Uncrewed aerial vehicle (UAV) networks are increasingly exposed to widespread and various network attacks due to their fully distributed nature and the limited defensive capabilities of individual devices. Existing defense strategies rely on network connectivity and UAV status information, which overlook information of network traffic. Malicious traffic detection offers a promising solution to achieve fine-grained attack detection. However, the dynamic nature and complexity of UAV networks limit the effectiveness of traditional traffic detection methods. Current approaches either fail to fully exploit the raw characteristics of traffic or do not consider the timeliness requirements of UAV networks. To address these challenges, we propose a novel twin graph encoder neural network, which can extract features of raw traffic bytes for efficient traffic detection. First, we propose a decoupled architecture for model training and inference to enable efficient detection of malicious traffic in UAV networks. Second, we propose a novel modeling method that models traffic as the co-occurrence graph and word frequency graph based on raw bytes. Then, we propose TGE-ETD, a Twin Graph Encoder for Encrypted Traffic Detection. TGE-ETD consists of a set of twin graph encoders that effectively extract intrinsic traffic features from graphs constructed from raw bytes. In addition, TGE-ETD employs a global attention pooling mechanism to effectively distinguish the feature contributions of different bytes. Finally, we conducted extensive experiments on a real UAV traffic dataset and four real-world network traffic datasets. TGE-ETD achieved an improvement of 1%-20% over the baseline methods by reducing the number of parameters by 20 times. Tested on multiple UAV hardware devices, TGE-ETD can achieve millisecond-level traffic detection.

Scrap steel detection is vital for optimizing recycling and manufacturing, but deploying machine learning models in distributed industrial settings faces challenges from data privacy and bandwidth constraints. Federated learning (FL) enables collaborative model training across dispersed clients without sharing raw data, preserving privacy. However, feature drift and bidirectional knowledge forgetting, driven by diverse imaging conditions and non-independent identically distributed data, remain critical bottlenecks that conventional FL cannot effectively address. This article proposes FedCLAM, a novel FL framework that explicitly tackles these challenges through three key innovations: 1) feature decoupling to separate shared and personalized representations and mitigate coupling bias; 2) feature-driven adaptive aggregation to alleviate parameter aggregation bias and enhance global robustness; and 3) knowledge inheritance to preserve both global and local knowledge across communication rounds, effectively reducing forgetting. Extensive experiments on real-world and open scrap steel datasets demonstrate that FedCLAM consistently outperforms state-of-the-art FL baselines, achieving superior classification accuracy, stability, and generalization in heterogeneous and noisy industrial scenarios.

Electric scooters (ESs) serve as smart city terminals by uploading real-time charging data to improve mobility and energy efficiency. Battery swapping service has emerged as a transformative solution for energy replenishmen of ESs. Time-of-use (TOU) pricing is commonly adopted to encourage off-peak electricity usage by offering different rates. However, the rigid work schedules of ES drivers always hinder the optimization of battery swapping, leading to escalated charging costs. Existing literature fails to account for the practical constraints of battery charging in real-world scenarios. To tackle thisparadox, we studied battery swap stations in Chengdu. Our findings reveal that the cost of swapping stations could be remarkably reduced by optimizingintrinsic, energy-greedy charging power strategies rather than by altering theextrinsic, rigid swapping schedules. Further analysis indicates 63% of charged batteries remain idle, presenting an opportunity to optimize charging power for cost savings without disrupting swapping schedules. Motivated by these insights, we propose PowerRL, anewadaptive charging power strategy empowered by spatio-temporal reinforcement learning. Specifically, we employ a spatio-temporal gated recurrent unit to predict the demand for battery swapping, which is then used to optimize charging power through reinforcement learning. Applied to the large-scale dataset, PowerRL could reduce per-unit electricity costs by 25.7% while still accommodating driver swapping schedules, thereby harmonizing TOU pricing with the rigidity of ES swapping service.

