SDN data plane generally utilizes TCAM to accommodate flow tables for fast packet classification, which also leads to serious problem of high energy consumption. Existing techniques are difficult to stably achieve satisfactory energy-saving effect especially under network traffic jitters. To address this problem, this paper first designs an elastic energy-saving cache to always keep sufficient number of active exact flows for stably high cache hit rates even under network traffic jitters. Particularly, we adaptively adjust the number of cache segments, in terms of the relationship between current cache hit rate and its preset expected range, to maintain high cache hit rate. Meanwhile, we regulate the threshold of packet inter-arrival time for identifying active exact flows, in accordance with current cache occupancy rate, to match the number of active exact flows with cache capacity. Furthermore, we theoretically derive cache occupancy rate based on randomly mapping assumption of each active exact flow and cache hit rate based on the assumption of flow activity degree model. Subsequently, we build an elastically energy-saving flow table storage architecture, by applying the elastic energy-saving cache to always enable a majority of incoming packets to bypass energy-hungry TCAM flow table lookups. Finally, we set up an experimental SDN platform to evaluate its performance on real network traffic traces. Experimental results indicate that our built flow table storage architecture steadily achieves high energy saving rates around 82.43% even under network traffic jitters, with the increase of 6.32˜7.55% compared to the state-of-the-art one.

As a popular service management system, the spatial crowdsourcing (SC) server is responsible for allocating nearby workers to perform tasks based on outsourced locations. However, protecting the sensitive information contained in these outsourced locations is crucial. Traditional differential privacy (DP) methods suffer from two limitations: 1) they usually rely on a trusted third party, failing to protect both worker and task location privacy simultaneously, thus risking privacy breaches; 2) they ignore the personalized privacy demands of different users. In this paper, we propose a personalized local DP-based location obfuscation (PLDPLO) scheme, thereby providing personalized privacy-preserving both worker and task locations locally while allocating high-quality tasks. To achieve this, we introduce a personalized location indistinguishability (PLI) model, a new personalized Laplace mechanism achieving local DP, to jointly provide the protection of worker locations and different privacy levels for different workers. To address task privacy, we present a spatial mapping indistinguishability (SMI) algorithm to obfuscate task locations based on a random response mechanism, thereby ensuring data utility. Additionally, we propose a Zipf-Poisson model-based task allocation graph (ZPTAG) algorithm to perform one-task-multiple-workers allocation and achieve a high competitive ratio, which reduces the move distance of workers. Our PLDPLO scheme guarantees ϵ-LDP. Extensive experiments over real datasets demonstrate that our scheme achieves over 89% data utility for task allocation and outperforms state-of-the-art methods while providing personalized privacy levels.

Heterogeneous software systems, which often combine closed-source libraries with exported interfaces, embedded assembly, and components in multiple languages, present significant challenges for formal verification. Our prior work, Hetrify, addressed this by converting RISC-V binaries into semantically equivalent C code, making such programs amenable to verification. However, its unified memory model required frequent dynamic computation of stack addresses, which significantly increased the size of the generated logical formulas, along with high memory usage and longer verification times. To address this, we propose memory access specialization, a static analysis and transformation technique that recovers fixed stack offsets during binary conversion to reduce verification overhead. By replacing symbolic stack accesses with fixed-offset memory references, it eliminates dynamic pointer arithmetic and reduces symbolic encoding complexity. This technique is integrated into Hetrify+, an enhanced verification tool for heterogeneous programs. To validate the effectiveness of our approach, we conduct both formal analysis and extensive empirical evaluation. Formal analysis guarantees the correctness of our method. In our evaluation, Hetrify+ demonstrates the same verification accuracy as the original Hetrify on 100 low-level RISC-V assembly programs, achieving up to 2.5× speedup and 4.9× reduction in memory usage. For 30 large-scale heterogeneous programs that include binary-only components, Hetrify+ maintains a 100% success rate, reducing verification time by 1.9× and memory consumption by 1.2×. These results demonstrate that memory access specialization is key to scaling the verification of heterogeneous programs.

