Large Language Model (LLM) inference services like ChatGPT are popular for enabling diverse tasks via prompts, yet they exacerbate privacy risks due to the potential exposure of sensitive data in user inputs. Existing local differential privacy (LDP)-based text sanitization mechanisms offer lightweight protection suitable for cloud-based LLM inference. Nevertheless, uniform privacy budget allocation and generalized sanitization mechanisms neglect the critical protection needs of sensitive user data, such as Personally Identifiable Information (PII). Empirical evidence of this work reveals that even with a strict privacy budget (ϵ=0.1), the sensitive information leakage rate can reach an alarmingly high 71.74%. To address these challenges, this paper proposes Rap-LI, a risk-aware privacy preservation framework for LLM inference, designed to be plug-and-play. Rap-LI performs risk identification and personalized labeling on user prompts, then develops a risk-aware LDP mechanism for text sanitization, formally proven to satisfy both token-level and sentence-level LDP guarantees. Extensive experimental results demonstrate Rap-LI’s superior privacy-utility balance. It improves privacy protection against sensitive information leakage by an average of 51.68% compared to methods with comparable utility. Our code is available at https://github.com/Cristliu/RapLI.

Text-to-image person re-identification (TIReID) aims to retrieve semantically related images from a large gallery given a text query. Most existing TIReID methods train the model on an ideal assumption that positive (negative) samples are semantically correlated (uncorrelated) on the visual and textual modal. However, we observe that incorrect annotation is unavoidable and ambiguous textual description is often taken as the input in practice, which leads to false negatives that ruin the feature aligning between modals for deviated identification. This work presents a novel False Negative Mitigation (FNM) method by identifying potential false negatives through distribution differences and mitigating them with dedicated loss. Specifically, the false negative mitigation loss is designed to adaptively adjust the optimization margin for potential false negatives in the latent space. Moreover, a Dual-Level Feature Representation method is introduced to leverage both global and local features during the mitigation, whereas a Momentum Contractive (MoC) module is plugged to enrich the training data for accurate similarity distribution estimation. We conduct extensive experiments on three public benchmark datasets and the results demonstrate that our FNM method achieves State-of-The-Art (SoTA) performance at all evaluation metrics.

Turning the multi-round vanilla Federated Learning into one-shot FL (OFL) significantly reduces the communication burden and makes a big leap toward practical deployment. However, this work empirically and theoretically unravels that existing OFL falls into a garbage (inconsistent one-shot local models) in and garbage (degraded global model) out pitfall. The inconsistency manifests as divergent feature representations and sample predictions. This work presents a novel OFL framework FAFI that enhances the one-shot training on the client side to essentially overcome inferior local uploading. Specifically, unsupervised feature alignment and category-wise prototype learning are adopted for clients’ local training to be consistent in representing local samples. On this basis, FAFI uses informativeness-aware feature fusion and prototype aggregation for global inference. Extensive experiments on three datasets demonstrate the effectiveness of FAFI, which facilitates superior performance compared with 11 OFL baselines (+10.86% accuracy). Code available at https://github.com/zenghui9977/FAFI_ICML25

Classifying failures for automated diagnosis is critical for maintaining the reliability of ever-increasing microservice systems.Existing failure classification methods focus solely on analyzing one specific data modality (e.g., metrics) or feeding multi-modal data (e.g., logs and traces) in a one-fit-all model.This would lead to biased inference as different modalities have differentiated proportions and manifest distinct sensitivities and patterns in failures of varied runtime contexts.This work proposes a novel failure classification framework for microservices, named AgileMFC, with the basic idea of flexibly learning and assembling the expertise of multi-modal monitoring data.The flexibility is attained with the split-and-merge design of data representation, where it assigns each modality a different gate that routes it to dedicated expert networks and uses a sharing gate to facilitate controllable knowledge interaction among modalities.Meanwhile, AgileMFC trains the representation networks and a classifier in consecutive stages with concatenated and weighted feature fusion, thus decoupling task-agnostic feature extraction with task-specific classification.Experiments on three datasets show that AgileMFC outperforms four baselines w.r.t., precision, recall, and F1-score on failure classification.

Large Language Models (LLMs) are increasingly gaining traction in the healthcare sector, yet expanding the threat of sensitive health information being easily exposed and accessed without authorization. These privacy risks escalate in regions like China, where privacy awareness is notably limited. While some efforts have been devoted to user surveys on LLMs in healthcare, users' perceptions of privacy remain unexplored. To fill this gap, this paper contributes the first user study (n=846) in China on privacy awareness and expectations in LLM-based healthcare consultations. Specifically, a healthcare chatbot is deployed to investigate users' awareness in practice. Information flows grounded in contextual integrity are then employed to measure users' privacy expectations. Our findings suggest that the prevalence of LLMs amplifies health privacy risks by raising users' curiosity and willingness to use such services, thus overshadowing privacy concerns. 77.3% of participants are inclined to use such services, and 72.9% indicate they would adopt the generated advice. Interestingly, a paradoxical “illusion” emerges where users' knowledge and concerns about privacy contradict their privacy expectations, leading to greater health privacy exposure. Our extensive discussion offers insights for future LLM-based healthcare privacy investigations and protection technology development.

Using hyperscale surveillance cameras, seamless target tracking can be accomplished in urban security scenarios, significantly enhancing public security and emergency response capabilities. In spite of the advantage of edge computing, tracking multiple targets using multiple cameras would incur prohibitive high computation costs. Based on the deployment of real-world cameras, this research finds that existing tracking scheduling is inefficient as a result of redundant and excessive activation of cameras. As a follow-up, the research proposes a hierarchical tracking framework called CityTrac that leverages fine-grained target movement predictions to provide efficient tracking in hyperscale cameras. First, CityTrac uses a specially designed camera selection strategy that ensures accurate tracking with a minimum number of cameras. After that, CityTrac constructs a probabilistic target movement graph by using historical tempo-spatial correlation information. Using the graph as a model, the tracking scheduling and camera selection problem are formulated as an optimization problem with efficiency-accuracy tradeoff constraints. The research addresses this NP-hard problem using greedy optimization. The experiments conducted with the Cityflow and Geolife datasets demonstrate that, compared with two baselines, CityTrac requires significantly fewer computation resources (over 90%) in order to track the same number of targets with the same level of accuracy.

Time series data from monitoring applications captures the behaviours of objects, which can often be split into distinguishable segments that reflect the underlying state changes. Despite the recent advances in time series state detection, the indicator selection for state detection is rarely studied, most of state detection work assumes the input indicators have been properly or manually selected. However, this assumption is disconnected from practice, on one hand, manual selection is not scalable, there can be up to thousands of indicators for the runtime monitoring of certain objects, e.g., supercomputer systems. On the other hand, performing state detection on a large amount of raw indicators is both inefficient and redundant. We argue that indicator selection should be made an upstream task for selecting a subset of indicators to facilitate state analysis. To this end, we propose ISSD ( I ndicator S election for S tate D etection), an indicator selection method for time series state detection. At its core, ISSD attempts to find an indicator subset that has as much high-quality states, which is measured by the channel set completeness and quality we invent based on segment-level sampling statistics. Such an indicator selection process is transformed into a multi-objective optimization problem and an approximation algorithm is designed to solve the NP-hard searching for specific end point in the Pareto front. Experiments on 5 datasets and 4 downstream methods show that ISSD has significant selection superiority compared with 6 baselines. We also elaborate on two observations of selection resilience and channel sensitivity of existing state detection methods and appeal to further research on them.