Federated learning (FL) is a promising paradigm of distributed machine learning, since it enables the collaborative training of machine learning models across multiple edge devices without compromising privacy. However, when deploying the FL framework in real-world scenarios, the global model is plagued by the aggregate loss issue, which can result in inefficient and unfair model training. Existing works in the literature either fail to ensure high efficiency in model updates or can not achieve fair model updates. Inspired by the art of “Less is more”, we propose FedBFO, a Federated learning framework with Bi-level First-order Optimization. The key idea of FedBFO is to treat model aggregation as the second optimization of local models, which utilize knowledge trimming to design a more tailored global model for each device. To this end, we first propose a knowledgeaware trimming scheme to adaptively aggregate tailored global models for local devices. We then present a anomaly-aware trimming mechanism to iteratively select qualified local models. The theoretical analysis shows that FedBFO can ensure the efficiency, fairness, and convergence of models. Experimental results demonstrate that FedBFO outperforms six state-of-the-art baselines with a 6.7x convergence speed than FedAvg, achieving better tradeoff convergence performance, efficiency, and fairness.

Based on emerging artificial intelligence (AI) tasks, cloud-edge–terminal architecture can provide powerful computing, intelligent interconnection, and real-time response, which can also be regarded as AI-driven network. Unfortunately, multiple network layers in the AI-driven network usually face various types of network threats, such as malicious network reconnaissance, side-channel attacks, and distributed denial of service (DDoS). Traditional security solutions respond to network threats after the occurrence of attacks. To solve this problem, the concept of moving target defense (MTD) has been proposed as a proactive defense mechanism that aims to defend against cyber attacks before they occur. In this article, we first provide a thorough analysis of the threats in the cloud-edge–terminal network. Then, we conduct a comprehensive survey to discuss the concept, design principles, and main classifications of MTD. Next, we further introduce the development potential in terms of AI-powered MTD on each network layer. Meanwhile, we also explore how MTD improves the security of AI algorithms. Lastly, we describe the existing challenges and research directions of MTD. The aim of this article is to provide an in-depth understanding for the readers on how to realize the integration between MTD and AI-driven network.

Unmanned Aerial Vehicles (UAVs) equipped with cameras provide a promising way for large-scale urban crowd surveillance due to their convenient deployment and flexible mobility. However, UAVs are constrained by limited power and computing resources, hindering existing work in achieving efficient UAV-based crowd surveillance,i.e., long flight time, high accuracy, and low latency. To this end, we propose EagleEye, alow-power,high-precision, andlow-latencycrowd surveillance system empowered by edge-assisted UAVs. It leverages lightweight devices on UAV-side to compress video information edge-independently, then transmits key video information instead of raw high-definition videos. Furthermore, we propose anovel spatio-temporal Compressive-Sensing-based video feature compression algorithmto achieve efficient, low-latency video compression. It can reduce video volumes greatly while minimizing the loss of crowd-surveillance-related information. Specifically, inspired by the Compressive Sensing theory, we compress the video content from both thetemporalandspatialperspectives by accounting forinter-frame redundancyandintra-frame information saliency, respectively. Finally, we implement a prototype system and conduct extensive experiments based on four large-scale datasets with over 20,000 frames. The experimental results demonstrate that EagleEye can reduce transmission latency by 31.4% only with no more than 4% of accuracy loss in urban crowd detection.

Federated learning (FL) is a promising artificial intelligence framework that enables clients to collectively train models with data privacy. However, in real-world scenarios, to construct practical FL frameworks, several challenges have to be addressed, including statistical heterogeneity, constrained resources, and fairness. Therefore, we first investigate anaggregation gapcaused by statistical heterogeneity during local model initialization, which not only causes additional computational overhead for clients but also leads to the degradation of fairness. To bridge this gap, we proposepFedCal, a novelpersonalizedfederated learning with lightweight adaptivecalibration strategy that performs calibration compensation through the prior knowledge of clients. Specifically, we introduce compensation for each client at the model initialization, with the compensation derived from the global gradient and the latest gradient bias. To enhance the calibration effect, we introduce a smoothing-based calibration strategy, and we design an adaptive calibration strategy. A representative example demonstrates that the proposed calibration and smoothing strategies improve fairness for clients. The theoretical analysis indicates that with an appropriate learning rate, pFedCal converges to a first-order stationary point for non-convex loss functions. Comprehensive experimental results show that pFedCal achieves faster convergence, higher accuracy, and improved fairness than the state-of-the-art methods.