Abstract Merge sort as a divide-sort-merge paradigm has been widely applied in computer science fields. As modern reduced instruction set computing architectures like the fifth generation (RISC-V) regard multiple registers as a vector register group for wide instruction parallelism, optimizing merge sort with this vectorized property is becoming increasingly common. In this paper, we overhaul the divide-sort-merge paradigm, from its register-level sort to the cache-aware merge, to develop a fine-grained RISC-V vectorized merge sort (RVMS). From the register-level view, the inline vectorized transpose instruction is missed in RISC-V, so implementing it efficiently is non-trivial. Besides, the vectorized comparisons do not always work well in the merging networks. Both issues primarily stem from the expensive data shuffle instruction. To bypass it, RVMS strides to take register data as the proxy of data shuffle to accelerate the transpose operation, and meanwhile replaces vectorized comparisons with scalar cousin for more light real value swap. On the other hand, as cache-aware merge makes larger data merge in the cache, most merge schemes have two drawbacks: the in-cache merge usually has low cache utilization, while the out-of-cache merging network remains an ineffectively symmetric structure. To this end, we propose the half-merge scheme to employ the auxiliary space of in-place merge to halve the footprint of naïve merge sort, and meanwhile copy one sequence to this space to avoid the former data exchange. Furthermore, an asymmetric merging network is developed to adapt to two different input sizes. Experiments on the RISC-V processor SG2042 show that four fine-grained optimization schemes including register strided transpose, hybrid merging network, half-merge strategy, and asymmetric merging network, improve performance by 4.05%, 19.88%, 12.23%, and 11.04% respectively. Importantly, the overall performance is 1.34x faster than the parallel sorting in the Boost C++ library, and 1.85x faster than std::sort.

Abstract The machine learning as a service (MLaaS) paradigm has been widely adopted across various applications. However, it also raises significant privacy concerns, particularly regarding the exposure of input data and trained models. Two-party computation in convolutional neural network (CNN) inference has emerged as a promising solution to address these privacy issues in MLaaS. Nevertheless, most existing privacy-preserving CNN architectures rely on computationally expensive encryption methods, resulting in prolonged inference times and increased communication overhead. In this paper, we propose F2PQNN, a fast and secure two-party inference framework for quantized CNNs. To minimize reliance on computationally intensive encryption, F2PQNN utilizes two non-colluding servers and integrates secret sharing with oblivious transfer techniques. Furthermore, F2PQNN incorporates quantization techniques, along with batching and asynchronous computation, to significantly accelerate inference predictions. We evaluate the performance of F2PQNN on the MNIST, Fashion-MNIST, CIFAR-10, and STL-10 datasets. Experimental results demonstrate that F2PQNN outperforms existing solutions, achieving a $9.14\times $ speedup and reducing communication overhead by $59.8\times $ on the MNIST dataset.

RGBT tracking is a task that utilizes images from visible (RGB) and thermal infrared (TIR) modalities to continuously locate a target, which plays an important role in various fields including intelligent transportation systems. Most existing RGBT trackers do not achieve high precision and real-time tracking speed simultaneously. To address this challenge, we propose an innovative RGBT tracker, the Siamese Triple-stream Feature Aggregation Network (SiamTFA). Firstly, a triple-stream backbone is presented to implement multi-modal feature extraction and fusion, which contains two parallel Swin Transformer feature extraction streams, and one feature fusion stream composed of joint-complementary feature aggregation (JCFA) modules. Secondly, our proposed JCFA module utilizes a joint-complementary attention to guide the aggregation of multi-modal features. Specifically, the joint attention can focus on spatial location information and semantic information of the target by combining the features of two modalities. Considering the complementarity between RGB and TIR modalities, the complementary attention is introduced to enhance the information of beneficial modality and suppress the information of ineffective modality. Thirdly, in order to reduce the computational complexity of the joint-complementary attention, we propose a depthwise shared attention structure, which utilizes depthwise convolution and shared features to achieve lightweight attention. Finally, we conduct extensive experiments on four official RGBT test datasets and the experimental results demonstrate that our proposed tracker outperforms some state-of-the-art trackers and the tracking speed reaches 37 frames per second (FPS). The code is available athttps://github.com/zjjqinyu/SiamTFA.

