Federated learning enables edge devices to collaboratively train models in edge–cloud environments without sharing raw sensor data. In practical edge deployments, however, devices often have access to different subsets of modalities due to hardware heterogeneity, varying sensing conditions, or temporary sensor failures. Such modality heterogeneity makes conventional federated aggregation unreliable, as parameters learned from different modalities evidence may not be directly comparable. Meanwhile, this setting also raises privacy concerns, since image, depth, and signal data may reveal sensitive user or environmental information. To address these challenges, we propose FedHyperM, a privacy-preserving federated multimodal learning framework for edge devices. FedHyperM performs hypergraph-based multimodal prediction by treating modality-specific features as nodes and adaptively constructing hyperedges to capture high-order correlations among multimodal features. On the cloud server, FedHyperM further introduces modality-sensitive aggregation (MSA), which uses the minimum edit distance (MinED) between available modality names to weight uploaded hypergraph learning parameters during aggregation. Experiments on the UR-FALL dataset under different modality existence rate (MER) settings demonstrate that FedHyperM consistently outperforms representative federated multimodal baselines, including FedAvg, FedCor, FedProx, AutoFed, and FedInMM. Compared with the five baselines, FedHyperM improves average accuracy by 3.64
Mobile video streaming serves as a cornerstone for advanced 5G applications and is projected to account for over 75
Deep reinforcement learning (DRL) has significantly advanced autonomous navigation for unmanned aerial vehicles (UAVs) serving as critical mobile nodes in future communication networks; yet, ensuring its security against adversarial threats remains a critical challenge for AI-enhanced network reliability. Existing backdoor attack paradigms primarily rely on explicit action overrides or direct modifications of environment reward signals, both of which leave conspicuous statistical anomalies in policy distributions that are easily captured by conventional defenses. However, the structural vulnerability within the internal critic return bootstrapping process has been largely overlooked. This paper uncovers this hidden attack surface and proposes Value Function Spoofing (VFS), a stealthy backdoor mechanism that injects localized biases into poisoned return targets during policy evaluation. By maintaining near zero mean advantage estimates, VFS achieves successful behavioral hijacking while avoiding the policy entropy collapse characteristic of explicit action overrides, thereby reducing the effectiveness of collapse-based distribution monitors compared with action-poisoning baselines. To counter this threat, we introduce Value Distribution Auditing and Pruning (VSAP), a closed-loop defense framework that extends security auditing with complementary internal state-value analysis alongside external behavioral tracking. VSAP leverages optimal transport metrics to detect global value shifts, deploys differential neuron localization to isolate sparse, pathological network circuits, and executes targeted structural pruning to restore estimation accuracy. Evaluations on the discrete grid-based UAV navigation benchmark demonstrate that VFS achieves high attack success with exceptional stealth, while VSAP effectively mitigates value-function backdoors while preserving benign operational utility, supporting more reliable DRL-based navigation in future network infrastructures.
The satellite-ground integrated network (SGIN) has emerged as a key sixth-generation (6 G) communication architecture, utilizing inter-satellite links (ISLs) and ground-satellite links (GSLs) to enable end-to-end routing. However, uneven traffic distribution causes traffic to converge on specific ground nodes, creating severe bottlenecks that lead to queue saturation and significant packet drops. Existing centralized approaches mitigate overload but suffer from massive signaling overhead and route-recovery latency under GSL dynamics. Conversely, distributed schemes are responsive but fail to resolve bottlenecks globally. This paper proposes a routing framework that retains the global optimization capability of centralized approaches while addressing their limitations under GSL dynamics. We propose congestion-aware dual cost routing (CADR), which estimates global load from active flows without global state collection to resolve bottlenecks via a dual cost. We also integrate a fast route recovery (FRR) procedure, which suppresses drops by constructing a local detour at the failure node and performing asynchronous synchronization. Simulation results demonstrate that the proposed scheme achieves the best overall performance under heavy uneven traffic load. Notably, it reduces the packet drop rate from 38.44
This work investigates the resource allocation for a parasitic symbiotic radio (PSR) network, in which a primary transmitter performs long-packet communication while multiple Internet of Things (IoT) nodes sequentially backscatter their short-packet data to a cooperative receiver. To enhance the primary performance, we propose a novel transmission framework that enables each IoT node not only to backscatter its own short-packet data, but also to assist the primary transmission by reflecting the primary signal without modulating its own information. Within this framework, we formulate an optimization problem to maximize the primary achievable throughput under reliability constraints for all IoT nodes, which yields a highly non-convex mixed-integer nonlinear programming problem. To solve it efficiently, we first employ the block coordinate descent (BCD) method to decouple the problem into two subproblems: the first is solved optimally via the bisection method, and the second is proved to be convex and handled with standard convex tools. A BCD-based iterative algorithm is then developed, which alternately updates the two subproblems until convergence. Simulation results verify the fast convergence of the proposed algorithm and demonstrate that, compared with the existing scheme, the proposed scheme achieves higher primary throughput while satisfying each IoT node’s reliability requirement.
