Federated unlearning has become a necessary mechanism for removing poisoned, sensitive, or withdrawn data from federated models. However, existing unlearning protocols often treat a client's acknowledgment of deletion as evidence of compliance. This assumption leaves a critical false unlearning vulnerability: malicious clients can return updates that appear consistent with unlearning while retaining backdoor knowledge, which may subsequently contaminate the global model through aggregation. The central problem is therefore to distinguish genuine forgetting from deceptive compliance during the unlearning process and to prevent untrustworthy clients from participating in later training rounds. To address this problem, this paper proposes a multidimensional dynamic trust evaluation framework that transforms unlearning verification into privacy-preserving, evidence-driven trust inference.To prevent attackers from identifying and overfitting the trap probes, the probe reference gradient is protected by CKKS approximate homomorphic encryption, while Johnson-Lindenstrauss random projection compressing. The resulting behavioral and quality trust values are fused through entropy-based weighting and evolved across rounds using Markov trust updating.Experimental results under false unlearning attack demonstrate that the proposed framework effectively identifies malicious clients, suppresses residual backdoor effects, and provides a more robust trust-evaluation solution for securing federated unlearning.
The Digital Twin (DT) is at risk of becoming a major source of energy consumption and performance degradation in 6G. The high-fidelity data flows used in DT synchronisation can overload the network, increase latency, and reduce the utility of DT. To address this, the literature proposes communication-centric optimisation, which treats data flows as fixed workloads. However, in practice, fidelity determines which aspects of the physical network are synchronised and, hence, the achievable DT accuracy, while also affecting data volume and latency. To address these, we introduce Synchronisation Energy Efficiency (SEE), a unified performance metric that jointly captures Twinning Rate, Twinning Accuracy, and energy consumption. We then propose a Mixed-Fidelity Graph DT that combines high- and low-fidelity data via inference-based reconstruction. This reduces the reliance on high-fidelity data while preserving DT accuracy. Then, to optimise SEE, we design a two-stage control policy that schedules synchronisation and selects its fidelity based on expected marginal gains. We conduct comparative evaluations across varying network load, deadline constraints, and fidelity ratios. The results show that the proposed approach achieves an 84.1% SEE gain. Moreover, while baseline policies experience SEE collapse under strict deadlines and increased load, the proposed approach maintains a positive SEE.
Indoor Visible Light Positioning (VLP) systems rely on stable optical wireless links, yet they are susceptible to signal blockage and communication interruptions caused by human obstruction, changes in receiver posture, and environmental dynamics. These effects lead to partial link failures and incomplete optical observations; most existing learning-based positioning methods, which rely on complete signal inputs, exhibit significant performance degradation when optical links are interrupted. This paper proposes GeoGated-PiTNet, a robust indoor VLP framework designed to maintain stable positioning under partial signal loss. The framework incorporates the spatial geometric relationships of light sources as prior knowledge into a graph structure through a graph attention mechanism, supporting geometry-aware feature modulation during optical link interruptions. A sample-level energy normalization strategy is introduced to alleviate amplitude scale inconsistency under different channel states, and a dual-view progressive training scheme is developed to jointly optimize unobstructed and obstructed signal views, improving robustness under varying link availability. Experiments conducted on two indoor datasets with mixed signal-to-noise ratios demonstrate that the proposed method consistently outperforms baseline methods. At high obstruction rates (e.g., 20%–30%), the average positioning error is reduced by 40%–50%, while maintaining stable performance without error divergence. Incorporating structure-aware modeling and link-state robustness into learning-based VLP frameworks is effective in enhancing positioning reliability in dynamic indoor optical wireless systems.
Verifiable traceability constitutes a critical security requirement across numerous domains, such as supply chain management. Blockchain technology, renowned for its immutability, presents a promising foundation for achieving such traceability in untrusted environments. However, existing blockchain-based approaches exhibit inherent limitations. They are primarily designed for unidirectional tracing of simple, linear data chains and lack the capability for bidirectional tracing across multi-origin and multi-destination data flow graphs. Existing solutions adaptable to these complex scenarios generally suffer from significant inefficiencies. To address these challenges, this paper introduces a verifiable bidirectional traceability index for blockchain systems. Our proposed scheme enables efficient forward and backward tracing queries. For forward traceability, we employ a hash aggregation technique to compress the verification object (VO). For backward traceability, the data flow graph is partitioned into time-based subgraphs, and a Sorted Merkle Patricia Trie (SMPT) is introduced to construct a cross-subgraph index. Furthermore, an active data replication mechanism is proposed to minimize VO size and associated verification costs. A comprehensive security analysis confirms the verifiability of our approach. Extensive experimental evaluation demonstrates that our scheme achieves an order-of-magnitude reduction in verification time and a 98% decrease in VO size compared to baseline solutions. The optimizations further improve system performance by 65% in terms of overall time cost.
The rapid evolution of multi-orbit satellite systems promises global coverage and highly reliable communications. Nevertheless, efficient and trustworthy multi-orbit satellite communication remains hindered by highly dynamic network topologies and heterogeneous latency. To address these challenges, this paper proposes a novel resource-optimization framework for multi-orbit satellite networks that integrates decentralized blockchain technology with hierarchical-graph-attention-driven multi-agent deep reinforcement learning (HGA-MADRL). Specifically, we introduce a spatiotemporal hierarchical graph attention transformer (ST-HGAT) to structurally encode the states of satellite nodes across different orbital layers. Moreover, a dynamic redundancy control (DRC) mechanism is designed to jointly optimize adaptive meta-reinforcement learning (AMeta-RL) and multi-agent reinforcement learning (MARL) according to environmental complexity and mission demands. The framework supports coordinated optimization of resource allocation, link selection, and energy management, thereby achieving a dynamic balance between performance and computational efficiency. Finally, simulation results demonstrate that the proposed framework exhibits significant advantages over existing baseline algorithms in convergence speed, training stability, and resource utilization.