In the paradigm of Software-Defined Networking (SDN), its data plane generally perform packet forwarding based on flow table lookup on TCAM with high energy consumption. Popular energy-saving methods employ caching techniques for most packets to bypass energy-intensive TCAM lookups. However, existing energy-saving caches cannot adapt to network traffic fluctuation with sufficient utilization of cache space due to non-negligible hash conflicts. To overcome this issue, we design a traffic-aware energy-saving cache with optimal utilization for TCAM flow tables in SDN data plane. In particular, we first devise a nearly conflict-free hashing algorithm for the cache called FelisCatus, which provides three candidate locations for each incoming flow by adjacent hopping, and searches for an empty or replaceable entry for each conflicting flow by co-directional kicking. Then, we propose an adaptive adjustment mechanism of flow activity criterion, i.e., packet inter-arrival time threshold, for enabling the cache to consistently accommodate the most active exact flows in network traffic. Furthermore, we build an energy-efficient SDN flow table storage architecture by applying the above cache and exploiting the accessing features of different memories. Finally, we verify the performance of our designed energy-saving cache and flow table storage architecture by experiments with backbone network traffic traces. Experimental results indicate that, our designed energy-saving cache obtains stable and high hit rates around 75% even under network traffic fluctuation, and our proposed flow table storage architecture achieve high energy saving rates around 71%, with the increase of 7.89% compared to state-of-the-art ones.

Objective assessment of taste sensation is essential for medical diagnosis, food development, and multisensory interaction. Human taste sensation can be characterized through biosignals such as electroencephalography (EEG) and electromyography (EMG). However, taste sensation recognition on multiple-subject datasets remains challenging due to the low signal-to-noise ratio and substantial individual variability of biosignals. To address these problems, we propose SwinTaste for accurate and generalized taste sensation recognition from bimodal biosignals. The Transformer is introduced to extract hierarchical features. A two-stage patch partition module is optimized for the characteristics of biosignals. Moreover, a multi-task learning strategy is adopted to improve the generalization and subject adaptation abilities. The SwinTaste is evaluated on a multiple-subject taste sensation dataset. Comparison experiments and ablation studies demonstrate the superior performance of SwinTaste, indicating the potential for generalized application in biosignal recognition.

Internet of Things (IoT) devices generate vast amounts of real-time data across diverse sectors, offering lucrative opportunities in the data trading market. This facilitates the conversion of raw data into valuable products and services, resulting in significant economic and social benefits. To issue data security problems in trading mechanisms, several solutions based on local differential privacy (LDP) provide lightweight methods for privacy preservation and efficient data exchange. However, most solutions lack effective mechanisms for longitudinal data. Moreover, LDP needs to address the data quality problem in IoT. At last, the pricing of privacy-preserving longitudinal data remains unresolved. To address these problems, we propose a privacy-preserving data trading (PPDT) scheme for IoT in this article. Specifically, to guarantee the security of longitudinal data, we utilize two perturbation techniques to accommodate data owners with varying privacy preferences. To enhance data availability, we devise a binary tree-based aggregation algorithm combined with a weighted average strategy and maximum likelihood estimation. Additionally, we derive an optimal contract that considers different levels of privacy preservation and data trading prices. In scenarios with incomplete information, the contract can provide appropriate incentives to the involved parties. Finally, we demonstrate the efficiency and effectiveness of the proposed data trading scheme through theoretical analysis and extensive experiments.

Abstract Recommendation systems have been effectively utilized in various fields, but their internal decision-making methods are still largely unknown. This opaque decision-making method can greatly affect users’ trust in the recommendation system. Therefore, finding a way to explain the reasons for model decisions has become an urgent task. Previous studies often used LSTM and other models to generate recommendation explanations and explain the reasons for recommendations in text form. However, traditional methods cannot effectively use the ID information of users and items, and the text generated is highly repetitive. To solve this problem, this paper uses the method of prompt learning combined with a graph encoder to design a recommendation explanation generation model. In order to narrow the semantic gap between the ID information of users and items and natural language and capture high-level interaction information, this paper designs a graph encoder based on user similarity to learn the interactive semantic information of user and item IDs, and to construct a continuous prompt. Then, the discrete prompt composed of discrete features of users and items is combined with the continuous prompt to construct a hybrid prompt to input into the pre-trained model to generate the recommended explanation. This paper experiments on three publicly available datasets and compares them with several state-of-the-art methods to demonstrate the personalization and text quality of the generated explanations.

