Xuangou Wu

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 this paradox, we studied battery swap stations in Chengdu. Our findings reveal that the cost of swapping stations could be remarkably reduced by optimizing intrinsic, energy-greedy charging power strategies rather than by altering the extrinsic, 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, a new adaptive 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.

With the rapid development of intelligent transportation and low-altitude economy, collaborative distillation has emerged as an efficient distributed learning paradigm that enables multi-terminal cooperation without sharing raw data. However, the open participation nature of such systems also exposes them to label-flipping attacks, where adversarial devices intentionally alter the labels of their local datasets to mislead the global model aggregation, thus compromising the reliability and trustworthiness of collaborative intelligence. To address this challenge, this paper proposes a label-flipping attack defense framework based on temporal feature modeling and unsupervised anomaly detection. In the early global training phase, each device’s multi-round prediction results on a shared public dataset are collected to extract statistical descriptors, including class distribution, prediction entropy, and inter-class transition strength—forming a short-term temporal feature sequence. An autoencoder is then employed to learn the evolution pattern of benign devices, and the reconstruction error combined with statistical deviation is used to compute anomaly scores. These scores are further analyzed by Isolation Forest model to identify potential attackers in an unsupervised manner. Finally, the anomaly scores are mapped to trust weights for weighted aggregation during subsequent distillation rounds, dynamically suppressing the influence of malicious devices. The proposed framework requires no access to raw data or model parameters, achieving accurate and robust attack detection while preserving privacy. Experimental results demonstrate that the method effectively identifies label-flipping devices under varying attack intensities, significantly enhancing the reliability and robustness of collaborative distillation systems, and showing strong potential for deployment in secure and scalable collaborative learning scenarios.

Edge computing involves distributed devices operating in dynamic environments with diverse resource constraints and highly heterogeneous data distributions. In these settings, personalized federated learning (PFL) provides a collaborative learning framework that preserves the unique data characteristics of each device. However, PFL models often encounter bidirectional catastrophic forgetting. During consecutive training rounds, the personalized characteristics learned by local models are readily overwritten by updates from the global model, while the global model’s shared representations are degraded by distribution shifts arising from heterogeneous local data. This challenge is further amplified in edge computing settings, where devices must adapt to fast-changing heterogeneous data. To address this issue, we propose FedSSM, a novel framework that leverages state-space models to mitigate forgetting in PFL. By capturing the temporal evolution of local model parameters through hidden states, the framework enhances the retention of critical knowledge throughout training rounds. Extensive experiments on multiple benchmark datasets demonstrate that FedSSM outperforms various state-of-the-art PFL algorithms, particularly with high data heterogeneity.

Uncrewed 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, a low-power, high-precision, and low-latency crowd 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 a novel spatio-temporal Compressive-Sensing-based video feature compression algorithm to 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 the temporal and spatial perspectives by accounting for inter-frame redundancy and intra-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-based human activity recognition (FL-HAR) has emerged as a significant research topic due to its extensive applications, enabling the collective integration of local client data and knowledge without compromising user privacy. A major challenge within FL-HAR is addressing the heterogeneity of data across distributed clients, which leads to varying feature distributions for the same activity across different clients. To overcome this limitation, we propose a novel algorithm for FL-HAR, termed FedPMP, with Adaptive Model Partitioning and Personalized Model Aggregation. This approach customizes individual models for each client, aligned with their individual data distribution. Specifically, our partitioning strategy categorizes the model into globally shared components and locally personalized components, based on layer-wise feature drift. The global components capture general features across the majority of user groups, while the local components extract user-specific personalized features. Furthermore, our personalized model aggregation approach separately consolidates the shared and personalized components, facilitating collaboration among clients in accordance with feature relevance. Notably, we prove that FedPMP converges under feature drift conditions, providing a solid theoretical foundation for its practical deployment. Extensive experiments on various real-world HAR benchmarks demonstrate that FedPMP can effectively address the issue of feature drift in HAR, boosting recognition accuracy for clients and reducing the variance in accuracy among them. Our method enhances recognition accuracy by an average factor of 5%, proving its efficacy and robustness under feature drift conditions.

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 an aggregation gap caused 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 propose pFedCal, a novel personalized federated learning with lightweight adaptive calibration 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.

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 paper, we firstly 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 paper is to provide an in-depth understanding for the readers on how to realize the integration between MTD and AI-driven network.

Owing to its outstanding parallel computing capabilities, quantum computing (QC) has been a subject of continuous attention. With the gradual maturation of QC platforms, it has increasingly played a significant role in various fields such as transportation, pharmaceuticals, and industrial manufacturing, achieving unprecedented milestones. In modern society, wireless communication stands as an indispensable infrastructure, with its essence lying in optimization. Although artificial intelligence (AI) algorithms such as reinforcement learning (RL) and mathematical optimization have greatly enhanced the performance of wireless communication, the rapid attainment of optimal solutions for wireless communication problems remains an unresolved challenge. QC, however, presents a new alternative. This paper aims to elucidate the fundamentals of QC and explore its applications in wireless communications and networking. First, we provide a tutorial on QC, covering its basics, characteristics, and popular QC algorithms. Next, we introduce the applications of QC in communications and networking, followed by its applications in miscellaneous areas such as security and privacy, localization and tracking, and video streaming. Finally, we discuss remaining open issues before concluding.