Sequential data learning is vital to harnessing the encompassed rich knowledge for diverse downstream tasks, particularly in healthcare (e.g., disease prediction). Considering data sensitiveness, privacy-preserving learning methods, based on federated learning (FL) and split learning (SL), have been widely investigated. Yet, this work identifies, for the first time, existing methods overlook that sequential data are generated by different patients at different times and stored in different hospitals, failing to learn the sequential correlations between different temporal segments. To fill this void, a novel distributed learning frameworkSTSLis proposed by training a model on the segments in order. Considering that patients have different visit sequences,STSLfirst implements privacy-preserving visit ordering based on a secure multi-party computation mechanism. Then batch scheduling participates patients with similar visit (sub-)sequences into the same training batch, facilitating subsequent split learning on batches. The scheduling process is formulated as an NP-hard optimization problem on balancing learning loss and efficiency and a greedy-based solution is presented. Theoretical analysis proves the privacy preservation property ofSTSL. Experimental results on real-world eICU data show its superior performance compared with FL and SL ($5\% \sim 28\%$better accuracy) and effectiveness (a remarkable 75% reduction in communication costs).

Deep cross-modal retrieval, with its effective and efficient search capabilities, has gained widespread adoption in today’s media-sharing practices yet raises concerns regarding potential threats to user data privacy. The cutting-edge data-centric countermeasures usually adopt adversarial learning, i.e., laboriously crafting the proper perturbation for each image, resulting in the noticeable noise in adversarial examples that greatly undermines the aesthetic appeal of image sharing. To address this issue, we propose a novel Model-centric Cross-modal Privacy-preserving framework (MCP), wherein the pre-defined invisible backdoor is seamlessly integrated into the global retrieval model via backdoor learning, thereby effectively preventing shared images containing such triggers from being retrieved. Specifically, we introduce a simple yet effective cross-modal backdoor learning algorithm that alternately optimizes two losses: 1) a privacy-preserving loss for perturbing retrieval with a user-injected trigger and 2) the standard utility loss for maintaining normal retrieval performance. Compared to state-of-the-art methods, MCP excels in providing excellent stealthiness, manifesting in a notable improvement of approximately 100% in SSIM metrics. Furthermore, it achieves an outstanding privacy-preserving (backdoor) success rate, as evidenced by a substantial mAP reduction of 22.3% (for FashionVC), 11.5% (for NUS-WIDE), and 21.8% (for MIRFlickr-25K) in poisoned retrieval, while maintaining similar normal retrieval performance. Additionally, MCP exhibits robust resistance against potential black-box defenses (e.g., trigger filtering) and white-box defenses (e.g., fine-tuning and model pruning). The code and data are available athttps://github.com/lqsunshine/MCP.

Cloud environments enhance diffusion model efficiency but introduce privacy risks, including intellectual property theft and data breaches. As AI-generated images gain recognition as copyright-protected works, ensuring their security and intellectual property protection in cloud environments has become a pressing challenge. This paper addresses privacy protection in diffusion model inference under cloud environments, identifying two key characteristics—denoising-encryption antagonism and stepwise generative nature—that create challenges such as incompatibility with traditional encryption, incomplete input parameter representation, and inseparability of the generative process. We propose PPIDM (Privacy-PreservingInference forDiffusionModels), a framework that balances efficiency and privacy by retaining lightweight text encoding and image decoding on the client while offloading computationally intensive U-Net layers to multiple non-colluding cloud servers. Client-side aggregation reduces computational overhead and enhances security. Experiments show PPIDM offloads 67% of Stable Diffusion computations to the cloud, reduces image leakage by 75%, and maintains high output quality (PSNR = 36.9, FID = 4.56), comparable to standard outputs. PPIDM offers a secure and efficient solution for cloud-based diffusion model inference.

Leveraging trace-rich features within embedded spaces has been established as effective in image manipulation localization (IML). Nevertheless, the feature of manipulated traces frequently comprises substantial redundant information only loosely related to IML tasks. This complexity has hindered existing methods in fully comprehending the essence of trace features. In light of this challenge, we introduce a novel decoupling representation learning network (DRN) tailored for IML. The DRN excels at decoupling intricate multidomain information and transforming it into representations directly pertinent to IML objectives. This is achieved through a meticulously designed frequency decoupling representation learning module (FDM) and spatial decoupling representation learning module (SDM). Specifically, the FDM operates by acquiring distinct low and high-frequency components to effectively decouple redundant information. The decoupled high-frequency components are then harnessed as intricate trace complements, enhancing the overall aggregation process. In addition, the redundant information is expertly separated into authentic and manipulated representations through the use of channel activation maps in SDM. Through extensive experimentation on three public benchmarks including CASIA, NIST, and Coverage, our method consistently demonstrates superior performance and enhanced robustness compared with existing state-of-the-art methods.

Time-Sensitive Networking (TSN) is an emerging real-time Ethernet technology that provides deterministic communication for time-sensitive (TS) traffic. At its core, TSN utilizes Per-Stream Filtering and Policing (PSFP) gates to mitigate the disruption of unavoidable frame drift. However, as first identified in this work, the naive PSFP gate design results in heavy memory usage, which hinders normal switching functions. This work proposes an efficient PSFP gate design called FooDog. FooDog employs a two-stage structure and a dual-engine policing mechanism to realize memory-efficient, logic-compact, and fast policing while maintaining minimal latency and jitter for TS traffic. Results on FPGA prototypes show that FooDog consumes only hundreds of kilobits of memory, reducing on-chip memory overheads by more than 90% compared to the unoptimized PSFP gate design. Additionally, it maintains end-to-end latency in the microsecond range and jitter below 150 nanoseconds under abnormal traffic conditions, comparable to typical TSN performance without anomalies.

Text-based Person Re-identification (T-ReID), which aims at retrieving a specific pedestrian image from a collection of images via text-based information, has received significant attention. However, previous research has overlooked a challenging yet practical form of T-ReID: dealing with image galleries mixed with occluded and inconsistent personal visuals, instead of ideal visuals with a full-body and clear view. Its major challenges lay in the insufficiency of benchmark datasets and the enlarged semantic gap incurred by arbitrary occlusions and modality gap between text description and visual representation of the target person. To alleviate these issues, we first design an Occlusion Generator (OGor) for the automatic generation of artificial occluded images from generic surveillance images. Then, a fine-granularity token selection mechanism is proposed to minimize the negative impact of occlusion for robust feature learning, and a novel multi-granularity contrastive consistency alignment framework is designed to leverage intra-/inter-granularity of visual-text representations for semantic alignment of occluded visuals and query texts. Experimental results demonstrate that our method exhibits superior performance. We believe this work could inspire the community to investigate more dedicated designs for implementing T-ReID in real-world scenarios. The source code is available at https://github.com/littlexinyi/MGCC.

Alongside the high performance built on massive hardware, ever-larger computer systems bear tons of hardware alerts every day during reliability maintenance. Based on an exploratory study on a representative supercomputer system, this work first characterizes supercomputer alerts as an overload of continuous bursts for the operators. Yet, existing similarity-based aggregation solutions, tuned for in-band textual alerts, are myopic by finding dissimilar representatives instead of looking into the semantics in the supercomputer context. To fill the void of supercomputer alert aggregation, we propose the SuperAgg framework to extract the hierarchical patterns of real-world alerts and use them for online alert management. SuperAgg jointly integrates unsupervised state detection of time series and expert analysis to successfully discover 4 categories of sensor-tier alert patterns and exploits primary-and-secondary statistics between sensors for system-tier correlation patterns. With such extracted knowledge, SuperAgg then identifies the formulated patterns online and uses spatiotemporal combined strategies to reduce the alert influx. Evaluations on alerts generated from a production supercomputer show that SuperAgg provides over 98% aggregation rate and significantly higher aggregation accuracy (over 83.8% and 43.2% on different datasets) than 3 baselines. Production deployment further demonstrates its effectiveness from the perspective of system operators. The source code is available at: https://github.com/Txh-User/SuperAgg.