Federated learning has emerged as a promising approach for fault diagnosis, as its ability to learn from decentralized data while preserving client privacy for industry. Yet, it also brings the problem of nonidentically and independently distributed (Non-IID) data, which can result in model convergence delay and performance degradation. Recent research aims to alleviate the problem caused by cross-domain without considering by cross-location. However, it is common in industrial production to have devices across different monitoring locations. Furthermore, experimental results indicate that the diagnostic models' performance of the latest techniques is significantly affected. To address the cross-location Non-IID data problem, we propose DecFFD, a personalized federated fault diagnosis framework that decouples global and personalized features. In DecFFD, we design a reconstructor for each client that acts as a supervisor and decoupler to disentangle global and personalized features. We then present a client alignment algorithm to eliminate the differences in global features among clients. In addition, we provide a theoretical analysis of fairness and generalization capability, offering a theoretical guarantee for model convergence. Finally, extensive experiments are conducted on two real-world datasets. Experimental results show that the accuracy of DecFFD outperforms the accuracy that of the state-of-the-art approach by 14.67% and converges at a faster rate.

Federated learning (FL) opens a new promising paradigm for the Industrial Internet of Things (IoT) since it can collaboratively train machine learning models without sharing private data. However, deploying FL frameworks in real IoT scenarios faces three critical challenges, i.e., statistical heterogeneity, resource constraint, and fairness. To address these challenges, we design a fair and efficient FL method, termed FedASA, which can address the challenge of statistical heterogeneity in resource-constrained scenarios by determining the shared architecture adaptively. In FedASA, we first present a cell-wised shared architecture selection strategy, which can adaptively construct the shared architecture for each device. We then design a cell-based aggregation algorithm for aggregating heterogeneous shared architectures. In addition, we provide a theoretical analysis of the federated error bound, which provides a theoretical guarantee for the fairness. At the same time, we prove the convergence of FedASA at the first-order stationary point. We evaluate the performance of FedASA through extensive simulation and experiments. Experimental results in cross-location scenarios show that FedASA outperformed the state-of-the-art approaches, improving accuracy by up to 13.27% with better fairness and faster convergence and communication requirement has been reduced by 81.49%.

The integration of Roadside Units (RSUs) is vital for the development of autonomous driving technologies. Challenges arise from sinking computing capabilities into RSUs and vehicles in the paradigm of Vehicle Edge Computing (VEC), particularly due to heterogeneous computation and communication capacities of network nodes and multiple sources of computing tasks (node-mounted and offloading tasks). These challenges complicate network stability from the perspective of a long-term optimization evolving over time, considering unpredictable task distribution and environmental states. To tackle these challenges, we approach the problem of partial task offloading to minimize task delay while meeting the demand of system stability over time as a dynamic long-term optimization. Utilizing Lyapunov stochastic optimization tools, we successfully decouple the long-term delay minimization and stability constraint, transforming it into a per-slot scheduling problem. Since the per-slot scheduling problem with complicated Lyapunov drift functions can not be solved by numerical optimization at each time step, our solution leverages a proposed deep reinforcement learning algorithm, leading to extensive simulations that demonstrate the superior effectiveness and efficiency of our proposal compared to existing schemes.

Encrypted traffic identification is critical for detecting network attacks and improving the quality of service, making it an essential aspect of cyberspace security. However, existing methods for classifying Tor applications are limited by their reliance on single packet size or time-related features of traffic flows, which limits the potential for improving classification accuracy. To address this challenge, we propose an end-to-end classification framework called the Matrix of Fused Features Network (MFFN). This framework fuses direction, packet size, and time-related features of the raw flows to generate matrices that represent Tor applications. We then use the residual net-work, a popular deep learning model, to automatically extract potential features from these matrices and classify them into Tor applications. We evaluate our approach on the public dataset ISCXTor2016, and experimental results demonstrate that MFFN achieves 97% accuracy for Tor application classification, outperforming existing methods.

