forgery traces, leading to significant improvements in deepfake detection performance. On the FakeAVCeleb benchmark, AURORA achieves an accuracy (ACC) of 94.32% and an area under the curve (AUC) of 93.66%, demonstrating superior performance.

A wide range of user perception applications leverage inertial measurement unit (IMU) data for online prediction. However, restricted by the non-i.i.d. nature of IMU data collected from mobile devices, most systems work well only in a controlled setting (e.g., for a specific user in particular postures), limiting application scenarios. To achieve uncontrolled online prediction on mobile devices, referred to as the flexible user perception (FUP) problem, is attractive but hard. In this paper, we propose a novel scheme, called Prism, which can obtain high FUP accuracy on mobile devices. The core of Prism is to discover task-aware domains embedded in IMU dataset, and to train a domain-aware model on each identified domain. To this end, we design an expectation-maximization (EM) algorithm to estimate latent domains with respect to the specific downstream perception task. Finally, the best-fit model can be automatically selected for use by comparing the test sample and all identified domains in the feature space. We implement Prism on various mobile devices and conduct extensive experiments. Results demonstrate that Prism can achieve the best FUP performance with a low latency.

The cognitive impairment caused by cerebrovascular disease is one of the most common causes of dementia in the elders. Cerebrovascular diseases are traditionally diagnosed by neuroimaging techniques, e.g., MRI, which are however high-priced and can only identify the structural change of cerebral vessels. Questionnaire form-based diagnosis, e.g., Mini-Mental State Examination (MMSE) or Cognitive Assessment (MoCA), is commonly used to discover cognition impairment, however subjective, highly relying on the experience of doctors. Electroencephalogram (EEG) waveform changes with physiological conditions and it is closely related to cognitive, so it is promising to identify cerebrovascular disease. In this work, we propose an EEG-based diagnosis approach, Cog-EEG, to distinguish between healthy people, cerebrovascular patients with and without cognitive impairment, effectively, which contains the following key designs. First, after pre-processing of EEG signals (i.e., electrode positioning, re-reference, filtering, segmentation and independent component analysis), Cog-EEG exploits weighted minimum norm estimation to enable source localization with low-density (i.e., 19 channels) EEG signals, and then the recovered source signals are analyzed in time-frequency domain. Second, Cog-EEG conducts cross wavelet transform between brain regions, and extracts the corresponding functional connectivity matrices between them. Third, the functional connectivity matrices are utilized as key features and fed into a classifier for prediction. We collect an EEG signals dataset of 78 subjects, among whom 19 are healthy, 30 are cerebrovascular patients with cognitive impairment, and the rest are cerebrovascular patients without cognitive impairment. Experimental results demonstrate Cog-EEG can achieve a classification accuracy of 96.7%.

3D face recognition systems are frequently susceptible to spoofing attacks, with 3D face presentation attacks being particularly notorious. Attackers commonly exploit 3D scanning and printing techniques to generate masks of target individuals, a method proven successful in various real-world scenarios. A defining characteristic of these attacks involves acquiring 3D face models via 3D scanning, a process that is notably more expensive and cumbersome compared to obtaining 2D photos. In this work, we introduce DREAM, a novel method for recovering 3D face models from a single 2D image. Specifically, our approach adopts a black-box strategy, reconstructing sufficient depth information to compromise target recognition models—such as face identification and authentication systems—by merely accessing their output and the corresponding RGB photo. Our key insight is that achieving successful attacks doesn’t necessitate restoring the precise ground-truth depth values; instead, it only requires recovering the essential features that are salient to the target model’s decision-making process. We evaluate DREAM’s effectiveness using four public 3D face datasets. Experimental results indicate that DREAM achieves a 94% success rate on face authentication models, even in cross-dataset testing. For face identification models, the success rate is 36%. Building upon DREAM, we further propose DREAM-3D, which leverages a 3D GAN to reconstruct depth images for deceiving 3D face recognition systems. Additionally, we evaluate DREAM-3D’s effectiveness on two datastes. Experimental results indicate that DREAM-3D achieves attack success rates exceeding 90% and approximately 50% against different models.

Detecting 3D objects from a monocular camera in mobile applications, such as on a vehicle, drone, or robot, is a crucial but challenging task. The monocular vision's near-far disparity and the camera's constantly changing position make it difficult to achieve high accuracy, especially for distant objects. In this paper, we propose a new Mono3D framework named MoGDE, which takes inspiration from the observation that an object's depth can be inferred from the ground's depth underneath it. MoGDE estimates the corresponding ground depth of an image and utilizes this information to guide Mono3D. We use a pose detection network to estimate the camera's orientation and construct a feature map that represents pixel-level ground depth based on the 3D-to-2D perspective geometry. To further improve Mono3D with the estimated ground depth, we design an RGB-D feature fusion network based on transformer architecture. The long-range self-attention mechanism is utilized to identify ground-contacting points and pin the corresponding ground depth to the image feature map. We evaluate MoGDE on the KITTI dataset, and the results show that it significantly improves the accuracy and robustness of Mono3D for both near and far objects. MoGDE outperforms state-of-the-art methods and ranks first among the pure image-based methods on the KITTI 3D benchmark.

Model updates are exchanged between server(s) and participants in Federated Learning (FL), which can result in excessive delay, especially for large models. Existing communication-efficient FL approaches such as quantization and top-k sampling apply compression to gradients assuming that gradients are sparse and can tolerate small deviations. This can hardly be applied to down-link transmission. In this work, we employ compressive sensing on model parameters instead of gradients and propose a two-way adaptive compression scheme, Cepe-FL, which exploits dictionary learning to project non-sparse model parameters into sparse representations to ensure reconstruction accuracy. Cepe-FL supports joint model reconstruction with drastic reduction in computational complexity from$O(n)$to$O(1)$. Cepe-FL adjusts the compression ratio adaptively according to the training loss, achieving the best trade-off between communication and model precision. Furthermore, it demonstrates efficacy in defending against membership inference attacks since only compressed models are exchanged. We conduct extensive experiments on three image classification tasks and compare with three communication-efficient approaches including FedPAQ, FedAvg and T-FedAvg. Cepe-FL presents the best performance in all tasks under IID and non-IID scenarios. We also implement white-box membership inference attacks, and the results show Cepe-FL can significantly suppress success ratio of inference in comparison with other approaches.

Voice assistants, increasingly integrated into wearable devices with limited human-computer interaction modalities, are susceptible to voice spoofing attacks. Such attacks exploit pre-recorded or synthesized voice commands to trick the assistants into executing actions unauthorized by legitimate users. In this work, we propose GyroTalk, a novel approach extracts individual and reliable features from speech movement sequences of users, using built-in gyroscopes in wearables, to differentiate between legitimate users and malicious attackers. GyroTalk is inspired by two critical insights. First, speech, as a highly intricate motor task, necessitates the synchronized coordination of multiple respiratory, laryngeal, lingual and mandibular muscles. These collective muscle movements propagate throughout the body, providing unique movement signatures. Second, the distinctive speech movement sequences of individual speakers, essential for generating specific words, can be grabbed by embedded IMU of wearables. We conduct a comprehensive evaluation of GyroTalk across various COTS Android devices, including smart phones, watches and glasses. Our experimental results demonstrate that GyroTalk can achieve a mean FAR of 2.23% and a FRR of 2.48%, even in the face of complicated voice spoofing attacks.

Low earth orbit (LEO) constellation integrated maritime networks have recently attracted much interest due to the rapid development of maritime applications and services. LEO satellites have the advantages of wide coverage to provide seamless connection for maritime wireless devices. However, due to the limited battery and computing capacity of unmanned aerial vehicles (UAVs) for ocean information perception and processing, the computing-intensive and delay-sensitive oceanic data suffer from long latency and high energy consumption, which degrades the efficiency of maritime services. In this paper, to enhance the perception and offloading endurance of UAVs in maritime networks, we propose an energy efficient multi-access edge computing scheme for heterogeneous satellite-maritime networks, with the objective of minimizing the cumulative transmitted energy for UAVs. Specifically, we first present a heterogeneous satellite-maritime network framework in which LEO satellites and unmanned surface vehicles (USVs) equipped with edge servers can process workloads simultaneously. Next, considering the limited battery supply of UAVs, we propose a hybrid harvesting-and-offloading scheme for resource allocation, where UAVs first harvest energy from solar power and radio frequency power from USV, and then UAVs determine the offloading strategy for task processing. Moreover, a joint optimization problem is formulated to optimize the offloading decision, the time scheduling, and the transmitting power. We also exploit a vertical architecture to solve the formulated problem. Regarding each decomposed sub-problem, we propose efficient algorithms to derive the corresponding solutions. Finally, we provide numerical results to validate the performance of our proposed algorithms in comparison with several benchmark algorithms.

With the development of new generation communication technologies, the future maritime information networks pave the way to promote the exploration of ocean resources. Moreover, the underwater data center (UDC) is considered to be a significant data storage and computing unit in future maritime networks for providing ocean services. However, the current deployment of UDC faces the critical issues, i.e., the long-distance underwater transmission is unreliable and the energy consumption and resources of underwater transmission are overloaded. To address the two critical issues of unreliable data transmission and high resource overheads, in this paper, we present a hybrid semantic data transmission architecture in air-ocean integrated networks, which can perceive the sea surface data accurately and transmit it to the UDC for processing. Specifically, in surface layer, uncrewed aerial vehicles (UAVs) perceive ocean environment and send data to the buoy via non-orthogonal multiple-access (NOMA) transmission to improve the channel utilization. In underwater layer, the buoy sends the collected data to UDC via semantic transmission, while the semantic fidelity metric is utilized to improve the transmission efficiency. A resource allocation problem for energy minimization is formulated to jointly optimize the semantic scaling factor, the NOMA decoding order, the communication and computing resource allocations. We exploit a decomposition approach to transform the problem into two sub-problems, where the optimal resource allocations are obtained by proposing efficient algorithms. Finally, we provide simulations to verify the effectiveness and efficiency of our proposed scheme. The results demonstrate that our proposal has the advantages of lower energy consumption compared to several baseline schemes.