Transforming the multi-round vanilla Federated Learning (FL) into one-shot FL (OFL) significantly reduces the communication burden and makes a big leap toward practical deployment. However, we note that existing OFL methods all build on model lossy reconstruction (i.e., aggregating while partially discarding local knowledge in clients' models), which attains one-shot at the cost of degraded inference performance. By identifying the root cause of stressing too much on finding a one-fit-all model, this work proposes a novel one-shot FL framework by embodying each local model as an independent expert and leveraging a Mixture-of-Experts network to maintain all local knowledge intact. A dedicated self-supervised training process is designed to tune the network, where the sample generation is guided by approximating underlying distributions of local data and making distinct predictions among experts. Notably, the framework also fuels FL with flexible, data-free aggregation and heterogeneity tolerance. Experiments on 4 datasets show that the proposed framework maintains the one-shot efficiency, facilitates superior performance compared with 8 OFL baselines (+5.54% on CIFAR-10), and even attains over ×4 performance gain compared with 3 multi-round FL methods, while only requiring less than 85% trainable parameters. Our code will be available at https://github.com/zenghui9977/IntactOFL.

Abstract Accurate and current comprehension of network status is crucial for efficient network management. Nevertheless, direct network measurement strategies entail substantial traffic overhead and demand intricate coordination among network entities, making them impractical. Network tomography, an indirect measurement approach, utilizes insights garnered from measured parts to deduce characteristics of the entire network. Past studies frequently depend on acquiring challenging-to-access information, such as the complete network topology or support from specialized protocols. Unfortunately, these constraints pose challenges in non-cooperative scenarios where obtaining such information is difficult. Recent endeavors pursue emancipating tomography from dependence on copious information, striving to predict unmeasured path performance using limited data. Nevertheless, the disparity between the measured data and actual performance has hindered the accuracy. In response, we introduce an innovative tomography framework named DRL-Tomo, designed to alleviate potential biases. DRL-Tomo initiates by generating augmented data through deep reinforcement learning, gradually approximating the genuine performance of unmeasured paths. Subsequently, a neural network model is trained using this augmented data, enabling precise inferences. Our experiments, encompassing both real-world and synthetic datasets, vividly demonstrate DRL-Tomo’s remarkable enhancement. Specifically, it achieves a substantial 10%–67% improvement in path delay prediction and an impressive 30%–98% enhancement in path loss rate prediction.

Abstract Nowadays, with the continuous increase in internet traffic, the demand for real-time and high-speed traffic analysis has grown significantly. However, existing traffic analysis technologies are either limited by specific applications or data, unable to expand for widespread implementation, or in offline mode are unable to keep up with dynamic adjustments required in certain network management scenarios. A promising approach is to utilize sketch technology to enhance real-time traffic analysis. Unfortunately, existing technologies suffer from defects, such as overly coarse-grained statistics that cannot perform precise application-level traffic analysis, and irreversibility, which cannot support real-time queries in a friendly way. To achieve real-time fine-grained application traffic analysis in general scenarios, we propose AppSketch, a real-time network traffic measurement tool. AppSketch adopts a one-pass approach to classify and label the application information of each packet in the network flows. It then hashes the flow, identified with the application tag, into a carefully designed multiple-key sketch, for gathering application-specific statistics. We conducted extensive experiments using a real-world network traffic dataset collected on a university campus. The results showed that AppSketch achieved high accuracy while requiring less update time than other alternatives. Moreover, AppSketch occupies limited memory ($ {\leq }$64KB), making it suitable for online network devices.

Designers widely accept AI as a partner in the design process for its efficient and intelligent decision-making. However, AI is often not perfect, and AI error often makes humans dumbfounded. Literature has pointed out the value of such AI error, while still leaving its inspiration essence and application strategies uncharted from the practice perspective. This work focuses on bridging the practice gap by looking into and exploiting the imaginative "mislabeled" objects of object detection models. To gain insights into the inspiration of AI "error", we collected a dedicated AI "error" dataset from object detection and invited eight designers to share divergent comments on the "mislabeled" objects. Coding was then performed on the comments, which summarizes the inspiration of AI "error" into six atomic dimensions. Subsequently, we took a step further to an exploratory study, a comparative ideation experiment with 20 designers, investigating how to apply these inspiration dimensions to create ideas. Questionnaire and interview results revealed that essential inspiration of AI "error" could positively activate creativity, especially the "Outline" dimension. A design model CETR is then formulated by summarizing the application of atomic inspiration of "error" into four forms of creativity, which could be taken as a guideline for cooperative design with AI "error". In addition, we also sketch two approaches to generate more inspiring and applicable AI "error", elaborate on two principal characteristics of AI "error" for promoting creativity, and propose three strategies for better co-creating with AI "error". Finally, we provide insight into design research about AI self-awareness and human-AI collaboration.

Increasing artwork plagiarism incidents stresses the urgent need for proper copyright protection on behalf of the creators. The latest development in this context focuses on embedding watermarks via deep encoder-decoder networks. However, we find that deep watermarking has a serious vulnerability on its robustness when facing deliberate plagiarism. To manifest it, we construct an attack that misuses watermarking encoder as a plagiarism lookout for bypassing copyright detection. As a remedy, we propose a patch-level deep watermarking framework (DIPW) to retain copyright evidence in essential patches with plagiarism resistance, inspired by a user study observation that subject elements in artworks are the principal plagiarism entities. Technically, DIPW adaptively finds the embedding patches by identifying a subset of non-overlapping and feature-rich objects; and tailors the model with dual-distortion losses and adversarial plagiarism noise injection for robustness. Experimental results demonstrate the superiority of DIPW in facilitating better robustness, secrecy, and imperceptibility with acceptable time burden.

Remote sensing image-text retrieval (RSITR) has become a research hotspot in recent years for its wide application. Existing methods in this context, based either on local or global feature matching, overlook the sensing variation-leaded visual deviation and geographically nearby image-text mismatching problems of remote sensing (RS) images. This work notes that this would limit the retrieval accuracy for RSITR. To handle this, we present IERR, an implicit-explicit relation reasoning framework that learns relations between local visual-textual tokens and enhances global image-text matching without requiring additional prior supervision. Specifically, masked image modeling (MIM) and masked language modeling (MLM) are used for symmetric mask reasoning consistency alignment. Meanwhile, masked features (i.e., implicit relation) and unmasked features (i.e., explicit relation) are fed into a multimodal interaction encoder to enhance the representations of the textual-visual features. Extensive experimental results on the RSICD and RSITMD datasets demonstrate the superiority of IERR compared with 17 baselines.

Text-based person re-identification (ReID) has enabled canonical applications in searching for and tracking targets from large-scale surveillance images with textual descriptions. Yet, existing text-based person ReID systems employ centralized model training that gathers images captured by different institutes' cameras into one place, which poses severe privacy threats to sensitive institutional information. This work is then devoted to exploring privacy-preserving text-based person ReID and proposes the framework of FedSH by tailoring the federated learning paradigm for distributed searching knowledge extraction. Specifically, FedSH resolves the local model generalization and entity boundary obscuring limitations, caused by inner-institute data homogeneity and inter-institute data heterogeneity, via building multi-granularity feature representation and a semantically self-aligned network. Meanwhile, it reduces the communication burden introduced by the embedding for multiple modals by updating common representation subspaces during federated learning. Experimental results on two public benchmarks demonstrate that our method can achieve at most 16.47% and 16.02% person ReID performance improvement by the Rank-1 metric, compared with 6 State-of-The-Art (SoTA) baselines and 6 ablation studies. We believe that our work will inspire the community to investigate the potential of implementing Federated Learning in real-world image retrieval and ReID scenarios.

