Lei Shi

Heterogeneous edge‐intelligence systems require federated models that preserve useful shared representations while adapting to persistent client differences. Existing decoupled personalized federated learning methods separate a shared encoder from private prediction heads and coordinate encoder information across clients. This paper proposes DecFedFreq, a framework that makes encoder residual aggregation adaptive. The method uses a low‐frequency discrete cosine transform descriptor as a deterministic, compressed summary of encoder‐update residuals for similarity estimation, applies affinity‐based grouping to infer dynamic client clusters, and restricts complete encoder‐residual aggregation to those clusters while prediction heads remain local. Across 10 matched random seeds under severe heterogeneous partitions with full participation, DecFedFreq obtains final‐round accuracies of on Fashion‐MNIST, on CIFAR‐10, and on CIFAR‐100. Relative to FedRep, the corresponding paired gains are 1.33, 2.96, and 5.13 percentage points, respectively, with all two‐sided paired ‐tests yielding . Additional five‐seed Fashion‐MNIST analyses of heterogeneous availability, unequal local‐computation budgets, representation alternatives, and a single‐global‐cluster control indicate that the main measured gain arises from similarity‐based grouping and cluster‐restricted aggregation. Discrete Cosine Transform is therefore positioned as the deterministic, empirically stable default similarity representation in the tested setting, with competitive accuracy and an eightfold reduction in similarity‐input dimension relative to the complete raw residual, rather than as an independently optimal accuracy component; server‐only profiling quantifies the added computation and memory over 25–100 active clients.

Federated learning (FL) is a distributed machine learning framework that coordinates clients to train models on their private datasets via a centralized server, thereby mitigating data privacy risks. However, the communication channels involved in this process are untrusted, leaving FL vulnerable to model poisoning attacks launched by adversaries through man-in-the-middle techniques. Such attacks can degrade the accuracy of the global model and ultimately cause the entire FL training process to fail. In this paper, we propose a defense mechanism that integrates secure verification with watermarking, with the primary goal of ensuring the integrity of models transmitted over communication channels and enabling highly reliable FL deployment. Our mechanism leverages a dual watermarking method: first, models are marked using specially generated samples, and then these samples are further watermarked based on a class histogram-inspired approach. This dual strategy enhances both model detection and watermark stealthiness. The key innovation of our method lies in its sensitivity to subtle tampering while imposing no loss in model accuracy. Experimental results demonstrate that our defense mechanism significantly strengthens the resilience of FL models against sophisticated model poisoning attacks, while maintaining high accuracy and reliability.

The Contrastive Language-Image Pre-training (CLIP) model uses contrastive learning to align image and text representations, and fine-tuning CLIP with federated learning can extend its application to professional fields. However, federated CLIP fine-tuning faces two key challenges: insufficient alignment of fine-grained semantics between vision and text modalities and poor adaptability to non-independent and identically distributed (non-IID) data. This paper proposes the Optimal Transport Dual Prompt Personalization (OTDPP) framework, injects prompt parameters into the deep networks of both visual and text encoders, achieves fine-grained cross-modal alignment through optimal transport, and designs a dual prompt tuning mechanism. The framework splits prompt parameters into a shared global part aggregated by the server and a private local part reserved by clients, and it enables personalized adaptation without updating large backbone encoders. Extensive experiments show that compared with classic prompt tuning baseline methods, OTDPP reduces computational and communication overhead, retains client-specific personalized features, significantly improves model adaptability and performance, and thus demonstrates broad application prospects.

Clustered federated learning (CFL) has emerged as an effective solution to address data heterogeneity in traditional federated learning (FL). However, the intrinsic cluster-based structure of CFL introduces new privacy risks, making it more vulnerable to client-level inference attacks. In this article, we propose a novel client-level data reconstruction attack based on conditional generative adversarial networks (cGANs), which exploits intracluster similarities to enhance the quality of reconstructed private data. Unlike prior works, our attack requires only partial access to a victim’s model updates through passive eavesdropping, thereby reflecting a more realistic threat model in decentralized and resource-constrained environments, such as the Internet of Things (IoT). To mitigate this threat, we developed a lightweight and adaptive defense mechanism grounded in local differential privacy (LDP). Our design incorporates dynamic privacy budget decay, selective layerwise noise injection, and real-time similarity-guided adaptation. This approach achieves a favorable privacy–utility trade-off, while explicitly addressing the computational, communication, and latency constraints inherent in IoT environments. The experimental results demonstrate that our proposed attack improves reconstruction similarity by up to 20% compared with existing baselines, while the defense reduces attack success rate by 27.2% with only a 3.3% accuracy drop. However, it significantly lowers computational cost—reducing FLOPs by 42.7%, memory usage by 23.4%, and DP noise processing time by 45.5%—without introducing additional communication overhead. These findings highlight the underestimated privacy vulnerabilities in CFL and underscore the necessity of efficient, context-aware defense strategies.

Blockchain technology has garnered significant attention from academia and industry, with scalability remaining a key challenge. Sharding is a promising solution, dividing the blockchain into smaller partitions called shards, each processing a portion of the transactions to increase throughput. This approach is critical for enabling efficient Proof of Stake (PoS) consensus mechanisms, as demonstrated by the transition of Dogecoin to PoS, where sharding reduces the computational burden on validators and enhances scalability. However, sharding introduces high storage redundancy, as nodes in each shard must collectively maintain a copy of the entire blockchain, imposing substantial storage pressure. To address this, segments are introduced to divide the main chain into smaller parts distributed across nodes. Existing methods, however, randomly assign segments to nodes, resulting in high costs for node setup and segment queries. This paper investigates the optimal allocation of segments within shards to minimize these costs, proposing a Segment Allocation algorithm based on Cost Clustering (SACC). Theoretical analysis and simulations demonstrate that SACC achieves lower setup, query, and total costs while maintaining security and scalability, offering a more efficient solution for sharding-based PoS blockchains like Dogecoin.