With the development of blockchain technology, de-centralized applications (DApps) are increasingly being developed and deployed on blockchain platforms. However, the complex data validation mechanism and strict encryption protocol settings of blockchain often lead to sparse traffic behavior of DApps. This sparsity poses a challenge for existing encrypted traffic identification methods to extract distinguishable DApps traffic features. In this study, we propose a novel approach for identifying DApps traffic features by observing the differences in burst timing features of DApps. We introduce a continuous burst feature matrix (CBFM) method based on burst feature aggregation that can aggregate sparse features and express the burst timing differences of DApps encrypted traffic. Additionally, we design a deep learning classifier to automatically extract the features contained in the CBFM. Our experimental results on real datasets demonstrate that the proposed CBFM method achieves a classification accuracy of 94%, outperforming state-of-the-art methods.

Website Fingerprinting (WF) is a traffic analysis attack that can identify the encrypted network activities of clients, allowing for the determination of which web pages the clients are accessing. Recently, many WF defense approaches have been proposed, such as WTF-PAD and Walkie-Talkie, which have now become Tors primary deployed defense mechanisms. However, these WF defense approaches have proven less effective when faced with deep learning technology. The identification accuracy of DF-based attack approaches exceeds by over 90%. To address the above problem, we present time-assisted traffic-shaping defense approach. Our approach integrates multi-website traffic based on timestamp information to incorporate all burst-level features. To create more confusion in website features, we use random time thresholds to segment bursts in different combinations for better shaping. We perform performance evaluations on DApps and find that our approach can significantly decrease the accuracy of DApps fingerprint attacks to less than 50%. Moreover, our approach is also effective in Tor scenarios.

Abstract Today's internet applications generate massive temporal data anywhere and anytime. Although some disk‐based temporal systems are currently available, they suffer from poor I/O performance, especially when applied in intelligent applications deployed in the cloud and edge environments. Therefore, how to process temporal operations with low latency and high throughput becomes a crucial problem for efficient data processing. This paper proposes Timo, a distributed in‐memory temporal query and analytic model for big temporal data. Firstly, a space‐efficient temporal index is proposed to support more efficiently query performance with less memory space than the state‐of‐art methods. Secondly, based on the temporal locality feature of temporal queries, Timo proposes a partitioner mechanism, which utilizes forward scan algorithm, to improve query throughput. Thirdly, some optimal strategies are proposed to improve the execution process of temporal query and analysis, which reduce intermediate result data size and improve the throughput much further. Lastly, we implement the Timo system on the Apache Spark platform, which extends Spark dataset API with temporal query and analysis API functions for users. Extensive experimental results show that the Timo system outperforms other Spark‐based temporal systems in terms of both query latency and throughput.

Fuzzing is a popular technique for finding soft-ware vulnerabilities. Despite their success, the state-of-art fuzzers will inevitably produce a large number of low-quality inputs. In recent years, Machine Learning (ML) based selection strategies have reported promising results. However, the existing ML-based fuzzers are limited by the lack of training data. Because the mutation strategy of fuzzing can not effectively generate useful input, it is prohibitively expensive to collect enough inputs to train models. In this paper, propose a generative adversarial networks based solution to generate a large number of inputs to solve the problem of insufficient data. We implement the proposal in the American Fuzzy Lop (AFL), and the experimental results show that it can find more crashes at the same time compared with the original AFL.

In-air handwriting is a new and more humanized human-computer interaction way. The existing in-air handwriting systems are mainly based on three-dimensional sensors, which are expensive, too large and not conducive for integration and application promotion. To solve this problem, this paper proposes a new in-air handwriting system using cheap and portable monocular camera which allows users writing freely in the air. Additionally we develop an end-to-end fingertip detection algorithm based on improved YOLOv5 algorithm to form in-air handwritten characters. Concretely, we first build a fingertip images dataset. After preprocessing and fingertip labeling, we use the dataset to train the improved YOLOv5 model, and then use the trained model to detect the coordinates of the fingertip in each video frame. After that, we connect the coordinates of the fingertip of each frame to form the character, and finally utilize the classifiers to recognize characters. The experimental results show that proposed in-air handwriting system allows user write freely in the air, and can obtain over 92 % in fingertip detection and character recognition.