Recently, a hierarchical fine-grained fusion mechanism has been proved effective in cross-modal retrieval between videos and texts. Generally, the hierarchical fine-grained semantic representations (video-text semantic matching is decomposed into three levels including global-event representation matching, action-relation representation matching, and local-entity representation matching) to be fused can work well by themselves for the query. However, in real-world scenarios and applications, existing methods failed to adaptively estimate the effectiveness of multiple levels of the semantic representations for a given query in advance of multilevel fusion, resulting in a worse performance than expected. As a result, it is extremely essential to identify the effectiveness of hierarchical semantic representations in a query-adaptive manner. To this end, this article proposes an effective query-adaptive multilevel fusion (QAMF) model based on manipulating multiple similarity scores between the hierarchical visual and text representations. First, we decompose video-side and text-side representations into hierarchical semantic representations consisting of global-event level, action-relation level, and local-entity level, respectively. Then, the multilevel representation of the video-text pair is aligned to calculate the similarity score for each level. Meanwhile, the sorted similarity score curves of the good semantic representation are different from the inferior ones, which exhibit a "cliff" shape and gradually decline (see Fig. fig1 as an example). Finally, we leverage the Gaussian decay function to fit the tail of the score curve and calculate the area under the normalized sorted similarity curve as the indicator of semantic representation effectiveness, namely, the area of good semantic representation is small, and vice versa. Extensive experiments on three public benchmark video-text datasets have demonstrated that our method consistently outperforms the state-of-the-art (SoTA). A simple demo of QAMF will soon be publicly available on our homepage: https://github.com/Lab-ANT.

Personalized Federated Learning (FL) handles the data heterogeneous problem by tailoring local models for each distributed data owner. Previous studies first train a highly-adaptable global model and then transfer it for personalization. However, the additional training aggravates burden of resource-limited end devices. Training a personalized local sub-network is a promising efficient solution. It normally prunes the global model by parameters’ scalar magnitude. In this paper, we found that the vector magnitude, i.e. the parameter stability, could further promote personalized FL. Driven by the local data characteristics, the values of some model parameters are hardly changed in their updates. But they consume the same resources as the changed ones. Thus, we propose Star-PFL, a STability-AwaRe algorithm for efficient FL Personalization. In Star-PFL, the data owner focuses on training non-stabilized parameters, and decreases the resource wastes on stabilized ones. Experimental results on two real-world biomedical datasets demonstrate that Star-PFL improves the accuracy (3.1%↑) and decreases the resource costs (communication 36.3%↓, computation 18.3%↓) than 5 typical baselines. The code is available at https://github.com/Guoyeting/Star-PFL.

Hyper-scale surveillance cameras have enabled seamless target (e.g., vehicles, individuals) tracking in urban scenarios, significantly improving everyday safety and emergency response capacity. However, practically tracking multiple targets among cameras would incur prohibitive huge computation costs, even with the acceleration utility of edge computing. This work observes with real-world camera deployment that existing tracking scheduling is unfortunately inefficient due to the redundant and over-activation of cameras. We then propose the design of a hierarchical tracking framework HyMOT that exploits fine-grained target movement prediction for efficient tracking in hyper-scale cameras. At its core, HyMOT builds a probabilistic target movement graph with tempo-spatial correlation knowledge extracted from historical statistics. With the graph, we formulate tracking scheduling as an optimization problem with efficiency-accuracy tradeoff constraints and solve this NP-hard problem with a greedy strategy. Experiments based on the Cityflow and Geolife datasets demonstrate that, compared with two baselines, Hy-MOT requires significantly less (over 90%) computation workloads to track the same amount of targets with similar accuracy.

Global-scale IPv6 scan, critical for network measurement and management, is still a mission to be accomplished due to its vast address space. To tackle this challenge, IPv6 scan generally leverages pre-defined seed addresses to guide search directions. Under this general principle, however, the core problem of effectively using the seeds is largely open. In this work, we propose a novel IPv6 active search strategy, namely HMap6, which significantly improves the use of seeds, w.r.t. the marginal benefit, for large-scale active address discovery in various prefixes. Using a heuristic search strategy for efficient seed collection and alias prefix detection under a wide range of BGP prefixes, HMap6 can greatly expand the scan coverage. Real-world experiments over the Internet in billion-scale scans show that HMap6 can discover 29.39M unique /80 prefixes with active addresses, an 11.88% improvement over the state-of-the-art methods. Furthermore, the IPv6 hitlists from HMap6 include all-responsive IPv6 addresses with rich information. This result sharply differs from existing public IPv6 hitlists, which contain non-responsive and filtered addresses, and pushes the IPv6 hitlists from quantity to quality. To encourage and benefit further IPv6 measurement studies, we released our tool along with our IPv6 hitlists and the detected alias prefixes.

Increasing artwork plagiarism incidents underscores the urgent need for reliable copyright protection for high-quality artwork images. Although watermarking is helpful to this issue, existing methods are limited in imperceptibility and robustness. To provide high-level protection for valuable artwork images, we propose a novel invisible robust watermarking framework, dubbed as IRWArt. In our architecture, the embedding and recovery of the watermark are treated as a pair of image transformations’ inverse problems, and can be implemented through the forward and backward processes of an invertible neural networks (INN), respectively. For high visual quality, we embed the watermark in high-frequency domains with minimal impact on artwork and supervise image reconstruction using a human visual system(HVS)-consistent deep perceptual loss. For strong plagiarism-resistant, we construct a quality enhancement module for the embedded image against possible distortions caused by plagiarism actions. Moreover, the two-stagecontrastive training strategy enables the simultaneous realization of the above two goals. Experimental results on 4 datasets demonstrate the superiority of our IRWArt over other state-of-the-art watermarking methods. Code: https://github.com/1024yy/IRWArt.

Massive amounts of high-quality data are the prerequisite and support for AI technologies. Due to the nature of privacy-preserving and low communication overheads, federated learning (FL) has garnered considerable attention in comparison with traditional data collection methods. However, the performance of FL is hampered by the lack of interested clients and limited local data due to selfishness and individual behavioral preferences. To this end, we propose PractFL, a proactive and distributed framework that incorporates the concept of mobile crowdsensing into the FL paradigm. Specifically, we design an incentive mechanism in the form of virtual red packets, which are a widely used way of monetary reward and gift-giving in social lives. In this article we extend this further by giving meaning to the locations, i.e., the red packets are only accessible at specific places. The virtual red packets’ locations and monetary amounts can be dynamically updated by the cloud center to encourage clients to collect additional data that may benefit the FL process. Further, we propose a distributed behavioral decision engine based on multiarmed bandits (i.e., choose which red packet to go for) in response to the incentive mechanism enforced by the cloud. Considering the movement cost and conflicts with other clients,$K$-anonymity and probabilistic selection are introduced in the distributed behavioral decision to recommend the optimal red packet choice for clients without revealing their privacy. The experimental results demonstrate that PractFL outperforms the baselines in terms of classification accuracy. We also find that PractFL can effectively alleviate the overfitting problem caused by class imbalance during the training.

Time series data from monitoring applications reflect the physical or logical states of the objects, which may produce time series of distinguishable characteristics in different states. Thus, time series data can usually be split into different segments, each reflecting a state of the objects. These states carry rich high-level semantic information, e.g., run, walk, or jump, which helps people better understand the behaviour of the monitored objects. Nevertheless, these states are latent and hard to discover, because the characteristic of time series is complicated and the computational cost is high. This paper develops an efficient and effective unsupervised approach for inferring the latent states of massive multivariate time data. To reduce the computational cost, we present Time2State, a scalable framework that utilizes a sliding window and an encoder to greatly reduce the length of raw time series. To train the encoder, we propose a novel unsupervised loss function, LSE-Loss. Extensive experiments show that compared to the state-of-the-art time series representation learning methods of the same kind, LSE-Loss brings a performance improvement of up to 15% in accuracy.