The Transformer has been successfully used in medical image segmentation due to its excellent long-range modeling capabilities. However, patch segmentation is necessary when building a Transformer class model. This process ignores the tissue structure features within patch, resulting in the loss of shallow representation information. In this study, we propose a Heterogeneous Swin Transformer with Multi-Receptive Field (HST-MRF) model that fuses patch information from different receptive fields to solve the problem of loss of feature information caused by patch segmentation. The heterogeneous Swin Transformer (HST) is the core module, which achieves the interaction of multi-receptive field patch information through heterogeneous attention and passes it to the next stage for progressive learning, thus complementing the patch structure information. We also designed a two-stage fusion module, multimodal bilinear pooling (MBP), to assist HST in further fusing multi-receptive field information and combining low-level and high-level semantic information for accurate localization of lesion regions. In addition, we developed adaptive patch embedding (APE) and soft channel attention (SCA) modules to retain more valuable information when acquiring patch embedding and filtering channel features, respectively, thereby improving model segmentation quality. We evaluated HST-MRF on multiple datasets for polyp, skin lesion and breast ultrasound segmentation tasks. Experimental results show that our proposed method outperforms state-of-the-art models and can achieve superior performance. Furthermore, we verified the effectiveness of each module and the benefits of multi-receptive field segmentation in reducing the loss of structural information through ablation experiments and qualitative analysis.

In modern datacenter networks (DCNs), mainstream congestion control (CC) mechanisms essentially rely on Explicit Congestion Notification (ECN) to reflect congestion. The traditional static ECN threshold performs poorly under dynamic scenarios, and setting a proper ECN threshold under various traffic patterns is challenging and time-consuming. The recently proposed reinforcement learning (RL) based ECN Tuning algorithm (ACC) consumes a large number of computational resources, making it difficult to deploy on switches. In this paper, we present a lightweight and hierarchical automated ECN tuning algorithm called LAECN, which can fully exploit the performance benefits of deep reinforcement learning with ultra-low overhead. The simulation results show that LAECN improves performance significantly by reducing latency and increasing throughput in stable network conditions, and also shows consistent high performance in small flows network environments. For example, LAECN effectively improves throughput by up to 47%, 34%, 32% and 24% over DCQCN, TIMELY, HPCC and ACC, respectively.

Taste sensation can be objectively measured using electroencephalography (EEG) or electromyography (EMG). How-ever, it is still challenging to effectively utilize the complementary information from EEG and EMG signals in taste sensation recognition. This paper proposes a bimodal fusion network (Bi-FusionNet) for recognizing basic taste sensations (sour, sweet, bitter, salty, umami, and blank). Two convolutional backbones with similar structures are designed to separately extract the single-modal features of EEG and EMG. Then, EEG and EMG features are concatenated for bimodal interaction and complementarity. Finally, three loss functions are adopted: a center loss for aggregating intra-class samples, a mean squared error loss for sequence positions for minimizing the difference between signals during the stimulation, and a softmax loss for minimizing the entropy of prediction and true labels. The results on the taste sensation dataset show that bimodal fusion improves recognition performance, and Bi-FusionNet outperforms single-modal methods and other fusion methods. Bi-FusionNet paves the way for the application of multimodal fusion in taste sensation recognition.

Abstract Deep learning approaches to interactive image segmentation are typically formulated as a binary labeling problem. A model trained to make predictions within a fixed set of labels (i.e., foreground and background labels) cannot be used to directly predict the binary masks of multiple objects of interest, which greatly limits its flexibility and adaptivity. The use of different classes of clicks as input is opted for and the first end‐to‐end learning model for multi‐object segmentation, based on a new designed neural network, is developed. The network consists of a visual feature extractor, a recurrent attention module and a dynamic segmentation head, extracts user click‐adapted appearance embedding features and spatial attention features, and then learns to transform this information into a segmentation of multiple objects. It is also proposed to train the network using a joint loss function, taking the embedding learning into account for segmentation. Comprehensive experiments are conducted on three benchmark datasets to demonstrate the effectiveness of the proposed method. It performs favorably against state‐of‐the‐art approaches on the multiple object segmentation task, for example, with 0.15 s per image, 0.06 s per object and mean IoU & F1 score of 84.90% on Pascal VOC 2012 validation set. It is further shown that the method can be used in numerous vision applications such as image recoloring and colorization.