In this paper, the state estimation problem for singularly perturbed switched complex dynamic networks (CDNs) is addressed, in which the persistent dwell-time (DT) switching mechanism is employed to depict the switchings among parameters of each node. In the aforementioned switching mechanism, the concept of stage consisting of the persistent portion and the DT portion is introduced. Based on the singular perturbation theory, a two-time-scaling variables containing fast and slow states are taken into account on the CDNs simultaneously, which make the constructed networks more realistic. Furthermore, the main aim is to design a mode-dependent estimator to track the state information that cannot be directly obtained and further ensures the global uniform exponential stability of the investigated systems with a generalized dissipativity property. Some sufficient criteria on the existence of the desired mode-dependent state estimator are established by utilizing a modified matrix decoupling method. Finally, the availability of the estimator design procedures is verified by a simulation example.

It is an inspiring way to provide emergence communication services in natural disaster areas by deploying wireless routers on the ground and multiple unmanned aerial vehicles (UAVs) in the air. The wireless routers serve as access points. UAVs relay data from routers and themselves to a remote base station in a safe place. Thus, people can communicate with others outside the disaster. The network lifetime is restricted to battery lifetime of routers which are scattered over a complex post-disaster area. There is a potential to prolong the network lifetime by utilizing UAV mobility. We consider the trajectory planning of UAVs with the goal of maximizing the network lifetime, which is modeled as a cooperative game with complete information. However, the time complexity of the problem increases exponentially with the number of UAVs as well as UAV candidate strategies. In our proposal, we boost the game process by excluding some of candidates from the strategy space for each UAV. Specifically, there is no influence for a given UAV, taking the strategies excluded, over the other UAVs. In addition, these strategies are dominated by another strategy at least. Our proposed model is verified through simulations that show its advantage on time complexity over others.

Compressive Sensing (CS) has been recognized as a promising technique to reduce and balance the transmission cost in wireless sensor networks (WSNs). Existing efforts mainly focus on applying CS to reliable WSNs, namely, each wireless link is 100% reliable. However, our experimental results show that traditional compressive data gathering (CDG) could result in arbitrarily bad recovery performance, when the wireless links are lossy. In this paper, we study the impact of packet loss on compressive data gathering and ways to improve its robustness using sparsest random scheduling (SRS). The key idea of our scheme is to treat each sampling value as one CS measurement, which helps us to reduce the impact of packet loss on the recovery accuracy. Our scheme also outperforms the tradition CDG in reliable WSNs in that our scheme has significantly lowered transmission cost. To achieve this, we present a sparsest measurement matrix where each row has only one nonzero element. More importantly, we propose a representation basis to sparsify the gathering data, and prove that our measurement matrix satisfies the restricted isometric property (RIP) with high probability. Extensive experimental results show our scheme can recover the data accurately with packet loss ratio up to $15\%$, while traditional CDG can hardly recover the data under similar or even better conditions.

Compressive sensing (CS)-based in-network data processing is a promising approach to reduce packet transmission in wireless sensor networks. Existing CS-based data gathering methods require a large number of sensors involved in each CS measurement gathering, leading to the relatively high data transmission cost. In this paper, we propose a sparsest random scheduling for compressive data gathering scheme, which decreases each measurement transmission cost from O(N) to O(log(N)) without increasing the number of CS measurements as well. In our scheme, we present a sparsest measurement matrix, where each row has only one nonzero entry. To satisfy the restricted isometric property, we propose a design method for representation basis, which is properly generated according to the sparsest measurement matrix and sensory data. With extensive experiments over real sensory data of CitySee, we demonstrate that our scheme can recover the real sensory data accurately. Surprisingly, our scheme outperforms the dense measurement matrix with a discrete cosine transformation basis over 5 dB on data recovery quality. Simulation results also show that our scheme reduces almost 10 × energy consumption compared with the dense measurement matrix for CS-based data gathering.