The popularization of smart Internet of Things (IoT) devices has facilitated the development of fog/edge computing. However, these infrastructure-based service paradigms may fail to complete tasks successfully due to computation and communication overload, or damage in challenging scenarios such as disasters or traffic jams. Noticing that a crowd of devices with considerable idle resources could be available, we investigate the problems of addressing the computation and communication unavailability with peer assistance in this work. To this end, we propose a dispersed service framework for resource-exhausted scenarios that adaptively offloads users’ data to available network computation points. However, the users may not be able to achieve the offloading due to geographical hindrances. Consequently, the relay is introduced as a bridge for data offloading between the users and the network computation points. Furthermore, a game-based incentive-driven offloading mechanism is designed by analyzing and balancing the cost and gain factors of three main entities (users, relays, and network computation points). Considering the interactions among the entities, a two-level Stackelberg game is established for efficiently allocating potential computation resource, as well as balancing the utility conflicts due to the data offloading. Given the hierarchical interaction structure, the upper level game involves network computation points as followers and the relay as a leader, while the lower level game includes the relay as a follower and users as leaders. Moreover, to facilitate applicability in large-scale scenarios with multiple relays, we decompose multiple relays into multiple single relay problems using a tripartite matching strategy that assigns appropriate relays to users and network computation points. The simulation results demonstrate the effectiveness of the proposed game-based incentive-driven mechanism and show that it outperforms the baselines in terms of the overall utilities of the involved entities and the average energy consumption of users.

Cross-modal retrieval aims to enable a flexible bi-directional retrieval experience across different modalities (e.g., searching for videos with texts). Many existing efforts tend to learn a common semantic representation embedding space in which items of different modalities can be directly compared, wherein the positive global representations of video-text pairs are pulled close while the negative ones are pushed apart via pair-wise ranking loss. However, such a vanilla loss would unfortunately yield ambiguous feature embeddings for texts of different videos, causing inaccurate cross-modal matching and unreliable retrievals. Toward this end, we propose a multimodal contrastive knowledge distillation method for instance video-text retrieval, called MCKD, by adaptively using the general knowledge of self-supervised model (teacher) to calibrate mixed boundaries. Specifically, the teacher model is tailored for robust (less-ambiguous) visual-text joint semantic space by maximizing mutual information of co-occurred modalities during multimodal contrastive learning. This robust and structural inter-instance knowledge is then distilled, with the help of explicit discrimination loss, to a student model for improved matching performance. Extensive experiments on four public benchmark video-text datasets (MSR-VTT, TGIF, VATEX, and Youtube2Text) demonstrate that our MCKD can achieve at most 8.8%, 6.4%, 5.9%, and 5.3% improvement in text-to-video performance by the$\text{R}\text{@}1$metric, compared with 14 SoTA baselines.

Due to the immense benefits of guaranteeing user privacy, popular messaging platforms have shown enthusiasm for deploying End-to-End Encryption (E2EE). However, E2EE could be misused for bypassing media moderation, opening a shortcut for the viral spreading of harmful media. Private hash-matching techniques are proposed to identify harmful content in E2EE. Unfortunately, the pioneering solution incurs prohibitively high latency due to redundant user-cloud interactions for a private inspection. In this paper, we designEntbergenfor efficient inspection of E2EE media by differentially handling harmless and harmful ingredients. For this, a novel Private-2D BloOm filter with Fuzzy Query (PBO-FQ) is designed for local, agile, and private media hash matching. It is proposed as the first structure that adapts inverted index and differential privacy (DP) towards seamless integration of sketch and mask encoding. With PBO-FQ,Entbergencan instantly filter out harmless media and only pays attention to the small-scale counterparts by scrutinizing them privately based on homomorphic encryption. Security analysis shows thatEntbergencan effectively fulfil the desired privacy requirements. Extensive evaluations demonstrate thatEntbergenis sufficiently efficient (w.r.t. computation and communication overhead) for working on mobile devices and can easily scale to real-world inspection with a large database.

The pervasive view of the mobile crowd bridges various real-world scenes and people's perceptions with the gathering of distributed crowdsensing photos. To elaborate informative visuals for viewers, existing techniques introduce photo selection as an essential step in crowdsensing. Yet, the aesthetic preference of viewers, at the very heart of their experiences under various crowdsensing contexts (e.g., travel planning), is seldom considered and hardly guaranteed. We propose CrowdPicker, a novel photo selection framework with adaptive aesthetic awareness for crowdsensing. With the observations on aesthetic uncertainty and bias in different crowdsensing contexts, we exploit a joint effort of mobile crowdsourcing and domain adaptation to actively learn contextual knowledge for dynamically tailoring the aesthetic predictor. Concretely, an aesthetic utility measure is invented based on the probabilistic balance formalization to quantify the benefit of photos in improving the adaptation performance. We prove the NP-hardness of sampling the best-utility photos for crowdsourcing annotation and present a (1-1/e) approximate solution. Furthermore, a two-stage distillation-based adaptation architecture is designed based on fusing contextual and common aesthetic preferences. Extensive experiments on three datasets and four raw models demonstrate the performance superiority of CrowdPicker over four photo selection baselines and four typical sampling strategies. Cross-dataset evaluation illustrates the impacts of aesthetic bias on selection.

Word-character lattice models have been proved to be effective for some Chinese natural language processing (NLP) tasks, in which word boundary information is fused into character sequences. However, due to the inherently unidirectional sequential nature, prior approaches have only learned sequential interactions of character-word instances but fail to capture fine-grained correlations in word-character spaces. In this article, we propose a lattice-aligned attention network (LAN) that aims to model dense interactions over word-character lattice structure for enhancing character representations. By carefully combining cross-lattice module, gated word-character semantic fusion unit, and self-lattice attention module, the network can explicitly capture fine-grained correlations across different spaces (e.g., word-to-character and character-to-character), thus significantly improving model performance. Experimental results on three Chinese NLP benchmark tasks demonstrate that LAN obtains state-of-the-art results compared to several competitive approaches.

Efficient Internet-wide scanning plays a vital role in network measurement and cybersecurity analysis. While Internet-wide IPv4 scanning is a solved problem, Internet-wide scanning for IPv6 is still a mission yet to be accomplished due to its vast address space. To tackle this challenge, IPv6 scanning generally needs to use pre-defined seed addresses to guide further IPv6 scanning directions. Under this general principle, various solutions have been developed, but all suffer from two primary pitfalls, low hit rate and low probing speed, caused by the inherent sparse distribution of active IPv6 addresses and the high computational complexity of the search algorithms, respectively. We develop 6Scan, a novel asynchronous IPv6 scanner that effectively addresses the above two problems. To increase the hit rate, 6Scan infers the promising search directions by encoding the regional identifiers of the target addresses within the probing packets and recording the regional activities from the asynchronously arrived replies. It then dynamically adjusts the search directions according to the scanning result of the previous steps. To speed up the search algorithm, 6Scan leverages the regional identifier encoding to quickly adjust search direction without excessive computation. Real-world experiments over the IPv6 Internet in a billion-scale probing budget show that compared with the state-of-the-art solutions, on average 6Scan can discover 6% more active addresses with nearly the same scanning time.

In the global Internet, the paths between two autonomous systems (ASes), which are used for the exchange of traffic, are essential for understanding the behavior of the Internet routing system and they can help improve the performance of many applications of the Internet. Popular approaches to obtain the AS path between an AS pair (AP) are measurement based (e.g., Traceroute), but considering the size of the modern Internet and the limitations of measurement resources, only paths between a very small portion of APs can be measured. In recent years, a large body of path inference approaches has been proposed to bridge the gap in measurement resources. However, as we show with experiments, they perform poorly in accuracy and coverage. We propose a generative measurable path inference (GMPI) framework for AS-level path measurement, which performs well in accuracy and coverage. GMPI addresses two limitations of previous approaches: 1) Information incompleteness due to unrevealed real-world AS-level routing policies and insufficient measuring resources. 2) Knowledge isolation caused by distributed AS knowledge with different sources and inconsistent forms. To overcome these challenges, the data-driven GMPI framework invents heuristic path generation to address incompleteness and a dual-attention network to integrate the isolated knowledge. GMPI does not perform any measurement or impose any burden on the network. Our performance evaluation shows that our framework GMPI outperforms state-of-the-art approaches in terms of accuracy and coverage. In particular, compared to the state-of-the-art stitching-based baseline, GMPI provides a 42.45% improvement in coverage and a 39.97% improvement in accuracy. The experimental results demonstrate that GMPI can accurately infer paths for nearly arbitrary APs.