Emerging Artificial Intelligence of Things (AIoT) applications desire online prediction using deep neural network (DNN) models on mobile devices. However, due to the movement of devices,unfamiliartest samples constantly appear, significantly affecting the prediction accuracy of a pre-trained DNN. In addition, unstable network connection calls for local model inference. In this paper, we propose a light-weight scheme, calledAnole, to cope with the local DNN model inference on mobile devices. The core idea of Anole is to first establish an army of compact DNN models, and then adaptively select the model fitting the current test sample best for online inference. The key is to automatically identifymodel-friendlyscenes for training scene-specific DNN models. To this end, we design a weakly-supervised scene representation learning algorithm by combining both human heuristics and feature similarity in separating scenes. Moreover, we further train a model classifier to predict the best-fit scene-specific DNN model for each test sample. We implement Anole on different types of mobile devices and conduct extensive trace-driven and real-world experiments based on unmanned aerial vehicles (UAVs). The results demonstrate that Anole outwits the method of using a versatile large DNN in terms of prediction accuracy (4.5% higher), response time (33.1% faster) and power consumption (45.1% lower).

Registration of point clouds collected from a pair of distant vehicles provides a comprehensive and accurate 3D view of the driving scenario, which is vital for driving safety related applications, yet existing literature suffers from the expensive pose label acquisition and the deficiency to gen-eralize to new data distributions. In this paper, we propose EYOC, an unsupervised distant point cloud registration method that adapts to new point cloud distributions on the fly, requiring no global pose labels. The core idea of EYOC is to train a feature extractor in a progressive fashion, where in each round, the feature extractor, trained with near point cloud pairs, can label slightly farther point cloud pairs, enabling self-supervision on such far point cloud pairs. This process continues until the derived ex-tractor can be used to register distant point clouds. Par-ticularly, to enable high-fidelity correspondence label gen-eration, we devise an effective spatial filtering scheme to select the most representative correspondences to register a point cloud pair, and then utilize the aligned point clouds to discover more correct correspondences. Experiments show that EYOC can achieve comparable performance with state-of-the-art supervised methods at a lower training cost. Moreover, it outwits supervised methods regarding generalization performance on new data distributions.

The 6G empowered Internet of vehicles paves the way to autonomous driving era, where the ultra-low latency communication and ultra-reliable connections promote the quality of service (QoS) for vehicle users. However, the high data traffic load and communication resource constraint pose a heavy burden to autonomous driving. This paper proposes a secrecy oriented slicing resource allocation scheme in 6G green vehicular networks. We consider that cellular vehicular user (CUE) and vehicular user equipment (VUE) and eavesdropper coexist in the networks, where VUE can reuse the resource block non-orthogonally with CUE, and the eavesdropper may overhear the data transmission of CUE and VUE. To meet the QoS and green communication requirements, we formulate a joint optimization for energy-efficient resource allocation subject to the data rate and secrecy capacity. Despite the non-convex of the formulated problem, we propose corresponding algorithms to derive the optimal resource allocation strategies. Simulation performance validate the effectiveness of our proposal in comparison with benchmark algorithms.

Parameter-efficient fine-tuning (PEFT) on pre-trained models is a new paradigm for model training, where the majority parameters of a pre-trained model are frozen, left only a small number of unfrozen (or additional) parameters to be tuned. PEFT demonstrates its effectiveness in fitting the downstream tasks, while introducing a new surface for backdoor attacks. In this paper, we design a novel backdoor attack towards the Vision Transformer (ViT) PEFT, called Bad-Tuning. To mislead the target pre-trained model, Bad-Tuning first purposefully tailors the trigger for the frozen portion of the model, and then backdoors the unfrozen part by injecting the trigger into fine-tuning samples of PEFT. The main challenges are two-fold. First, backdooring PEFT should be highly efficient with a few backdoored samples. Second, the trigger should be hardly noticed by human beings as well as backdoor scanners. To deal with the challenges, Bad-Tuning optimizes the trigger by learning CLS sequences which represent rich deep semantics of samples, and introduces color loss to evaluate the invisibility of triggers. Extensive experiments on different datasets demonstrate the effectiveness, efficiency and invisibility of Bad-Tuning under both white-box and gray-box scenarios. Bad-Tuning achieves an average attack success rate (ASR) of over 99.9% even only 0.1% backdoored samples are injected. Moreover, Bad-Tuning outperforms the SOTA backdoor attacks on both ASR and invisibility (SSIM).

Federated learning (FL) collaboratively trains global models while preserving private data locally, making it an ideal privacy-preserving learning technique. However, recent studies have shown that FL poses risks of security attacks and privacy leaks during model parameter transfer. Existing research suggests that backdoor attacks cannot assist with membership inference attacks in machine learning. This paper proposes a joint attack of backdoor injection and membership inference in FL, M-Door, which can connect two independent work lines to ensure the security and privacy of FL. In M-Door, an attacker hidden within the client can not only perform backdoor attacks on the global model, but also perform membership inference attacks on the global model by analyzing the transmitted model parameters. This attack method can improve the success rate of backdoor attacks, and simultaneously increase the success rate of membership inference attacks. We conduct extensive experiments on three image classification tasks to evaluate the effectiveness of M-Door. Compared with the other two attack methods, the experimental results show that M-Door exhibits significant advantages in backdoor and membership inference attacks under both IID and Non-IID data settings.

Emerging Artificial Intelligence of Things (AIoT) applications desire online prediction using deep neural network (DNN) models on mobile devices. However, due to the movement of devices, unfamiliar test samples constantly appear, significantly affecting the prediction accuracy of a pre-trained DNN. In addition, unstable network connection calls for local model inference. In this paper, we propose a light-weight scheme, called Anole, to cope with the local DNN model inference on mobile devices. The core idea of Anole is to first establish an army of compact DNN models, and then adaptively select the model fitting the current test sample best for online inference. The key is to automatically identify model-friendly scenes for training scene-specific DNN models. To this end, we design a weakly-supervised scene representation learning algorithm by combining both human heuristics and feature similarity in separating scenes. Moreover, we further train a model classifier to predict the best-fit scene-specific DNN model for each test sample. We implement Anole on different types of mobile devices and conduct extensive trace-driven and real-world experiments based on unmanned aerial vehicles (UAV s). The results demonstrate that Anole outwits the method of using a versatile large DNN in terms of prediction accuracy (4.5 % higher), response time (33.1 % faster) and power consumption (45.1 % lower).

Federated learning is a mechanism that allows participating clients to train locally with their own data in order to receive rewards, thus avoiding the transfer of data to a central server and protecting users’ privacy. However, some “lazy” clients may adopt the strategy of fabricating false model local updates in an attempt to “free-riding” without actually contributing real data or consuming local computational resources. To address this issue, we propose FLoomChecker, an integrity detection scheme for federated learning training models. The scheme combines the techniques of trusted execution environments and Bloom filters to efficiently identify clients that do not train honestly by committing and proving. We conducted experimental evaluations of FLoomChecker, examining three main aspects: query time, build time, and memory footprint in trusted execution environment (TEE). The experimental results demonstrate the effectiveness of our scheme, and its performance improves as the number of local training rounds increases.

Multivariate time series anomaly detection is a longstanding but crucial technology, holding significant potential for system security and stability. Prior studies focus on designing sophisticated architectures, integrating advanced modules (e.g., CNN, LSTM, and Transformer), and identifying anomalies based on point-wise reconstruction errors, as it assumes anomalies cannot be correctly reconstructed. However, reconstruction-based methods are risk of over-generalization, making the assumption untenable. Moreover, the performance of complicated architectures is catastrophically overestimated by the flawed point-adjust protocol. In this paper, we propose a simple but efficient architecture comprising only feed-forward layers. The input time series is hierarchically transformed into multi-scale patches to discover complex temporal information. Triple-views are constructed to capture representation discrepancies among different views as the anomaly criterion, circumventing the overgeneralization issue of reconstruction-based methods. To further amplify the discrepancy, a loss function is designed to encourage the consistency among different views during the training phase. The proposed method is evaluated through quantitative and qualitative experiments, demonstrating its competitive performance on four benchmarks, and providing new baselines to the community without point-adjust protocol.

The quality of federated learning (FL) is highly correlated with the number and quality of the participants involved. It is essential to design proper contribution evaluation mechanisms. Shapley Value (SV)-based techniques have been widely used to provide fair contribution evaluation. Existing approaches, however, do not support dynamic participants (e.g., joining and departure) and incur significant computation costs, making them difficult to apply in practice. Worse, participants may be incorrectly valued as negative contribution under the Non-IID data scenarios, further jeopardizing fairness. In this work, we propose FairFed to address the above challenges. First, given that each iteration is of equal importance, FairFed treats FL as Multiple Single-stage Cooperative Games, and evaluates participants by each iteration for effectively coping with dynamic participants and ensuring fairness across iterations. Second, we introduce Cooperative Shapley Value (CSV) to rectify negative values of participants to improving the fairness while preserving true negative values. Third, we prove if participants are Strategically Equivalent, the number of participant combinations can be sharply reduced from exponential to polynomial, thus significantly reducing the computational complexity of CSV. Experimental results show that FairFed achieves up to 25.3 × speedup and reduces deviations by three orders of magnitude to two state-of-the-art approximation approaches.

Point cloud registration (PCR) can significantly extend the visual field and enhance the point density on distant objects, thereby improving driving safety. However, it is very challenging for vehicles to perform online registration between long-range point clouds. In this paper, we propose an online long-range PCR scheme in VANETs, called LoRaPCR, where vehicles achieve long-range registration through multi-hop short-range highly-accurate registrations. Given the NP-hardness of the problem, a heuristic algorithm is developed to determine best registration paths while leveraging the reuse of registration results to reduce computation costs. Moreover, we utilize an optimized dynamic programming algorithm to determine the transmission routes while minimizing the communication overhead. Results of extensive simulations demonstrate that LoRaPCR can achieve high PCR accuracy with low relative translation and rotation errors of 0.55 meters and 1.43°, respectively, at a distance of over 100 meters, and reduce the computation overhead by more than 50% compared to the state-of-the-art method.