Spatial crowdsourcing (SC) is a location-based outsourcing service whereby SC-server allocates tasks to workers with mobile devices according to the locations outsourced by requesters and workers. Since location information contains individual privacy, the locations should be protected before being submitted to untrusted SC-server. However, the encryption schemes limit data availability, and existing differential privacy (DP) methods do not protect the tasks’ location privacy. In this paper, we propose a differential privacy-based location protection (DPLP) scheme, which protects the location privacy of both workers and tasks, and achieves task allocation with high data utility. Specifically, DPLP splits the exact locations of both workers and tasks into noisy multi-level grids by using adaptive three-level grid decomposition (ATGD) algorithm and DP-based adaptive complete pyramid grid (DPACPG) algorithm, respectively, thereby considering the grid granularity and location privacy. Furthermore, DPLP adopts an optimal greedy algorithm to calculate a geocast region around the task grid, which achieves the trade-off between acceptance rate and system overhead. Detailed privacy analysis demonstrates that our DPLP scheme satisfies$\epsilon$-differential privacy. The extensive analysis and experiments over two real-world datasets confirm high efficiency and data utility of our scheme.

Trending topic recommendation (TTR) has become a popular social service for social users to obtain interesting topics based on public social content. Due to sensitive privacy, the traditional differential privacy (DP) methods are used to protect social contents. However, these DP methods rely on a fully trusted third party (TTP) without considering protecting the correlations of social keywords, and do not support the privacy-preserving online social content publication. In this article, we propose a novel local DP-based TTR (LDPTR) scheme to perform high-quality online TTR services while achieving local privacy-preserving social contents. Specifically, LDPTR first clusters the social keywords with high correlations into noisy graph classes based on a graph-based LDP (GLDP) algorithm for ensuring the keyword correlation privacy. Second, LDPTR adopts a novel mechanism called ∈2-compressive sensing indistinguishability (CSI) to generate noisy social topics, thereby preventing the user-linkage attack and breaking the curse of high-dimensional local differential privacy (LDP). Then, a dynamic graph-based CSI (DGCSI) algorithm is proposed to protect the online social content privacy while ensuring data usability. Furthermore, LDPTR calculates the trending topics of interest with high burstiness by using a topic distribution similarity model-based topic burstiness (TMTB) algorithm, which achieves effective TTR services. Our LDPTR scheme satisfies ∈-LDP by detailed security analysis. Extensive experiments over two real-world data sets show that the proposed LDPTR gets high data utility while ensuring high-level privacy.

A new method to generate color texture images is proposed in this paper, which derived from our previous works to generate the gray texture. The method is based on the olfactory system bionic model to generate the gray texture. The model architecture mimics that of mammal olfactory neural system. Period function is used as the activity function of nodes in the model to realize the periodic repetition of texture. Chaotic mapping is used to adjust the model parameters to assure the model being in non-convergence state. The previous input is introduced as the noise to simulate the background noise of neural system. One color image is used as seed image. In HSV space, the Hue (H), Saturation (S) and Value (V) of each pixel is used as the model input and the model output is composed as the H, S and V of corresponding pixel in generated texture. Experimental results show that the proposed method can generate many beautiful color textures, whose textures are different from the original texture.

Though the olfactory model entitled KIII has been widely used to pattern recognition, it only can give bare prediction. Combining EM model with the transductive confidence machine, a novel method to recognize hypoxia electroencephalogram (EEG) with a preset confidence level is proposed in this paper. This method can make prediction with confidence measure rather than bare prediction. The experimental results of classifying normal and hypoxia EEGs show that the method can set confidence level in advance for every prediction to control the risk of error effectively.

Modeling olfactory neural systems, the Kill model proposed by Freeman exhibits chaotic dynamic characteristics and has potential for pattern recognition. Fuzzy c-means clustering can classify an object to several classes at the same time but with different degrees based on fuzzy sets theory. Based on the Kill model, mandarin digital speech is recognized utilizing the features extracted by the fuzzy c-means clustering. Experimental results show that the Kill model can perform digital speech recognition efficiently and the fuzzy c-means clustering has better performance than the hard k-means clustering.

To model mammalian olfactory neural systems, a chaotic neural network entitled K-set has been constructed. This neural network with non-convergent "chaotic" dynamics simulates biological pattern recognition. This paper reports the characteristics of the KIII set and applies it to digital classification of the sounds of Mandarin spoken digits. Experimental results show that the KIII set performs digital speech recognition efficiently.