As a decentralized training paradigm, Federated learning (FL) promises data privacy by exchanging model parameters instead of raw local data. However, it is still impeded by the resource limitations of end devices and privacy risks from the ‘curious’ cloud. Yet, existing work predominately ignores that these two issues are non-orthogonal in nature. In this article, we propose a joint design (i.e., AHFL) that accommodates both the efficiency expectation and privacy protection of clients towards high inference accuracy. Based on a cloud-edge-end hierarchical FL framework, we carefully offload the training burden of devices to one proximate edge for enhanced efficiency and apply a two-level differential privacy mechanism for privacy protection. To resolve the conflicts of dynamical resource consumption and privacy risk accumulation, we formulate an optimization problem for choosing configurations under correlated learning parameters (e.g., iterations) and privacy control factors (e.g., noise intensity). An adaptive algorithmic solution is presented based on performance-oriented resource scheduling, budget-aware device selection, and adaptive local noise injection. Extensive evaluations are performed on three different data distribution cases of two real-world datasets, using both a networked prototype and large-scale simulations. Experimental results show that AHFL relieves the end's resource burden (w.r.t. computation time 8.58%$\downarrow$, communication time 59.35%$\downarrow$and memory consumption 43.61%$\downarrow$) and has better accuracy (6.34%$\uparrow$) than 3 typical baselines under the limited resource and privacy budgets. The code for our implementation is available athttps://github.com/Guoyeting/AHFL.

Photo capturing and sharing have become routine daily activities for social platform users. Alongside the entertainment of social interaction, we are experiencing tremendous visual violation and photo abusing. Especially, users may be unconsciously filmed and exposed online, which is termed as the non-consensual sharing issue. Unfortunately, this problem cannot be well handled with proactive access control or dedicated bystander detection, as users are unaware of their situations and may be filmed stealthily. We propose Videre on behalf of the privacy of the unaware parties in a way that they would be automatically identified and warned before such photos go public. For this, we first elaborate on the predominant features encountered in non-consensual captured photos via a thorough user study. Then we establish a dataset for this context and build a classifier as a proactive detector based on multi-deep-feature fusion. To relieve the burden of person-wise unawareness detection, we further design a signature-based filter for local pre-authorization, which can also implicitly avoid classification errors. We implement and test Videre in various field settings to demonstrate its effectiveness and performance.

The combination of Mobile Crowdsensing (MCS) and truth discovery has benefited the ubiquitous monitoring and analysis of the physical world. To address the concerns alongside user data collection, the literature has partially studied data privacy protection for truth discovery. Yet, the threats of location leakage remain overlooked in such contexts. For joint accommodating privacy protection and truth elaboration, we propose to leverage differential privacy for distributed user location obfuscation and explore spatial correlation for corresponding observation's value calibration. We form this process into a truth estimation deviation minimization problem under differential privacy and obfuscation requirements. By theoretically transforming it into probabilistic calibration residual optimization, the problem can be solved via linear programming. Evaluation on real-world temperature and humid sensing data shows its effectiveness on providing significant location distortion distance and practically acceptable time consumption. Results also reveal an up to 53% truth discovery accuracy improvement compared to the SCP baseline.

IPv6 target generation is the critical step for fast IPv6 scanning for Internet-wide surveys. Existing techniques, however, commonly suffer from low hit rates due to inappropriate space partition caused by the outlier addresses and short-sighted splitting indicators. To address the problem, we propose 6Forest, an ensemble learning-based approach for IPv6 target generation that is from a global perspective and resilient to outlier addresses. Given a set of known addresses, 6Forest first considers it as an initial address region and then iteratively divides the IPv6 address space into smaller regions using a maximum-covering splitting indicator. Before a round of space partition, it builds a forest structure for each region and exploits an enhanced isolation forest algorithm to remove the outlier addresses. Finally, it pre-scans samples from the divided address regions and based on the results generates IPv6 addresses. Experiments on eight large-scale candidate datasets indicate that, compared with the state-of-the-art methods in IPv6 worldwide scanning, 6Forest can achieve up to 116.5% improvement for low-budget scanning and 15× improvement for high-budget scanning.

Speech recognition systems, trained and updated based on large-scale audio data, are vulnerable to backdoor attacks that inject dedicated triggers in system training. The used triggers are generally human-inaudible audio, such as ultrasonic waves. However, we note that such a design is not feasible, as it can be easily filtered out via pre-processing. In this work, we propose the first audible backdoor attack paradigm for speech recognition, characterized by passively triggering and opportunistically invoking. Traditional device-synthetic triggers are replaced with ambient noise in daily scenarios. For adapting triggers to the application dynamics of speech interaction, we exploit the observed knowledge inherited from the context to a trained model and accommodate the injection and poisoning with certainty-based trigger selection, performance-oblivious sample binding, and trigger late-augmentation. Experiments on two datasets under various environments evaluate the proposal's effectiveness in maintaining a high benign rate and facilitating outstanding attack success rate (99.27%, ~4% higher than BadNets), robustness (bounded infectious triggers), feasibility in real-world scenarios. It requires less than 1% data to be poisoned and is demonstrated to be able to resist typical speech enhancement techniques and general countermeasures (e.g., dedicated fine-tuning). The code and data will be made available at https://github.com/lqsunshine/DABA.

The wide sharing and rapid dissemination of digital artworks has aggravated the issues of plagiarism, raising significant concerns in cultural preservation and copyright protection. Yet, modes of plagiarism are formally uncharted, causing rough plagiarism detection practices with duplicate checking. This work is thus devoted to understanding artwork plagiarism, with poster design as the running case, for building more dedicated detection techniques. As the first study of such, we elaborate on 8 elements that form unique posters and 6 judgement criteria for plagiarism using an exploratory study with designers. Second, we build a novel poster dataset with plagiarism annotations according to the criteria. Third, we propose models, leveraging the combination of primary elements and criteria of plagiarism, to find suspect instances in a retrieval process. The models are trained under the context of modern artwork and evaluated on the poster plagiarism dataset. The proposal is shown to outperform the baseline with superior Top-K accuracy (~33%) and retrieval performance (~42%).

The privacy-preserving nature of Federated Learning (FL) exposes such a distributed learning paradigm to the planting of backdoors with locally corrupted data. We discover that FL backdoors, under a new on-off multi-shot attack form, are essentially stealthy against existing defenses that are built on model statistics and spectral analysis. First-hand observations of such attacks show that the backdoored models are indistinguishable from normal ones w.r.t. both low-level and high-level representations. We thus emphasize that a critical redemption, if not the only, for the tricky stealthiness is reactive tracing and posterior mitigation. A three-step remedy framework is then proposed by exploring the temporal and inferential correlations of models on a trapped sample from an attack. In particular, we use shift ensemble detection and co-occurrence analysis for adversary identification, and repair the model via malicious ingredients removal under theoretical error guarantee. Extensive experiments on various backdoor settings demonstrate that our framework can achieve accuracy on attack round identification of ∼80% and on attackers of ∼50%, which are ∼28.76% better than existing proactive defenses. Meanwhile, it can successfully eliminate the influence of backdoors with only a 5%∼6% performance drop.