Decentralized and open features of federated learning provides opportunities for malicious participants to inject stealthy trojan functionality into deep learning models collusively. A successful trojan attack is desired to be effective, precise and imperceptible, which generally requires priori knowledge such as aggregation rules, tight cooperation between attackers, e.g. sharing data distributions, and the use of inconspicuous triggers. However, in realistic, attackers are typically lack of the knowledge and hardly to fully cooperate (for privacy and efficiency reasons), and out of scope triggers are easy to be detected by scanners. We propose FedTrojan, a zero-knowledge federated trojan attack. Each attacker independently trains a quasi-trojaned local model with a self-select trigger. The model behaves normally on both regular and trojaned inputs. When local models are aggregated on the server side, the corresponding quasi-trojans will be assembled into a complete trojan which can be activated by the global trigger. We choose existing benign features rather than artificial patches as hidden local triggers to guarantee imperceptibility, and introduce catalytic features to eliminate the impact of local trojan triggers on behaviors of local/global models. Extensive experiments show that the performance of FedTrojan is significantly better than that of existing trojan attacks under both the classic FedAvg and Byzantine-robust aggregation rules.

Federated Learning (FL) enables multiple participants to collaboratively train a globally shared model without the need of explicit data sharing. However, prior research indicates that local model updates released during the federated training may also jeopardize privacy of participants. To address this issue, local differential privacy (LDP) mechanism has been applied to FL systems. LDP provides privacy protection with rigorous mathematical proof by introducing random perturbations, e.g., Gaussian noise, to the released updates, however excessive noise compromises the utility of the updates. In this paper, we propose a novel Correlation-aware Adaptive LDP mechanism, Fed-CAD, for FL, which reduces the required scale of noise by leveraging the temporal correlation between consecutive local model updates belonging to the same participant, without increasing the privacy budgets (risks). We theoretically prove that Fed-CAD satisfies (ε, δ)-LDP as long as the difference between local models is smaller than the differential bound, and analyze the noise variance, a metric of utility. We implement Fed-CAD on image classification FL tasks. Experimental results demonstrate that Fed-CAD significantly outperforms the one-shot LDP baseline.

Adhesive adversarial patches have been common used in attacks against the computer vision task of monocular depth estimation (MDE). Compared to physical patches permanently attached to target objects, optical projection patches show great flexibility and have gained wide research attention. However, applying digital patches for direct projection may lead to partial blurring or omission of details in the captured patches, attributed to high information density, surface depth discrepancies, and non-uniform pixel distribution. To address these challenges, in this work we introduce DepthCloak, an adversarial optical patch designed to interfere with the MDE of vehicles. To this end, we first simplify the patch to a gray pattern because the projected ''black-and-white light'' has strong robustness to ambient light. We propose a generative adversarial network (GAN) based approach to simulate projections and deduce a projectable list. Then, we employ neighborhood averaging to fill sparse depth values, compress all depth values into a reduced dynamic range via nonlinear mapping, and use these values to adjust the Gaussian blur radius as weight parameters, thereby simulating depth variation effects. Finally, by integrating Moiré pattern and applying style transfer techniques, we customize adversarial patches featuring regularly arranged characteristics. We deploy DepthCloak in real driving scenarios, and extensive experiments demonstrate that DepthCloak can achieve an attack success rate of over 80% in the physical world.

3D face recognition is subject to frequent spoofing attacks, in which 3D face presentation attack is one of the most notorious attacks. The attacker takes advantages of 3D scanning and printing techniques to generate masks of targets, which has found success in numerous real-life examples. The salient feature in such attacks is to obtain 3D face models through 3D scanning, though relatively more expensive and inconvenient when comparing with 2D photos. In this work, we propose a new method, DREAM, to recover 3D face models from single 2D image. Specifically, we adopt a black-box approach, which recovers 'sufficient' depths to defeat target recognition models (e.g., face identification and face authentication models) by accessing its output and the corresponding RGB photo. The key observation is that it is not necessary to restore the true value of depths, but only need to recover the essential features relevant to the target model. We used four public 3D face datasets to verify the effectiveness of DREAM. The experimental results show that DREAM can achieve a success rate of 94% on face authentication model, even in cross-dataset testing, and a success rate of 36% on face identification model.

Studies have proven that applying patch stickers generated through adversarial training to target objects can effectively deceive classifiers or target detectors. These ’Print-and-paste’ adversarial attacks however have three shortcomings. First, touching the target object physically is required, which may be infeasible in practice. Second, stickers might be taken as evidence to identify attackers. Third, the attack effect decreases significantly in poor light, especially at long distances. To overcome above limitations, we introduce OptiCloak, a car vanishing attack, which fools the Object Detector (OD) of a vision-based autonomous driving systems with transient projection pattern. We establish three digital-to-physical mapping models to compensate the distortions caused by perspective deformation, double image and partial light reflection in real-world. Furthermore, to avoid adversarial functionality degeneration caused by the loss of patch details in long-range attacks, we utilize MeanShift Filtering to constrain the ’resolution’ of pixels in a patch during training. We propose a gradient-free patch updating approach, which utilizes ZO-AdaMM to approximate gradients and model parameters through confidence scores of OD, making OptiCloak can work well in both white-box and black-box scenarios. We deploy OptiCloak in real-world driving scenarios, and the extensive experimental results demonstrate that OptiCloak achieves similar Attack Success Rates (ASRs) as printed patches in bright environments, while significantly improving the attack performance in gloomy environments. This effect is validated across all settings, including different angles, imaging devices, and film transparency rates. In black-box settings, the average ASR can reach 71%, with a maximum attack distance of approximately 10m.

Continuous authentication is viewed to be increasingly important for mobile devices, which store a wide range of private data and sensitive information of users. Traditional continuous authentication methods need user inputs (e.g. typing, sliding). In this work, we present MAUTH, a zero-effect continuous authentication scheme for mobile devices. With the built-in motion sensors on commercial off-the-shelf (COTS) devices, MAUTH can continuously extract, classify and verify the unique tremor features of users on how their body intrinsically shakes during the normal use of such devices. As a result, it is extremely difficult if not impossible to reproduce the same set of tremors as individuals differ in their muscle development. We implement MAUTH as a software on Android-based smartphones, which demonstrates that MAUTH is light-weight and unobtrusive to its users. We conduct extensive real-world experiments and trace-driven simulations in controlled and uncontrolled environments on 21 volunteers. The results show that MAUTH is difficult to counterfeit and achieves a low average false positive rate (FPR) of 6.73% under real-world spoofing attacks. Moreover, MAUTH is comfortable to use and can achieve a low average false negative rate (FNR) of 2.2% during uncontrolled and continuous usage of devices, leveraging isolation-forest-based classifiers trained with only 40 training samples.

Efficient one-hop multicasting (OHM) of high-volume sensor data plays a pivotal role in the success of cooperative autonomous driving applications. Although millimeter-Wave (mmWave) bands demonstrate huge potential for high- bandwidth OHM data transmission, the challenge lies in enabling individual vehicles to locate and communicate with suitable neighbors in a fully distributed and highly dynamic scenario. This paper introduces mmV2V, a fully distributed OHM scheme designed for vehicular networks, comprising three tightly integrated protocols. Initially, synchronized vehicles perform a probabilistic neighbor discovery procedure, wherein randomly divided transmitters (or receivers) clockwise scan (or listen to) the surroundings in synchronization with heterogeneous Tx (or Rx) beams. This approach facilitates the identification of the vast majority of neighbors within a few repeated rounds. Subsequently, vehicles engage in negotiations with their neighbors to establish an optimal communication schedule in evenly distributed slots. Finally, matched pairs of neighboring vehicles commence high data rate transmissions using refined beams. We implement a prototype testbed to validate the feasibility of the main components of mmV2V. Extensive simulations based on generated and real-world traffic traces are conducted and the results demonstrate that mmV2V consistently achieves a high completion ratio in demanding OHM tasks across various traffic conditions.

Point cloud registration (PCR) can significantly extend the visual field and enhance the point density on distant objects, thereby improving driving safety. However, it is very challenging for vehicles to perform online registration between long-range point clouds. In this paper, we propose an online long-range PCR scheme in VANETs, called LoRaPCR, where vehicles achieve long-range registration through multi-hop short-range highly-accurate registrations. Given the NP-hardness of the problem, a heuristic algorithm is developed to determine best registration paths while leveraging the reuse of registration results to reduce computation costs. Moreover, we utilize an optimized dynamic programming algorithm to determine the transmission routes while minimizing the communication overhead. To the best of our knowledge, LoRaPCR is the first solution to achieve multi-vehicle point cloud long-range registration. Results of extensive experiments demonstrate that LoRaPCR can achieve high PCR accuracy with low relative translation and rotation errors of 0.55 meters and 1.43${}^{\circ }$, respectively, at a distance of over 100 meters, and reduce the computation overhead by more than 50% compared to the state-of-the-art method.

Mobile monocular 3D object detection (Mono3D) (e.g., on a vehicle, a drone, or a robot) is an important yet challenging task. Existing transformer-based offline Mono3D models adopt grid-based vision tokens, which is suboptimal when using coarse tokens due to the limited available computational power. In this paper, we propose an online Mono3D framework, called MonoATT, which leverages a novel vision transformer with heterogeneous tokens of varying shapes and sizes to facilitate mobile Mono3D. The core idea of MonoATT is to adaptively assign finer tokens to areas of more significance before utilizing a transformer to enhance Mono3D. To this end, we first use prior knowledge to design a scoring network for selecting the most important areas of the image, and then propose a token clustering and merging network with an attention mechanism to gradually merge tokens around the selected areas in multiple stages. Finally, a pixel-level feature map is reconstructed from heterogeneous tokens before employing a SOTA Mono3D detector as the underlying detection core. Experiment results on the real-world KITTI dataset demonstrate that MonoATT can effectively improve the Mono3D accuracy for both near and far objects and guarantee low latency. MonoATT yields the best performance compared with the state-of-the-art methods by a large margin and is ranked number one on the KITTI 3D benchmark.

