Deep crowd-counting models deployed in safety critical surveillance systems are vulnerable to adversarial patch attacks, yet existing patch attacks typically optimize the entire density field uniformly and overlook the spatial structure of density-estimation errors. We propose the Discrepancy-Aware Asymmetric Patch Attack (DAPA), an offline universal-patch framework that uses prediction–ground-truth discrepancy to prioritize spatially vulnerable output locations. DAPA combines a discrepancy-aware spatial weight map with a sign-aware regression objective that amplifies positive signed model-output responses while attenuating the gradient contribution of negative responses, which oppose count inflation; this branch is active only for models that expose signed outputs. Experiments on ShanghaiTech Parts A and B and two external benchmarks, involving both CNN-based and Transformer-based crowd-counting architectures, show that DAPA achieves the largest performance margins over the evaluated patch baselines in both white-box and black-box settings while generating patches approximately 4–17× faster than the compared baselines. The attack shows strong transferability across the evaluated CNN-based and Transformer-based architectures, transfers between datasets with differing density distributions, remains effective in physical-world deployment, and approaches peak performance with as few as 10% of the training images. DAPA remains the strongest of the compared attacks, although DAPA-based adversarial training largely neutralizes all attacks.
Traditional malware detection methods based on signature matching are limited to identifying only previously recognized malware and are susceptible to evasion techniques such as binary obfuscation. We propose an enhanced malware detection model utilizing higher-order attribute augmentation within heterogeneous graph networks. Firstly, acknowledging the benefits of heterogeneous graph networks in preserving the rich semantic features of nodes, the model is constructed by employing the embedded representation technique of meta-paths and meta-graphs. Then, to address the inadequate direct-neighbor representation for malware nodes and to recognize the importance of hidden relationships within higher-order neighbors, we propose a method that progressively enhances node embeddings at the node level, introducing a higher-order attribute-enhanced embedding learning technique called HoGCN2vec. Finally, by appropriately incorporating node-level local semantics to improve node-level aggregation, a new method is proposed to alleviate semantic confusion arising from noisy information stemming from higher-order neighbor interactions during deep network propagation. Experiments on the publicly available Drebin dataset show that the model is not only effective at malware detection but can also be used to obtain high-level semantic relationships among applications in complex malware networks.
Large-scale edge IoT systems increasingly rely on multi-agent deep reinforcement learning (MADRL) to coordinate distributed devices for compute-intensive tasks. However, existing MADRL paradigms often require a centralized critic or cross device information sharing, which can expose devices' local states and thus raise privacy concerns. This paper develops fedV, a privacy-aware federated deep reinforcement learning framework for volunteer-supported edge computing, where opportunistic volunteer devices are selected to elastically extend edge capacity. Due to intermittent and untrusted volunteer resources, MADRL-based coordination becomes a highly challenging long term optimization problem. To preserve multi-agent cooperation without raw-state exposure, fedV adopts a hierarchical learning architecture based on two-tier federated learning. Specifically, IoT-side agents are trained from local observations, while an edge-side agent aggregates privacy-preserving representations to make global decisions under strict device-side privacy protection. We further introduce a reversed-reward weighted aggregation rule to improve learning efficiency under heterogeneous and non-IID multi-agent interactions. Compared with representative baselines, fedV improves the long-term average reward by up to 30.9% and the converged reward by up to 54.6%. These simulations demonstrate that fedV achieves robust schedulability and scalability under different edge-capacity settings.
Zero Trust Architecture (ZTA) is crucial for 6G networks, which require continuous verification and strict access controls to eliminate implicit trust. The Radio Resource Control (RRC) protocol in 3GPP is a critical element for access control in 6G, defining the signaling procedures between mobile devices (user equipment or UE) and the radio access network (RAN). While RRC includes measures to protect the integrity and confidentiality of signaling messages, it does not fully align with zero-trust principles. For example, despite mutual authentication, encryption, and integrity protection, current RRC security assumes protocol adherence, which conflicts with the zero-trust principle of assuming breach. This paper assesses the current 3GPP release 17 RRC design for 6G against zero-trust standards, identifying weaknesses and vulnerabilities validated through an over-the-air (OTA) testbed. We introduce an intelligent fault injection approach to detect and predict vulnerabilities, considering various internal threat conditions. Three strategies-Listenand- Learn (LAL), Synchronize-and-Learn (SyAL), and Sourceand- Learn (SoAL) are proposed. Extensive experiment results demonstrate that LAL identifies 129 vulnerabilities with 0.072 seconds, SyAL improves precision and recall with fewer tests, and SoAL detects three man-in-the-middle vulnerabilities. Our solution effectively ensures 6G zero-trust security, providing realtime vulnerability detection and proactive defense.
In the context of federated learning, when a participant needs to withdraw from collaboration for specific reasons, it should have the right to completely remove its relevant data from the global model. Additionally, if a malicious party executes a poisoning attack, its harmful data must be removed from the global model. To achieve this, we propose a client-level federated unlearning method with asymptotic optimization (FedAOP), which effectively eliminates the influence of client data, maintains model accuracy, and avoids catastrophic forgetting. This method also introduces a calculation for the optimal unlearning round based on data influence, thereby balancing storage and communication costs. And by randomly initializing certain specific layers, the unlearning process is accelerated. We further assess the effectiveness of unlearning by introducing membership inference and data poisoning attacks during training. Experiments on the MNIST, Fashion-MNIST and CIFAR-10 datasets demonstrate that our method is both effective and efficient.