Smart cameras, as typical IoT devices, are widely adopted to provide surveillance on individuals, homes, and the environment. The unavoidably captured sensitive visuals via these cameras may raise significant security concerns, while the prevalent software defects and authentication misconfiguration issues aggravate the vulnerability of such devices. However, traditional cryptography techniques are inadequate to provide full protection of these devices due to the large computation overhead. In this context, realizing accountability for these surveillance systems shall be the last line of defense in the presence of fast-evolving and high-influential threats. We propose EviChain, a scalable blockchain-based solution to trace the operations on intelligent surveillance cameras and reserve the evidence for any misuse in tamper-proofing manipulation records. Building a blockchain over the distributed cameras is challenging due to the limited capacity of on-board memory. To tackle this challenge, we design a cooperative mechanism that enables cameras to adaptively join in groups and share storage for recording blocks. In addition, we present a computation efficiency and delay-aware block generation strategy to reduce the cost of the consensus process. We perform extensive simulations to validate the superior performance of EviChain over other baselines, for example, Practical Byzantine Fault Tolerance (PBFT).

A key performance indicator (KPI) consists of critical time series data that reflect the runtime states of network systems (e.g., response time and available bandwidth). Despite the importance of KPI, datasets for KPI anomaly detection available to the public are very limited, due to privacy concerns and the high overhead in manually labelling the data. The insufficiency of public KPI data poses a great barrier for network researchers and practitioners to evaluate and test what-if scenarios in the development of artificial intelligence for IT operations (AIOps) and anomaly detection algorithms. To tackle the difficulty, we develop a univariate time series generation tool called TSAGen, which can generate KPI data with anomalies and controllable characteristics for KPI anomaly detection. Experiment results show that the data generated by TSAGen can be used for comprehensive evaluation of anomaly detection algorithms with diverse user-defined what-if scenarios.

The emerging federated learning (FL) paradigm allows multiple distributed devices to cooperatively train models in parallel with the raw data retained locally. The local-computed parameters will be transferred to a centralized server for aggregation. However, the vanilla aggregation method ignores the heterogeneity of the distributed data, which may lead to slow convergence and low training efficiency. Yet, existing data scheduling and improved aggregation methods either incur privacy concerns or fail to consider the fine-grained heterogeneity. We propose FedCav, a contribution-aware model aggregation algorithm that differentiates the merit of local updates and explicitly favors the model-informed contributions. The intuition is that the local data showing higher inference loss is likely to facilitate better performance improvement. To this end, we design a novel global loss function with explicit optimization preference on informative local updates, theoretically prove its convex property, and use it to regulate the gradient descent process iteratively. Additionally, we propose to identify abnormal updates with fake loss by auditing historic local training statistics. The results of extensive experiments demonstrate that FedCav needs fewer training rounds (~34%) for convergence and achieves better inference accuracy (~2.4%) than the baselines (i.e., FedAvg and FedProx). We also observe that FedCav can actively mitigate the model replacement attacks with agile recovery capability towards the aggregation.

The rapid development of the Internet of Things (IoT) has promoted the wide adoption of mobile medical devices, which monitor patients’ body conditions in real-time. The collected health-related data are highly sensitive, requiring careful protection during the accessing and transmission for specialized analysis. Yet, existing efforts either rely on centralized authentication for the numerous end devices or are designed to be partially distributed for a closed institution, both lacking scalability for mobile healthcare scenarios. In this paper, we propose HealthTrust that provides a generalized and flexible authentication scheme for distributed medical devices in the cross-institution context. With blockchain as the building block, HealthTrust jointly exploits smart contracts and secure authentication to attain controlled transmission and secure exchanging of healthcare data between institutions. We have implemented the system functions with a prototype based on Ethereum. Experimental results and safety analysis demonstrate that HealthTrust can well satisfy both the safety and feasibility requirements.

Adaptive bitrate (ABR) algorithms are often used to optimize the quality of user experience (QoE) during video playback. In the client-side video player, the buffer size and predicted throughput are mainly used to improve user's QoE. However, due to the randomness of mobile network traffic and the heavy-tail effect of the network, it is very difficult to predict throughput. We innovatively use Bayesian neural network to dynamically evaluate video signals. Unlike previous neural network solutions, we use probability distributions instead of point estimates to predict throughput, which can effectively evaluate QoE metrics. Our contributions are to first (i) use of Bayesian neural network to guide video adaptive bitrate adaptation, and then (ii) propose a bitrate adaptive algorithm denoted 2prong, which utilizes high-dimensional contextual information such as buffer occupancy, predicted throughput and video quality to find the most valuable information for quality adaption in real time. We demonstrate the effectiveness of the 2prong algorithm using a simulation testbed. By comparing with other methods, it is demonstrated that 2prong can improve the video quality of video streaming transmission.

With the development of edge computing, edge storage solutions are attracting widespread attention. When facing the requirements of lower latency and faster access speed from end devices, edge storage solutions are considered to be an alternative to the cloud. However, edges are usually owned by small organizations which have limited operations and maintenance capabilities. This makes these edge devices can be easily disabled by external attacks or internal hardware failures. Besides, the heterogeneity of the edge devices will also make it difficult to price the edge resources uniformly. To tackle these problems, we propose SmartStore: an auction mechanism based on blockchain to allocate edge resources. Considering centralized solutions have access bottlenecks and trust issues, we built SmartStore on the smart contract. With Bayesian game theory, SmartStore can analyze how data owners (DO) and edges price the resources can maximize their benefits. From an economic perspective, both DO and edges can make full use of edge heterogeneous resources with SmartStore. Besides, a two-stage submission strategy is proposed to complete the sealed auction. Furthermore, considering the reliability of edge storage, we propose a cluster-based block distribution algorithm for SmartStore's intelligent edge recommendation process. SmartStore ensures the reliability of edge storage while maximizing the benefits and resource utilization of both parties. Finally, we conduct specific experiments on the proposed auction smart contract through “Ethereum” and the experimental results of implementation show the effectiveness and efficiency of our SmartStore.

Edge computing emerges as an alternative to cloud computing in the scenarios where the end devices require lower latency and faster access speeds. Edge nodes are deployed at the proximity of the end devices to reduce response time. On the other hand, the edge nodes are usually owned by small organizations that have limited operations and maintenance capabilities. Data on the edge may be easily damaged, due to external attacks or internal hardware failures. Therefore, it is essential to verify data integrity in edge computing. However, edge environment requires a different trust model compared with other computing and storage paradigm. Besides, compared with cloud storage, edge storage is decentralized and storage service participants may pose greater internal and external threats. This paper proposes a blockchain-based intelligent crowdsourcing audit approach (Crowdauditing) to achieve on-chain and off-chain credibility of audit results. The model relies on an untrusted auditor committee from the crowd to audit data integrity and uses smart contracts as the core of the intelligent system to ensure the reliability of result submission, the accuracy of the result judgment, and reasonable punishments and rewards. Specifically, an unbiased selection algorithm is proposed to achieve fairness during the auditor committee construction. An innovative two-stage submission strategy is proposed to ensure that the auditor committee can reach a consensus on the off-chain audit results. An incentive mechanism is carefully designed to force auditors providing audit services honestly to maximize their own rewards. Moreover, we modeled that as a game of n players, which proves the reliability of the result. Finally, we implement a prototype of Crowdauditing based on smart contracts. The extensive experimental results demonstrate the effectiveness of Crowdauditing.

With the explosive emergence of computation-intensive and latency-sensitive applications, data processing could be envisioned to perform closer to the data source. Similar to edge and fog computing, dispersed computing is considered as a complementary computing paradigm, which can excavate potential computation resources in the network to users, and serve as a supplement for sharing the computational burden when the edge is overloaded. In this article, we first make full use of idle and geographically dispersed computation resources via task offloading, contributing to conserve energy for mobile devices. Especially, a dispersed computing offloading framework concerning the interests of users and networked computation points is proposed. We further transform the initial problem into a multiobjective optimization problem subject to latency and resource constraints. To tackle such a complex problem, an energy-saving bilateral matching algorithm is designed to obtain the optimal task offloading strategy. The simulation results demonstrate that our proposed algorithm can outperform the benchmark schemes in terms of user fairness and can achieve a relatively balanced energy cost ratio. Furthermore, comparative experiments with edge computing are implemented in Amber Response and Disaster Relief scenarios, respectively, to reveal the advantages of the proposed framework.