Physical adversarial attacks directly apply adversarial perturbations to real-world objects. Perturbations usually are printed as patches and pasted on target objects. This requires attackers in the vicinity of targets, which may not be feasible in practice. In this paper, we propose a stealthy physical adversarial attack by taking advantage of the transient of light projection. The attacker utilizes a drone with a portable projector to project the adversarial light pattern on the rear windshield of a vehicle to obstruct the object detector (OD) in autonomous driving systems. This can lead to serious safety vulnerability. We train digital perturbations by back propagation on the OD in an iterative manner. Unfortunately, they do not work well in the form of light patterns due to distortion, double imaging and partial reflectance when projected as light pattern. Hence, we model the mapping from digital to light projections, and use the inverse of the mapping to compensate the projection distortion in each iteration. We employ four state-of-the-art ODs to demonstrate the effectiveness and robustness of our proposed attack.

Registration of distant outdoor LiDAR point clouds is crucial to extending the 3D vision of collaborative autonomous vehicles, and yet is challenging due to small overlapping area and a huge disparity between observed point densities. In this paper, we propose Group-wise Contrastive Learning (GCL) scheme to extract density-invariant geometric features to register distant outdoor LiDAR point clouds. We mark through theoretical analysis and experiments that, contrastive positives should be independent and identically distributed (i.i.d.), in order to train density-invariant feature extractors. We propose upon the conclusion a simple yet effective training scheme to force the feature of multiple point clouds in the same spatial location (referred to as positive groups) to be similar, which naturally avoids the sampling bias introduced by a pair of point clouds to conform with the i.i.d. principle. The resulting fully-convolutional feature extractor is more powerful and density-invariant than state-of-the-art methods, improving the registration recall of distant scenarios on KITTI and nuScenes benchmarks by 40.9% and 26.9%, respectively. Code is available at https://github.com/liuQuan98/GCL.

Check-in data widely published in mobile social networks (MSNs) pose serious privacy threats to users. Existing inference attack methods show that pairwise social relationship could be estimated by analyzing check-in user records. However, the efficacy of such attacks heavily depends on the density of check-in data, which is often not present in practice. In this work, we propose a new inference attack scheme, which for the first time can effectively reveal hidden social friendship in both the real world and in cyberspace among users with only sparse check-in data. Our attack method enjoys two salient features. First, it requires no prior knowledge about social connections, instead it estimates users' social proximity by exploiting both physical presence and social proximities. Second, our attack scheme can automatically learn representative features based on the significance of various check-in records, rather than relying on heuristic features. We conduct extensive trace-driven simulations, and the results demonstrate that our inference attack method can improve the efficacy of the state-of-the-art learning-based schemes up to 40%. Moreover, our proposed attack method is also robust against common data obfuscation mechanisms.

For many driving safety applications, it is of great importance to accurately register LiDAR point clouds generated on distant moving vehicles. However, such point clouds have extremely different point density and sensor perspective on the same object, making registration on such point clouds very hard. In this paper, we propose a novel feature extraction framework, called APR, for online distant point cloud registration. Specifically, APR leverages an autoencoder design, where the autoencoder reconstructs a denser aggregated point cloud with several frames instead of the original single input point cloud. Our design forces the encoder to extract features with rich local geometry information based on one single input point cloud. Such features are then used for online distant point cloud registration. We conduct extensive experiments against state-of-the-art (SOTA) feature extractors on KITTI and nuScenes datasets. Results show that APR outperforms all other extractors by a large margin, increasing average registration recall of SOTA extractors by 7.1% on LoKITTI and 4.6% on LoNuScenes. Code is available at https://github.com/liuQuan98/APR.

Due to information explosion, conflicting data on the same object among multiple sources is ubiquitous on the Web. To solve those conflicts while estimating source reliability, truth discovery has become a hot topic. However, when considering multi-value objects, the inevitable unbalanced data distribution is overlooked by the existing approaches. In particular, only a few sources make lots of claims while most sources only provide a few claims, which renders the source reliability estimated for “small” sources totally random; Some objects are covered by plenty of sources while some objects are claimed by only a few sources, which causes the value correctness calculated for “cold” objects unreasonable. To tackle the unbalanced data where multi-value objects exist, we propose a confidence interval based approach (CIMTD). We estimate source reliability from two aspects, i.e., the ability to claim the correct number of value(s) and specific value(s) on an object. To reflect the real reliability for both “big” and “small” sources, confidence intervals of enriched estimation are considered. While estimating source reliability, uncertainty degrees are introduced to model object differences. Confidence intervals are also considered to reflect the real uncertainty for both “hot” and “cold” objects. Experimental results on two real-world datasets demonstrate the effectiveness of our approach.

Few-shot named entity recognition is extremely important for the domain lacking annotation data. Existing approaches ignore that class descriptions can provide additional information and rich prior knowledge for the model. Most datasets do not provide class description information, and the same entity class may have different definitions in different datasets. In this paper, we proposed a simple and effective method, i.e., AutoDes, that can extract class descriptions from annotated data automatically. AutoDes uses examples to construct class descriptions and takes the prediction words of entity-oriented prompts as candidate examples. Experiments with different few-shot settings on multiple datasets show that AutoDes is superior to the state-of-the-art methods in low-resource settings, improving F1 scores by 1.2% to 7.5% absolute points.

Math word problems (MWPs) solving has achieved promising results recently. However, most existing methods focus only on learning the mapping function between problem text and the target equation, ignoring the logical semantic similarity among problem texts with same target prototype equation and different topic description. Under the condition of maintaining the logical semantics, modifying only the topic words of a question text will let these models generate completely different equations and answers. In this paper, we propose a novel approach called Logical Semantic Aggregator (LSA) which solves the math word problem efficiently and effectively by extracting logical semantics from problem texts for giving a guidance to equation generation. In addition, a Implicit Constants Predictor (ICP) mechanism is used to predict the corresponding numerical label, which is used as the information prompts to improve the semantic representations of MWPs. Experimental results on the Math23K dataset revealed that our proposed methods can achieve better performance than baselines and higher equation accuracy with the help of implicit numerical label prompts.

Relation Extraction (RE) is a fundamental task in Natural Language Processing (NLP), which aims to extract relations between entity pairs mentioned in a given sentence. Recently, relation extraction has achieved remarkable process with the development of deep neural network. However, most of the recent works focus on using external knowledge to pretrain model or encoding entity related information with graph neural network. In this paper, we propose a BERT-based model with entity-level contrastive learning module based only on the original input sentences for relation extraction. By optimizing the contrastive learning and relation classification objectives jointly, our model is more effective in encoding relation representation between two entities in a given sentence and achieves better relation extraction performance. Experiments conducted on two mainstream RE datasets show that our model significantly outperforms the baselines' results on two benchmark datasets. Meanwhile, the results of our model are very competitive with the state-of-the-art models which employ extra training datas or information.

With the emergence of multidrug-resistant bacteria, antimicrobial peptides (AMPs) offer promising options for replacing traditional antibiotics to treat bacterial infections, but discovering and designing AMPs using traditional methods is a time-consuming and costly process. Deep learning has been applied to the de novo design of AMPs and address AMP classification with high efficiency. In this study, several natural language processing models were combined to design and identify AMPs, i.e. sequence generative adversarial nets, bidirectional encoder representations from transformers and multilayer perceptron. Then, six candidate AMPs were screened by AlphaFold2 structure prediction and molecular dynamic simulations. These peptides show low homology with known AMPs and belong to a novel class of AMPs. After initial bioactivity testing, one of the peptides, A-222, showed inhibition against gram-positive and gram-negative bacteria. The structural analysis of this novel peptide A-222 obtained by nuclear magnetic resonance confirmed the presence of an alpha-helix, which was consistent with the results predicted by AlphaFold2. We then performed a structure-activity relationship study to design a new series of peptide analogs and found that the activities of these analogs could be increased by 4-8-fold against Stenotrophomonas maltophilia WH 006 and Pseudomonas aeruginosa PAO1. Overall, deep learning shows great potential in accelerating the discovery of novel AMPs and holds promise as an important tool for developing novel AMPs.

For ligand binding prediction, it is crucial for molecular docking programs to integrate template-based modeling with a precise scoring function. Here, we proposed the CoDock-Ligand docking method that combines template-based modeling and the GNINA scoring function, a Convolutional Neural Network-based scoring function, for the ligand binding prediction in CASP15. Among the 21 targets, we obtained successful predictions in top 5 submissions for 14 targets and partially successful predictions for 4 targets. In particular, for the most complicated target, H1114, which contains 56 metal cofactors and small molecules, our docking method successfully predicted the binding of most ligands. Analysis of the failed systems showed that the predicted receptor protein presented conformational changes in the backbone and side chains of the binding site residues, which may cause large structural deviations in the ligand binding prediction. In summary, our hybrid docking scheme was efficiently adapted to the ligand binding prediction challenges in CASP15.

Accurate and continuous breathing rate detection is crucial as it can help people to assess their physical health and provide early warning and diagnosis for potential human diseases. Traditional breathing detection approaches involving intrusive devices are uncomfortable for long-term continuous monitoring. While contactless detection approaches utilizing radio-frequency (RF) signals or acoustic signals mainly focus on sensing the changes of chest and abdomen displacements, which are not a good indicator recording breathing event due to existing false body movements. In this article, we present Wi-Tracker, a contactless breathing detection system based on commercial off-the-shelf (COTS) smartphones, which detects breathing event through capturing the Doppler effect caused by human exhaled airflow on the reflected acoustic wave. Specifically, Wi-Tracker uses the speaker on smartphone to transmit ultrasound signals and its microphone to receive the reflected acoustic signals recording breathing event. Then, we adopt a cumulative power spectral density (CPSD) method to extract fine-grained breathing pattern from the received signals. Finally, we design algorithms to accurately capture the breathing event from the extracted breathing pattern. We evaluate Wi-Tracker with six volunteers for a period of one month. Experimental results show that Wi-Tracker is able to achieve contactless breathing detection with a mean estimation error (MEE) of 0.17 bpm, which is even better as compared to RFID-based or WiFi-based approaches.