Surveillance cameras have been extensively used in smart cities and high security zones. However, with the exploding deployment of smart cameras, the rapid growth of cloud workloads from vision-based IoT applications are becoming a huge burden for all cloud service providers. Some researchers have proposed mechanisms, such as compression and deduplication to reduce the video traffic size, but these methods cannot offset the enormous growth of data volume. Most of the surveillance video data do not need to be proceeded in real time. By making use of the IoT camera’s onboard resources to store the data, the cloud workloads can be fundamentally reduced. However, recent incidents have posed a new, powerful geo-range attack, where the attacker may compromise a group of surveillance cameras located within an area. Existing simple onboard solutions cannot offer secure defense against such geo-range attacks. To tackle the problem, we developCentipede, a cooperative video data storage system that distributes video content across geographically dispersed surveillance cameras. It generates secure copies for the video content and enhances data security by judiciously distributing erasure-coded video blocks across optimally-chosen surveillance cameras. In this article, we implementCentipedeand evaluate its performance.Centipedeis the first solution that can fundamentally reduce the cloud workload and defend against geo-range attacks.

Existing mobile photo crowdsensing approaches focus on the participant-to-server photo pre-selection, i.e., reducing the photo redundancy from participants to a server. The server may still receive plenty of photos for a target area. Yet, another important problem is to select a proper photo subset of an area from the server to a requester. This is a challenging problem because the selected subset with a small size should attain both coverage on the PoIs - Points of Interest (i.e., photo coverage of the area) and quality on the views (i.e., view quality). In this paper, we propose a novel and generic server-to-requester photo selection approach even when there are neither photo shooting direction information nor reference photos. A utility model is designed to measure photo merits of coverage and quality by exploiting photos' spatial distribution and visual representativeness. We present two photo selection schemes, basic and PoI number-aware, to maximize the photo selection utility with multiple levels of granularity. Experimental results on real-world datasets show that our basic scheme outperforms the baselines by an average of 33% and 18.7% on photo coverage and view quality, respectively. Our PoI number-aware scheme can yield an additionally 44.8 percent improvement on the photo coverage performance.

Network anomalies, such as wide-area congestion and packet loss, can seriously degrade network performance. To this end, it is critical to accurately identify network anomalies on end-to-end paths for high quality network services in practice. In this work, we propose an unsupervised two-step method for the detection and characterization of general network anomalies. It first finds the change-points in large-scale RTT time series by formalizing an optimization problem in terms of data series segmentation. Then we mark the segments as normal or abnormal on different sides of a change-point through exploitation of their distribution statistics. After detecting an anomaly, a further step is introduced to analyze the relations between links with state changes and localize the entities (nodes or links) that most likely cause the corresponding event. We believe such unsupervised and light-weighed method can provide valuable insights on anomaly mining in large-scale time series data. Extensive experiments on both simulated (artificial time series with ground truth) and real-network (RIPE Atlas traceroute measurements) datasets are performed. The results demonstrate that the proposed method can achieve better performance, w.r.t. accuracy and efficiency, than existing solutions.

City-wide package delivery has become popular due to the dramatic rise of online shopping. It places a tremendous burden on the traditional logistics industry, which relies on dedicated couriers and is labor-intensive. Leveraging the ridesharing systems is a promising alternative, yet existing solutions are limited to one-hop ridesharing or need consignment warehouses as relays. In this paper, we propose a new package delivery scheme which takes advantage of multi-hop ridesharing and is entirely consignment free. Specifically, a package is assigned to a taxi which is guided to deliver the package all along to its destination while transporting successive passengers. We tackle it with a two-phase solution, named PPtaxi. In the first phase, we use the Multivariate Gaussian distribution and Bayesian inference to predict the passenger orders. In the second phase, both the computation efficiency and solution effectiveness are considered to plan package delivery routes. We evaluate PPtaxi with a real-world dataset from an online taxi-taking platform and compare it with multiple benchmarks. The results show that the successful delivery rate of packages with our solution can reach 95 percent on average during the daytime, and is at most 46.9 percent higher than those of the benchmarks.

Driven by the visions of Internet of Things and 5G communications, the edge computing systems integrate computing, storage, and network resources at the edge of the network to provide computing infrastructure, enabling developers to quickly develop and deploy edge applications. At present, the edge computing systems have received widespread attention in both industry and academia. To explore new research opportunities and assist users in selecting suitable edge computing systems for specific applications, this survey paper provides a comprehensive overview of the existing edge computing systems and introduces representative projects. A comparison of open-source tools is presented according to their applicability. Finally, we highlight energy efficiency and deep learning optimization of edge computing systems. Open issues for analyzing and designing an edge computing system are also studied in this paper.

Traditional mobile crowdsensing photo selection process focuses on selecting photos from participants to a server. The server may contain tons of photos for a certain area. A new problem is how to select a set of photos from the server to a smartphone user when the user requests to view an area (e.g., a hot spot). The challenge of the new problem is that the photo set should attain both photo coverage and view quality (e.g., with clear Points of Interest). However, contributions of these geo-tagged photos could be uncertain for a target area due to unavailable information of photo shooting direction and no reference photos. In this paper, we propose a novel and generic server-to-requester photo selection approach. Our approach leverages a utility measure to quantify the contribution of a photo set, where photos' spatial distribution and visual correlation are jointly exploited to evaluate their performance on photo coverage and view quality. Finding the photo set with the maximum utility is proven to be NP-hard. We then propose an approximation algorithm based on a greedy strategy with rigorous theoretical analysis. The effectiveness of our approach is demonstrated with real-world datasets. The results show that the proposal outperforms other approaches with much higher photo coverage and better view quality.

WiFi networks are vulnerable to rogue AP attacks in which an attacker sets up an imposter AP to lure mobile users to connect. The attacker can eavesdrop on the communication, severely threatening users' privacy. Existing rogue AP detection solutions are confined to some specific attack scenarios (e.g., by relaying the traffic to a target AP) or require additional hardware. In this paper, we propose a crowdsensing based approach, named CRAD, to detect rogue APs in camouflage without specialized hardware requirement. CRAD exploits the spatial correlation of RSS to identify a potential imposter, which should be at a different location from the legitimate one. The RSS measurements collected from the crowd facilitate a robust profile and minimize the inaccuracy effect of a single RSS value. As a result, CRAD can filter out abnormal samples sensed in the realtime by dynamically matching the profile. We evaluate our approach with both a public dataset and a real prototype. The results show that CRAD can yield 90% detection accuracy and precision with proper crowd presence, even when the rogue AP is launched close to the legitimate one (e.g., within 1m).

Mobile crowdsensing (MCS) is a new paradigm which takes advantage of pervasive mobile devices to collaboratively collect data and analyze physical phenomenon. As mobile devices are owned and controlled by individuals with various capabilities and intentions, a main challenge MCS applications face is to ensure the credibility of the crowd contributed data. Existed works attempt to increase confidence level of the sensory measurements by validating the location. However, the required infrastructure or neighbor support may not always be available, and the unreliable form containing false sensory data with a valid location is implicitly ignored. In this paper, we propose a novel Crowd-based Credibility Improving Scheme (CCIS) to improve the credibility of data in possible false forms leveraging crowd data property and crowd participants' reputation. Based on the data clusters generated using a lightweight fixed-width clustering algorithm, CCIS is able to adequately identify and filter out the clusters constituted mainly by false data using reputation information as the classifier. We conduct simulations on a publicly available trace with crowd contributed temperature measurements, the results show that CCIS yields an improvement of overall data credibility of around 1.2 with clustering accuracy over 96%.