Voice interaction, as an emerging human-computer interaction method, has gained great popularity, especially on smart devices. However, due to the open nature of voice signals, voice interaction may cause privacy leakage. In this paper, we propose a novel scheme, calledSeVI, to protect voice interaction from being deliberately or unintentionally eavesdropped. SeVI actively generates jamming noise of superior characteristics, while a user is performing voice interaction with his/her device, so that attackers cannot obtain the voice contents of the user. Meanwhile, the device leverages the prior knowledge of the generated noise to adaptively cancel received noise, even when the device usage environment is changing due to movement, so that the user voice interactions are unaffected. SeVI relies on only normal microphone and speakers and can be implemented as light-weight software. We have implemented SeVI on a commercial off-the-shelf (COTS) smartphone and conducted extensive real-world experiments. The results demonstrate that SeVI can defend both online eavesdropping attacks and offline digital signal processing (DSP) analysis attacks.

One-hop multicasting (OHM) of high-volume sensor data is essential for cooperative autonomous driving applications. While millimeter-Wave (mmWave) bands can be utilized for high-bandwidth OHM data transmission, it is very challenging for individual vehicles to find and communicate with a proper neighbor in a fully distributed and highly dynamic scenario. In this paper, we propose a fully distributed OHM scheme in vehicular networks, called mmV2V, which consists of three highly integrated protocols. Specifically, synchronized vehicles first conduct a probabilistic neighbor discovery procedure, in which randomly divided transmitters (or receivers) clockwise scan (or listen to) the surroundings in pace with heterogeneous Tx (or Rx) beams. In this way, the vast majority of neighbors can be identified in a few repeated rounds. Furthermore, vehicles negotiate with each of their neighbors about the optimal communication schedule in evenly distributed slots. Finally, each agreed pair of neighboring vehicles start high data rate transmissions with refined beams. We conduct extensive simulations and the results demonstrate that mmV2V can achieve a high completion ratio in rigid OHM tasks under various traffic conditions.

In Federated Learning (FL), two-way model exchanges are required between the server and the workers every training round. Due to the large size of machine learning models, communications between them lead to high training delay and economic cost. At present, communication-efficient FL methods, for examples, top-k sparsification and quantization, taking advantages of the sparseness of model gradients and the fact that gradient-based model updating can tolerance small deviations, effectively reduce the communication cost of single training round. However, these gradient-based communication-efficient schemes cannot be applied to downlink communication. In addition, they cannot be used in conjunction with those communicationfrequency-suppressed methods, e.g., FedAvg, which hinders them from further improving training efficiency. In this paper, we propose FedCS, a compressive sensing based FL method, which can effectively compress and accurately reconstruct non-sparse model (both local and global) parameters (iveights), and can reduce the overall communication cost up to 10 $\times$ as compared to FedAvg without decreasing test accuracy. We introduce 1) a dictionary learning scheme with a quasi-validation set, which helps to project non-sparse parameters onto a sparse domain; 2) ajoint reconstruction scheme, by using which the server recovers global model parameters by executing the reconstruction algorithm only once a round, regardless of the number of compressed local models; 3) a compression ratio adjustment strategy, which balances the trade-off between total communication cost and model accuracy. We perform FedCS on three image classification tasks, and compare it with FedAvg, FedPAQ and T-FedAvg (two improvements of FedAvg). Experimental results demonstrate that FedCS outperforms comparison methods in all tasks, and always maintains a comparable test accuracy to FedAvg, even using a small quasi-validation set and on Non-IId data.

Federated learning is a distributed machine learning paradigm that resolves the conflict between training requirements and client data privacy. There are some challenges for federated learning, such as data heterogeneity and communication load, which lead to the global model bias and slow convergence. In this work, we address the problem of data heterogeneity and communication load from a novel perspective, which is named FedSPARK. 1) We propose a new federated learning interaction training strategy, serial pipeline training (SPT). SPT changes the local training of a single client to serial training of multiple clients, which improves the performance of the global model on heterogeneous data. 2) We propose the global knowledge regularization (GKR) which is inspired by continuous learning. GKR can reduce the bias of the client local model by building global knowledge. Through theoretical analysis and experiments on multiple datasets, we show that our approaches greatly reduce the computation of clients and the amount of communication between clients and server, and improve the efficiency of federated learning compared to existing methods.

Monocular 3D object detection (Mono3D) in mobile settings (e.g., on a vehicle, a drone, or a robot) is an important yet challenging task. Due to the near-far disparity phenomenon of monocular vision and the ever-changing camera pose, it is hard to acquire high detection accuracy, especially for far objects. Inspired by the insight that the depth of an object can be well determined according to the depth of the ground where it stands, in this paper, we propose a novel Mono3D framework, called MoGDE, which constantly estimates the corresponding ground depth of an image and then utilizes the estimated ground depth information to guide Mono3D. To this end, we utilize a pose detection network to estimate the pose of the camera and then construct a feature map portraying pixel-level ground depth according to the 3D-to-2D perspective geometry. Moreover, to improve Mono3D with the estimated ground depth, we design an RGB-D feature fusion network based on the transformer structure, where the long-range self-attention mechanism is utilized to effectively identify ground-contacting points and pin the corresponding ground depth to the image feature map. We conduct extensive experiments on the real-world KITTI dataset. The results demonstrate that MoGDE can effectively improve the Mono3D accuracy and robustness for both near and far objects. MoGDE yields the best performance compared with the state-of-the-art methods by a large margin and is ranked number one on the KITTI 3D benchmark.

Voice assistants are popular to wearable devices with limited input and output capabilities, however vulnerable to voice attacks, which cheat a voice assistant by playing forged voice commands without user awareness. In this paper, we propose VOGUE, which captures unique yet stable pattern of speech movement sequences of speakers with embedded gyroscope in wearable devices, to distinguish between registered legal user and malicious attackers (human or machines). The design of VOGUE is based on two key observations. First, speech, as a type of highly complex motor task, inherently requires coordinated actions of many orofacial, laryngeal, pharyngeal, and respiratory muscles, and the collective movements of muscles propagate to distant body segments. Second, to generate a certain word, the speech movement sequence of a speaker is known to be distinctive, and can be captured by inertial sensors. We implement VOGUE on three kinds of COTS android devices including smart glasses, watches and phones, and conduct comprehensive evaluation on the performances. Experimental results show that VOGUE achieves a mean false-acceptance rate (FAR) and false- rejection rate (FRR) of 2.23% and 2.48%, respectively, even under sophisticated voice impersonation attacks.

MOTIVATION: The human major histocompatibility complex (MHC), also known as human leukocyte antigen (HLA), plays an important role in the adaptive immune system by presenting non-self-peptides to T cell receptors. The MHC region has been shown to be associated with a variety of diseases, including autoimmune diseases, organ transplantation and tumours. However, structural analytic tools of HLA are still sparse compared to the number of identified HLA alleles, which hinders the disclosure of its pathogenic mechanism. RESULT: To provide an integrative analysis of HLA, we first collected 1296 amino acid sequences, 256 protein data bank structures, 120 000 frequency data of HLA alleles in different populations, 73 000 publications and 39 000 disease-associated single nucleotide polymorphism sites, as well as 212 modelled HLA heterodimer structures. Then, we put forward two new strategies for building up a toolkit for transplantation and tumour immunotherapy, designing risk alignment pipeline and antigenic peptide prediction pipeline by integrating different resources and bioinformatic tools. By integrating 100 000 calculated HLA conformation difference and online tools, risk alignment pipeline provides users with the functions of structural alignment, sequence alignment, residue visualization and risk report generation of mismatched HLA molecules. For tumour antigen prediction, we first predicted 370 000 immunogenic peptides based on the affinity between peptides and MHC to generate the neoantigen catalogue for 11 common tumours. We then designed an antigenic peptide prediction pipeline to provide the functions of mutation prediction, peptide prediction, immunogenicity assessment and docking simulation. We also present a case study of hepatitis B virus mutations associated with liver cancer that demonstrates the high legitimacy of our antigenic peptide prediction process. HLA3D, including different HLA analytic tools and the prediction pipelines, is available at http://www.hla3d.cn/.

Acquiring the geographical distribution of neighbors can support more adaptive media access control (MAC) protocols and other safety applications in Vehicular ad hoc network (VANETs). However, it is very challenging for each vehicle to estimate its own neighbor distribution in a fully distributed setting. In this paper, we propose an online distributed neighbor distribution estimation scheme, called PeerProbe, in which vehicles collaborate with each other to probe their own neighborhood via simultaneous symbol-level wireless communication. An adaptive compressive sensing algorithm is developed to recover a neighbor distribution based on a small number of random probes with non-negligible noise. Moreover, the needed number of probes adapts to the sparseness of the distribution. We implement a prototype system to verify the feasibility of PeerProbe in various typical vehicular channel conditions. We further conduct extensive simulations and the results demonstrate that PeerProbe is lightweight and can accurately recover highly dynamic neighbor distributions in critical channel conditions.

Online semantic segmentation on a time series of point cloud frames is an essential task in autonomous driving. Existing models focus on single-frame segmentation, which cannot achieve satisfactory segmentation accuracy and offer unstably flicker among frames. In this paper, we propose a light-weight semantic segmentation framework for largescale point cloud series, called TempNet, which can improve both the accuracy and the stability of existing semantic segmentation models by combining a novel frame aggregation scheme. To be computational cost-efficient, feature extraction and aggregation are only conducted on a small portion of key frames via a temporal feature aggregation (TFA) network using an attentional pooling mechanism, and such enhanced features are propagated to the intermediate non-key frames. To avoid information loss from non-key frames, a partial feature update (PFU) network is designed to partially update the propagated features with the local features extracted on a non-key frame if a large disparity between the two is quickly assessed. As a result, consistent and information-rich features can be obtained for each frame. We implement TempNet on five state-of-the-art (SOTA) point cloud segmentation models and conduct extensive experiments on the SemanticKITTI dataset. Results demonstrate that TempNet outperforms SOTA competitors by wide margins with little extra computational cost.

Despite that deep neural networks have achieved satisfactory performance, they rely on powerful hardware for training, which is expensive and not easy to get. Therefore, the compression and acceleration of deep neural network is important. In addition, lightweight design of complex model usually leads to the decrease of model accuracy. In order to solve the problems, we propose a novel ensemble learning model called FA-GAL-ResNet, which is a lightweight residual network integrated focused attention mechanism and generative adversarial learning by knowledge distillation. In the training of knowledge distillation, the traditional single teacher network is replaced by the multi-teacher hybrid network, which enhances the learning span of student network. Meanwhile, the distilled network with attention scoring module focuses on the notable features related to the target task and the irrelevant features would be filtered by grade of scores and weights. Experimental results show that the proposed network structure improves the accuracy of image classification task, in comparison with the separate complex as well as lightweight neural networks.

Acquiring the geographical distribution of neighbors can support more adaptive media access control (MAC) protocols and other safety applications in Vehicular ad hoc network (VANETs). However, it is very challenging for each vehicle to estimate its own neighbor distribution in a fully distributed setting. In this paper, we propose an online distributed neighbor distribution estimation scheme, called PeerProbe, in which vehicles collaborate with each other to probe their own neighborhood via simultaneous symbol-level wireless communication. An adaptive compressive sensing algorithm is developed to recover a neighbor distribution based on a small number of random probes with non-negligible noise. Moreover, the needed number of probes adapts to the sparseness of the distribution. We conduct extensive simulations and the results demonstrate that PeerProbe is lightweight and can accurately recover highly dynamic neighbor distributions in critical channel conditions.

Barcodes and NFC have become the de facto standards in the field of automatic identification and data capture. These standards have been widely adopted for many applications, such as mobile payments, advertisements, social sharing, admission control, and so on. Recently, considerable demands require the integration of these two codes (barcode and NFC code) into a single tag for the functional complementation. To achieve the goal of "one tag, two codes" (OTTC), this work proposes CoilCode, which takes advantage of the printed electronics to fuse an NFC coil antenna into a QR code on a single layer. The proposed code could be identified by cameras and NFC readers. With the use of the conductive inks, QR code and NFC code have become an essential part of each other: the modules of the QR code facilitate the NFC chip in harvesting energy from the magnetic field, while the NFC antenna itself represents bits of the QR code. Compared to the prior dual-layer OTTC, CoilCode is more compact, cost-effective, flimsy, flexible, and environment-friendly, and also reduces the fabrication complexity considerably. We prototyped hundreds of CoilCodes and conducted comprehensive evaluations (across 4 models of NFC chips and 8 kinds of NFC readers under 13 different system configurations). CoilCode demonstrates high-quality identification results for QR code and NFC functions on a wide range of inputs and under different distortion effects.

Rhythmic movement disorder (RMD) is closely related to health problems like insomnia, daytime fatigue, anxiety disorder, and depression, or even causes severe injuries resulting from the movements. To obtain detailed information of RMD related abnormal movements for early diagnosis, there are generally three categories of solutions: 1) using camera to record image data; 2) wearing various smart devices; and 3) deploying dedicated hardware to capture sensor data. But none of such are widely accepted for different reasons due to privacy, inconvenience and excessive overhead. We believe one of the essential features in a feasible solution is nonintrusiveness, in which movement data collection should be carried out without the awareness of targets. In addition, it should be fairly accurate and low cost. In this work, we propose Wi-PSG, a contactless and nonintrusive sleep monitoring system, which exploits channel state information (CSI) from existing WiFi infrastructures to detect RMD related movements. Specifically, we introduce new set of sensitivity metrics and reconstruct the collected CSI into an ideal subcarrier sensitive to all target movements. With the estimated CSI background model derived from static propagation paths, nonmovement interference can be canceled from RMD movement detection.We then train the classifier for distinguishing different kinds of RMD movements using both time and frequency features extracted from CSI signals. We implement Wi-PSG with a pair of WiFi devices and wireless access point. We evaluate Wi-PSG with nine volunteers over a one-month period. The extensive experiments demonstrate that Wi-PSG can achieve a recognition accuracy of above 92%, even under challenging scenarios.

As mobile devices play increasingly important roles in our daily lives, it is of great significance to protect personal mobile devices from being lost. Noticing the trend that one person normally carries more than one mobile device, we propose an innovative scheme, calledCoSafe, to detect device loss by verifying the motion consistency between a pair of devices. The rationale is that the vibrations perceived on devices carried by the same person should be tightly coupled whereas a lost device would show distinct mobility characteristics from others. Specifically, CoSafe compares the mobility consistency between a pair of devices on three levels, where coarse features (i.e., the mobility state and motion periodicity) are first compared to give fast response and more complex comparison on subtle feature (i.e., the relative phase) is conducted only when needed. In this way, CoSafe can instantly respond and introduce very low computation and communication costs. We implement CoSafe on a Commercial-Off-The-Shelf Android smartphone and a smartwatch, and conduct both trace-driven simulations and real-world experiments to evaluate the performance of CoSafe. The results show that CoSafe achieves a mean false negative ratio and false positive ratio of 1.46 and 3.12 percent, respectively, even under sophisticated stealing attacks.

Finding a small object (e.g., earbuds, keys or a wallet) in an indoor environment (e.g., in a house or an office) can be frustrating. In this paper, we propose an innovative system, calledHyperEar, to localize such an object using only a single smartphone, based on enhanced time-difference-of-arrival (TDoA) measurements over acoustic signals issued from the object. One major challenge is the hardware limitations of a Commercial-Off-The-Shelf (COTS) phone with a short separation between the two microphones and the low sampling rate of such microphones. HyperEar enhances the accuracy of TDoA measurements by virtually increasing distances between microphones through sliding the phone in the air. HyperEar requires no communication for synchronization between the phone and the object and is a low-cost and easy-to-use system. We evaluate the performance of HyperEar via extensive experiments in various indoor conditions and the results demonstrate that, for an object of 7 m away, HyperEar can achieve a mean localization accuracy of about 15 cm when the object in normal indoor environments.

Voice interaction, as an emerging human-computer interaction method, has gained great popularity, especially on smart devices. However, due to the open nature of voice signals, voice interaction may cause privacy leakage. In this paper, we propose a novel scheme, called SeVI, to protect voice interaction from being deliberately or unintentionally eavesdropped. SeVI actively generates jamming noise of superior characteristics, while a user is performing voice interaction with his/her device, so that attackers cannot obtain the voice contents of the user. Mean-while, the device leverages the prior knowledge of the generated noise to adaptively cancel received noise, even when the device usage environment is changing due to movement, so that the user voice interactions are unaffected. SeVI relies on only normal microphone and speakers and can be implemented as light-weight software. We have implemented SeVI on a commercial off-the- shelf (COTS) smartphone and conducted extensive real-world experiments. The results demonstrate that SeVI can defend both online eavesdropping attacks and offline digital signal processing (DSP) analysis attacks.

MOTIVATION: The coronavirus disease 2019 (COVID-19) caused by a new type of coronavirus has been emerging from China and led to thousands of death globally since December 2019. Despite many groups have engaged in studying the newly emerged virus and searching for the treatment of COVID-19, the understanding of the COVID-19 target-ligand interactions represents a key challenge. Herein, we introduce COVID-19 Docking Server, a web server that predicts the binding modes between COVID-19 targets and the ligands including small molecules, peptides and antibodies. RESULTS: Structures of proteins involved in the virus life cycle were collected or constructed based on the homologs of coronavirus, and prepared ready for docking. The meta-platform provides a free and interactive tool for the prediction of COVID-19 target-ligand interactions and following drug discovery for COVID-19. AVAILABILITY AND IMPLEMENTATION: http://ncov.schanglab.org.cn. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

In mobile crowd sensing (MCS) applications, a public model of a system is expected to be derived from observations collected by mobile device users, through regression modeling. For example, a model describing the relationship between running speed, heart rate, height, and weight of runner can be constructed using MCS data collected from wristbands. Unique features of MCS data bring regression new challenges. First, observations are error-prone and private, making it of great difficulty to derive an accurate model without acquiring raw data. Second, observations are nonstationary and opportunistically, calling for an adaptive model updating mechanism. Last, mobile devices are resource-constrained, posing an urgent demand for lightweight regression. We propose an adaptive and blind regression scheme. The core idea is first to select an optimal `safe' subset of observations locally stored over all participants, such that the inconsistency between the subset and the corresponding regression model is minimized, and as many observations as possible are included. Then, based on the resulted regression model, more observations are checked and selected to refine the model. With observations constantly coming, newly selected `safe' observations are used to make the model updated adaptively. To preserve data privacy, one-time pad masking and blocking scheme are integrated.

Finding a small object (e.g., keys or a wallet) in an indoor environment (e.g., in a house or an office) can be frustrating. In this paper, we propose an innovative system, called HyperEar, to localize such an object using only one single smartphone, based on enhanced time-difference-of-arrival (TDoA) measurements over acoustic signals issued from the object. One major challenge is the hardware limitations of a Commercial-Off-The-Shelf (COTS) phone with a short separation between the two microphones and the low sampling rate of such microphones. HyperEar enhances the accuracy of TDoA measurements by virtually increasing distances between microphones through sliding the phone in the air. HyperEar requires no communication for synchronization between the phone and the object and is a low-cost and easy-to-use system. We evaluate the performance of HyperEar via extensive experiments in various indoor conditions and the results demonstrate that, for an object of 7m away, HyperEar can achieve a mean localization accuracy of about 15cm when the object in normal indoor environments.

Federated learning is inherently vulnerable to poisoning attacks, since no training samples will be released to and checked by trustworthy authority. Poisoning attacks are widely investigated in centralized learning paradigm, however distributed poisoning attacks, in which more than one attacker colludes with each other, and injects malicious training samples into local models of their own, may result in a greater catastrophe in federated learning intuitively. In this paper, through real implementation of a federated learning system and distributed poisoning attacks, we obtain several observations about the relations between the number of poisoned training samples, attackers, and attack success rate. Moreover, we propose a scheme, Sniper, to eliminate poisoned local models from malicious participants during training. Sniper identifies benign local models by solving a maximum clique problem, and suspected (poisoned) local models will be ignored during global model updating. Experimental results demonstrate the efficacy of Sniper. The attack success rates are reduced to around 2% even a third of participants are attackers.

Relative direction estimation among neighboring vehicles in urban environment is essential to a wide variety of driving safety applications. To obtain accurate direction information solely from vehicle-to-vehicle (V2V) communications is desirable but very challenging due to the highly dynamic vehicular environments. In this paper, we propose an online vehicular AoA estimation scheme, called DeepAoA, based on a deep learning method. More specifically, Channel state information (CSI) is estimated from a set of synchronized receiving radios by a receiver vehicle. By taking the CSI phase difference between a pair of such radios, CSI phase errors in baseband can be effectively eliminated, which makes CSI phase difference a compelling feature to represent the direction of incident radio frequency (RF) signals and the dynamic channel characteristics. A convolutional neural network (CNN) model is then trained with labeled samples of CSI phase difference. We implement a prototype of DeepAoA receiver using four synchronized USRPs with their antennas in uniform circular array (UCA) configuration for full field of view. We collect real-world CSI trace and conduct trace-driven simulations. DeepAoA can achieve AoA estimation errors of less than 3 degrees with a 98% confidence interval with four antennas. The results demonstrate the efficacy of DeepAoA.

Smartphones with dual cameras are increasingly popular due to the need of supporting 3D vision. The depth information is critical for 3D vision. However, the two cameras on a smartphone are too close to accurately estimate the depth information especially for objects beyond two meters. In this paper, we propose an innovative system, called HyperSight, to estimate the depth information of objects using a dual camera smartphone. HyperSight realizes a virtual longbaseline stereo vision rig by having a user to move the phone in the air. The phone movement is continuously tracked and estimated using the short-baseline dual camera seeing nearby objects. We implement HyperSight as software on a Commercial-Off-The-Shelf (COTS) smartphone and conduct real-world experiments. The results show that when measuring feature-rich objects at a distance of five meters, HyperSight achieves a mean depth error of 6cm, which is up to 10× and 18× improvement in the accuracy compared with the stereo vision system using the native dual cameras and the Measure app based on ARKit 1 on mobile devices, respectively.

Mobile crowd sensing (MCS) data have unique features, i.e., private, error-prone, non-stationary and opportunistically generated, and collected by resource-constrained mobile devices, which bring the regression task new challenges. First, it is of great difficulty to derive an accurate model without acquiring raw data. Second, adaptive model updating mechanism is urgently needed. Last, regression schemes should be lightweight. In this paper, we propose a blind regression scheme, called Sieve, in MCS settings. The core idea is first to let the server help in coordinating the selection of a small `clean' subset of observations locally stored over all volunteers. Based on the `clean' subset, an initial model can be established among volunteers in a distributed fashion. With observations constantly coming, instead of model re-estimating from the scratch, newly selected `clean' observations are used to update the established model. Sieve preserves data privacy by only exchanging aggregated information. With the incremental model updating strategy, it also minimizes the communication and computation overhead of mobile devices. Extensive trace-driven simulations are conducted and the results demonstrate the efficacy of the Sieve design.

In mobile crowd sensing (MCS) applications, a public model of a system or phenomenon is expected to be derived from sensory data, i.e., observations, collected by mobile device users, through regression modeling. Unique features of MCS data bring the regression task new challenges. First, observations are error-prone and private, making it of great difficulty to derive an accurate model without acquiring raw data. Second, observations are non-stationary and opportunistically generated, calling for an evolutionary model updating mechanism. Last, mobile devices are resource-constrained, posing an urgent demand for lightweight regression schemes. In this paper, we propose an evolutionary blind regression scheme, called Lotus, in MCS settings. The core idea is first to select a 'maximum- safe-subset' of observations locally stored over all participants, which refers to finding a subset containing half of observations, such that the corresponding regression model has a minimum value of residual sum of squares. It implies the inconsistency between observations in the subset is minimized. Since such a maximum-safe- subset selection problem is NP-hard, a distributed greedy hill- climbing algorithm is proposed. Then, based on the resulted regression model, more observations are checked. Selected ones will be used to refine the model. With observations constantly coming, newly selected 'safe' observations are used to make the model evolved. To preserve data privacy, a one-time pad masking mechanism, and a blocking scheme are integrated into the process of regression estimation. Intensive theoretical analysis and extensive trace driven simulations are conducted and the results demonstrate the efficacy of the Lotus design.

WiFi technology has gained a wide prevalence for not only wireless communication but also pervasive sensing. A wide variety of emerging applications leverage accurate measurements of the Channel State Information (CSI) information obtained from commodity WiFi devices. Due to hardware imperfection of commodity WiFi devices, the frequency response of internal signal processing circuit is mixed with the real channel frequency response in passband, which makes deriving accurate channel frequency response from CSI measurements a challenging task. In this paper, we identify non-negligible non-linear CSI phase errors and report that IQ imbalance is the root source of non-linear CSI phase errors. We conduct intensive analysis on the characteristics of such non-linear errors and find that such errors are prevalent among various WiFi devices. Furthermore, they are rather stable along time and the received signal strength indication (RSSI) but sensitive to frequency bands used between a transmission pair. Based on these key observations, we propose new methods to compensate both non-linear and linear CSI phase errors. We demonstrate the efficacy of the proposed methods by applying them in CSI splicing and indoor distance ranging. Results of extensive real-world experiments indicate that accurate CSI phase measurements can significantly improve the performance of splicing and the stability of the derived power delay profiles (PDPs). Moreover, the estimated distance errors are reduced by 5.7 times on average comparing to the state-of-the-art schemes.

WiFi technology has gained a wide prevalence for not only wireless communication but also pervasive sensing. A wide variety of emerging applications leverage accurate measurements of the Channel State Information (CSI) information obtained from commodity WiFi devices. Due to hardware imperfection of commodity WiFi devices, the frequency response of internal signal processing circuit is mixed with the real channel frequency response in passband, which makes deriving accurate channel frequency response from CSI measurements a challenging task. In this paper, we identify non-negligible non-linear CSI phase errors and report that IQ imbalance is the root source of non-linear CSI phase errors. We conduct intensive analysis on the characteristics of such non-linear errors and find that such errors are prevalent among various WiFi devices. Furthermore, they are rather stable along time and the received signal strength indication (RSSI) but sensitive to frequency bands used between a transmission pair. Based on these key observations, we propose new calibration methods to compensate both non-linear and linear CSI phase errors. We demonstrate the efficacy of the proposed methods by applying them in CSI splicing. Results of extensive real-world experiments indicate that accurate CSI phase measurements can significantly improve the performance of splicing and the stability of the derived power delay profiles (PDPs).

IEEE 802.11p based Dedicated Short Range Communication (DSRC) has been considered as a promising wireless technology for enhancing road safety and efficiency. However, there is lack of deep understanding about how IEEE 802.11p performs for vehicle-to- vehicle (V2V) communications in urban environments. In this demo paper, we first introduce the statistical analysis results on V2V communication performance, based on a large volume of real-world measurement trace collected in Shanghai city. With the key insights observed in our experiments, we propose a novel scheme to synthesize V2V communication traces which can be of great value for VANET-related research, such as network protocol design and simulations.

Smartphones have been widely used with a vast array of sensitive and private information stored on these devices. To secure such information from being leaked, user authentication schemes are necessary. Current password/pattern-based user authentication schemes are vulnerable to shoulder surfing attacks and smudge attacks. In contrast, stroke/gait-based schemes are secure but inconvenient for users to input. In this paper, we propose ShakeIn, a handy user authentication scheme for secure unlocking of a smartphone by simply shaking the phone. With embedded motion sensors, ShakeIn can effectively capture the unique and reliable biometrical features of users about howthey shake. In this way, even if an attacker sees a user shaking his/her phone, the attacker can hardly reproduce the same behavior. Furthermore, by allowing users to customize the way they shake the phone, ShakeIn endows users with the maximum operation flexibility. We implement ShakeIn and conduct both intensive trace-driven simulations and real experiments on 20 volunteers with about 530,555 shaking samples collected over multiple months. The results show that ShakeIn achieves an average equal error rate of 1.2 percent with a small number of shakes using only 35 training samples even in the presence of shoulder-surfing attacks.

With more and more trajectory traces available, conducting analysis and mining on those trajectories can obtain valuable information. Although the published traces are often made anonymous by substituting the true identities of mobile nodes with random identifiers, the privacy concern remains. In this paper, we propose a new de-anonymization attack based on the movement pattern of moving objects. Since moving objects are open to observe in public spaces, an attacker can easily learn information about a victim's movement either through direct observations or from third parties. After collecting a few trajectory segments of a mobile object, the movement pattern of the victim can be extracted, using an improved TF-IDF method. By comparing the movement pattern of the victim with those extracted from historical anonymous traces, it is possible to identify the victim from the anonymous traces. We conduct extensive trace-driven simulations and the results demonstrate that the attacker is able to de-anonymize anonymous trajectories with high probability.

Acquiring accurate location information of vehicles is of great importance. Global Positioning System (GPS) has been widely deployed and used to be the most convenient solution to outdoor localization. As more and more infrastructure such as elevated roads, tunnels and tall buildings is built, however, the ever-increasing complexity of urban environments makes vehicle localization especially in those urban canyons a new challenging problem. In this paper, we propose a novel scheme, called UPS, to tackle urban vehicle localization problem. Inspired by the observation from empirical study that the Received Signal Strength Indication (RSSI) values of cellular signals (e.g., GSM) perceived over a distance have ideal temporal-spatial characteristics for fingerprinting, UPS refines the location accuracy of a moving vehicle by matching its cellular-aware trajectory, which is an association between consecutive geographical positions and the corresponding wide-band GSM RSSI values, with a pre-constructed map. Moreover, UPS leverages large mobility of vehicles to construct large-scale maps. We implement a prototype system to validate the feasibility of the UPS design. We conduct extensive real-world experiments and results show that UPS can work stably in various urban settings and achieve an accuracy of 4.2 meters on average and 5.3 meters with a 90% precision.

Access to accurate relative front-rear distance information between vehicles can be of great interest to drivers as such information can be utilised to improve driving safety. Acquiring such information in urban settings is very challenging due to the high complexity of urban environments. In this paper, we propose a novel scheme, called RUPS, to tackle the relative distance fixing problem. We first investigate pervasive GSM signals and find that the received signal strength indicator (RSSI) of multiple GSM channels measured over a distance has ideal temporal-spatial characteristics for temporary fingerprinting. With this observation, an RUPS-enabled vehicle first perceives the information of its GSM-aware trajectory while moving. Then by exchanging and comparing its own trajectory with that of a neighbouring vehicle, the vehicle can identify common locations overlapped on both trajectories. Finally, the relative distance between this pair of vehicles can be obtained by further comparing their geographical trajectories since that common location. As a result, RUPS is a fully distributed and lightweight scheme, requiring only a minimum hardware deployment, and does not need synchronization between vehicles or any pre-constructed signal maps. Extensive trace-driven simulation results show that RUPS can work stably under complex urban environments and overwhelm the performance of GPS by 2.7 times on average.

Since the heterogeneous Internet Service Provider (ISP) router policies prevent the quality-of-service (QoS) multimedia applications which require IP layer multicasting from being widely deployed on the Internet, the mechanism of implementing such applications by application layer broadcasting through organizing the multicast group in a peer-to-peer overlay network is proposed. In this paper, we study the bandwidth-delay-constrained minimum spanning tree problem in an overlay network, which is NP-complete. We propose a novel genetic algorithm for resolving this problem and compare it with a state-of-the-art method. Simulation results demonstrate that the proposed algorithm is effective and efficient.

IEEE 802.11p based Dedicated Short Range Communication (DSRC) has been considered as a promising wireless technology for enhancing transportation safety and traffic efficiency. However, with limited literature available, there is lack of understanding about how IEEE 802.11p performs for vehicleto-vehicle (V2V) communications in urban environments. In this paper, we conduct intensive statistical analysis on V2V communication performance, based on the empirical measurement data collected from off-the-shelf IEEE 802.11p-compatible onboard units (OBUs). We have several key insights as follows. First, both line-of- sight (LoS) and non-line-of-sight (NLoS) durations follow power law distributions, which implies that the probability of having long LoS/NLoS conditions can be relatively high. Second, the packet inter-reception (PIR) time distribution follows an exponential distribution in LoS conditions but a power law in NLoS conditions. In contrast, the packet inter-loss (PIL) time distribution in LoS condition follows a power law but an exponential in NLoS condition. Third, the overall PIR time distribution is a mix of exponential distribution and power law distribution. The presented results provide solid ground to validate models, tune VANET simulators and improve communication strategies.

Participatory regression modeling is a cost-efficient mechanism to establish the relationships among multiple dimensions of sensory data collected from volunteers. Getting an accurate model estimate is challenging for two main reasons. First, with the concern of confidentiality of individual private data, the original data are nearly unavailable; second, low quality data with outliers are inherently embedded in the collected data. In this paper, we propose an innovative scheme, PURE, which can accurately estimate the global regression model without the need for knowing local private data (referred to as blind regression modeling) even when there is a large portion of outliers embedded. The wisdom of PURE is to let individual participants peer judge and further improve the global estimate via negotiations. Meanwhile, during the whole process, all information is exchanged in an aggregated way. By design, PURE is secure and can well protect individual privacy. Furthermore, PURE is a lightweight protocol suitable for mobile devices. Extensive trace-driven simulation results show that PURE can achieve an outstanding accuracy gain of two orders of magnitude even with random outliers near a ratio of 50 percent compared with the state-of-the-art least square estimator.

Advances in low-power wireless communications and micro-electronics make a great impact on a transportation system and pervasive deployment of road-side units (RSU) is promising to provide drive-thru Internet to vehicular users anytime and anywhere. Downloading data packets from the RSU, however, is not always reliable because of high mobility of vehicles and high contention among vehicular users. Using inter-vehicle communication, cooperative downloading can maximize the amount of data packets downloaded per user request. In this paper, we focus on effective data downloading for realtime applications (e.g., video streaming, online game) where each user request is prioritized by the delivery deadline. We propose a cooperative downloading algorithm, namely MMCD, which maximizes the amount of data packets downloaded from the RSU while minimizing delivery delay of each user request. The performance of MMCD is evaluated by extensive simulations and results demonstrate that our algorithm can reduce mean delivery delay while gaining downloading throughput as high as that of a state-of-the-art method.

With the popularity of intelligent mobile devices, enormous urban information has been generated and required by the public. In response, ShanghaiGrid (SG) aims to providing abundant information services to the public. With fixed schedule and urban-wide coverage, an appealing service in SG is to provide free message delivery service to the public using buses, which allows mobile device users to send messages to locations of interest via buses. The main challenge in realizing this service is to provide efficient routing scheme with privacy preservation under highly dynamic urban traffic condition. In this paper, we present an innovative scheme BusCast to tackle this problem. In BusCast, buses can pick up and forward personal messages to their destination locations in a store-carry-forward fashion. For each message, BusCast conservatively associates a routing graph rather than a fixed routing path with the message in order to adapt the dynamic of urban traffic. Meanwhile, the privacy information about the user and the message destination is concealed from both intermediate relay buses and outside adversaries. Both rigorous privacy analysis and extensive trace-driven simulations demonstrate the efficacy of BusCast scheme.

Vehicular networks consist of highly mobile vehicles communications, where connectivity is intermittent. Due to the distributed and highly dynamic nature of vehicular network, to minimize the end-to-end delay and the network traffic at the same time in data forwarding is very hard. Heuristic algorithms utilizing either contact-level or social-level scale of vehicular mobility have only one-sided view of the network and therefore are not optimal. In this paper, by analyzing three large sets of Global Positioning System (GPS) trace of more than ten thousand public vehicles, we find that pairwise contacts have strong temporal correlation. Furthermore, the contact graph of vehicles presents complex structure when aggregating the underlying contacts. In understanding the impact of both levels of mobility to the data forwarding, we propose an innovative scheme, named ZOOM, for fast opportunistic forwarding in vehicular networks, which automatically choose the most appropriate mobility information when deciding next data-relays in order to minimize the end-to-end delay while reducing the network traffic. Extensive trace-driven simulations demonstrate the efficacy of ZOOM design. On average, ZOOM can improve 30% performance gain comparing to the state-of-art algorithms.

In urban vehicular networks, where privacy, especially the location privacy of anonymous vehicles is highly concerned, anonymous verification of vehicles is indispensable. Consequently, an attacker who succeeds in forging multiple hostile identifies can easily launch a Sybil attack, gaining a disproportionately large influence. In this paper, we propose a novel Sybil attack detection mechanism, Footprint, using the trajectories of vehicles for identification while still preserving their location privacy. More specifically, when a vehicle approaches a road-side unit (RSU), it actively demands an authorized message from the RSU as the proof of the appearance time at this RSU. We design a location-hidden authorized message generation scheme for two objectives: first, RSU signatures on messages are signer ambiguous so that the RSU location information is concealed from the resulted authorized message; second, two authorized messages signed by the same RSU within the same given period of time (temporarily linkable) are recognizable so that they can be used for identification. With the temporal limitation on the linkability of two authorized messages, authorized messages used for long-term identification are prohibited. With this scheme, vehicles can generate a location-hidden trajectory for location-privacy-preserved identification by collecting a consecutive series of authorized messages. Utilizing social relationship among trajectories according to the similarity definition of two trajectories, Footprint can recognize and therefore dismiss “communities” of Sybil trajectories. Rigorous security analysis and extensive trace-driven simulations demonstrate the efficacy of Footprint.

Inter-contact times (ICTs) between moving vehicles are one of the key metrics in vehicular networks, and they are also central to forwarding algorithms and the end-to-end delay. Recent study on the tail distribution of ICTs based on theoretical mobility models and empirical trace data shows that the delay between two consecutive contact opportunities drops exponentially. While theoretical results facilitate problem analysis, how to design practical opportunistic forwarding protocols in vehicular networks, where messages are delivered in carry-and-forward fashion, is still unclear. In this paper, we study three large sets of Global Positioning System (GPS) traces of more than ten thousand public vehicles, collected from Shanghai and Shenzhen, two metropolises in China. By mining the temporal correlation and the evolution of ICTs between each pair of vehicles, we use higher order Markov chains to characterize urban vehicular mobility patterns, which adapt as ICTs between vehicles continuously get updated. Then, the next hop for message forwarding is determined based on the previous ICTs. With our message forwarding strategy, it can dramatically increase delivery ratio (up to 80%) and reduce end-to-end delay (up to 50%) while generating similar network traffic comparing to current strategies based on the delivery probability or the expected delay.

Providing real-time traffic information in metropolises is desired since it can not only facilitate the traffic management but also save the time of travelers on road as well as the vehicle fuel consumption which is crucial in low-carbon society. However, to obtain the traffic information is extremely difficult due to the high cost of deploying a tremendously large number of sensors on every road segments or intersections. Recently, the ShanghaiGrid (SG) project presents an innovative cost-efficient way to address this issue by deploying traffic sensors on several thousands mobile taxies. Traffic condition information perception from these sensory data is very challenging because individual taxi reports are error-prone and sparse in terms of temporal and spatial distribution. In this paper, we use a data aggregation approach to overcome the aforementioned challenge, i.e., the ”error-prone” problem and ”sparse” problem. We first extensively study the characteristics of the measurement data from over 3000 operational taxies in Shanghai City. Utilizing the spatial correlation of traffic conditions, we propose a correlation based traffic estimation algorithm to successfully expand the coverage of taxi sensors. Our experimental result demonstrates the significance of the proposed algorithm by providing the traffic information at any time and any location in Shanghai City.

No abstract available.