To mitigate the rising concern on privacy infringement in recommenders, federated recommender (FRec) boosted with local differential privacy (LDP) has been proposed, in which each user privately retains his (or her) dataset, and only exposes model parameters distorted by LDP to a parameter server (PS). Although such kind of solution can largely preserve data privacy, it overlooks the potential unfairness issue. For example, a user can become a free-rider by setting an extremely small privacy budget to completely distort his (or her) model such that the model training is mainly contributed by other users. To guarantee fairness, we design a fairness-aware federated recommender framework called FFRec when users adopt heterogeneous privacy budgets. In FFRec, fairness implies that clients with higher utilities will have more opportunities to participate in training. We integrate fairness into FFRec by taking both sample size and privacy requirements into account. Specifically, a user is selected by the PS to participate in model training based on his (or her) sample size and potential noise influence by LDP and cumulative participation counts. We formulate the fairness-aware client selection problem as an integer programming problem and then reduce it into a submodular problem which has been proven to be NP-hard. To resolve this problem, we employ an approximation algorithm with an approximation ratio of 1 − e−12. Extensive experiments with popular field-measured datasets (i.e., Movielens and Netflix) demonstrate that FFRec can improve fairness by 10.3%-38.5% while guaranteeing high recommendation accuracy

Ethereum 2.0 adopts a new publish/subscribe-based network protocol stack to support its peer-to-peer (P2P) architecture for exchanging consensus messages under the Proof-of-Stake mechanism. Each node scores its peers based on historical behavior and maintains a mesh set of high-scoring peers that can receive and forward full messages. As a result, message propagation time and bandwidth consumption are highly sensitive to the mesh set size. Despite its critical role, optimizing the mesh size to balance propagation delay and bandwidth cost has received limited attention. To fill this gap, we propose the Ethereum 2.0 Network Analyzer (Eth2.0-NA), a framework that optimizes mesh size through two key contributions: (1) a theoretical model that quantifies the relationship between message propagation hop count and mesh size, enabling mesh size optimization for a given hop count target; and (2) a distributed measurement architecture employing multiple probe nodes to empirically validate the model. Experimental results confirm that the theoretical model aligns closely with real-world network behavior. Leveraging this model, we recommend reducing the default mesh size from 8 to 7 on the Ethereum 2.0 MainNet, achieving a 14.1% reduction in bandwidth consumption without increasing propagation delay.

As users request their preferred metaversal media, e.g., 360-degree video, user requests tracked by metaverse content providers (MCPs) pose significant privacy leakage risks. Unfortunately, existing privacy-enhancing techniques are largely ineffective in protecting user privacy for metaversal content requests since these requests cannot be easily altered or concealed by users and must remain visible to MCPs to ensure accurate content delivery. To safeguard user privacy in metaverse multimedia services (MMS), one practical approach is pre-fetching multimedia content (e.g., short videos, video patches in 360$<^>\circ$degrees videos) that is not directly related to users' interests, thereby preventing MCPs from accurately inferring user preferences. However, plain pre-fetching strategies encounter a critical trade-off between privacy protection and edge caching performance given that MCPs often rely on edges for distributing metaverse content. In this paper, we propose a cache-friendly and privacy-aware content pre-fetching (CRACE) algorithm for user devices (UDs) along with a complementary caching algorithm for edge caches (ECs). CRACE effectively mitigates privacy leakage in metaverse content requests while minimally impacting caching performance. Specifically, we introduce a novel privacy model to guide pre-fetching decisions and formulate a Stackelberg game to analyze strategic interactions between UDs and ECs. We derive optimal strategies that maximize their respective utilities and demonstrate the existence and uniqueness of the Stackelberg equilibrium. Extensive experiments conducted with real-world data demonstrate that CRACE significantly enhances privacy protection, reducing privacy disclosure by up to 59.03% compared to baseline algorithms, with negligible impact on the edge caching performance.

Federated learning (FL) enables clients to learn a machine learning model collaboratively without sharing their private local data to the server. However, due to its distributed structure, FL is vulnerable to poisoning attacks where adversaries intentionally send the poisoned local model parameters to the server and further affect the behavior of the global model. Existing works on mitigating poisoning attacks in FL are difficult to accurately characterize data relationships of high-dimensional parameters. Moreover, they cannot guarantee Byzantine robustness when the majority of clients are compromised (Byzantine ratio) and clients’ local datasets are highly non-independent and identically distributed (non-IID). In this paper, we conduct a pioneering work to introduce deep one-class classification into mitigating poisoning attacks in FL, which can guarantee Byzantine robustness even under a Byzantine ratio and non-IID degree greater than 0.5 with the assistance of a root dataset. Our key idea is to adequately learn vector features of benign local model parameters using Deep Support Vector Data Description (Deep SVDD) and achieve the optimal classification by training a deep learning-based one-class classifier equipped with a proper decision boundary based on the root dataset. To further optimize the classifier, we employ a regularizer based on random noise injections, which can address the hypersphere collapse problem inherent in Deep SVDD. Exhaustive experiments on MNIST, F-MNIST, and CIFAR-10 demonstrate that compared with five typical Byzantine-robust methods, our defense strategy achieves excellent effectiveness in mitigating targeted/untargeted poisoning attacks and an adaptive attack in FL. Even with a similar but different root dataset, it still maintains a good Byzantine robustness.

Automatic Medical Coding (AMC), which assigns standardized International Classification of Diseases (ICD) codes to clinical notes, is essential for medical reimbursement, quality reporting, and clinical research. Existing pre-trained language model (PLM)-based methods typically formulate AMC as an extreme multi-label classification problem over a predefined code set, while recent large language model (LLM)-based approaches instead frame it as generation or multi-step reasoning. However, key challenges remain, including the extreme length of clinical notes that hinders effective interpretation, the vast ICD label space, and complex coding rules that are not explicitly captured by LLMs. In this work, we propose Knowledge-Guided Reasoning over Clinical Evidence with LLMs (KREL), a framework that leverages LLMs for clinical text understanding and reasoning while integrating external ICD coding guidelines as structured knowledge. This design enables tight coupling between domain knowledge and LLM reasoning, reducing hallucinations and improving compliance with coding standards. Experiments on benchmark datasets show that KREL consistently outperforms strong PLM-based and state-of-the-art LLM-based baselines.

Electricity theft is a global problem that can lead to significant economic losses and potential grid security risks. However, detecting electricity thefts remains a challenging task even with machine learning-based methods. The primary obstacle lies in the low quality of available data. First, the user categories may be incorrectly labelled. Second, the data distribution is highly imbalanced, with theft instances forming only a small fraction of the dataset. To address these challenges, we propose a novel electricity theft detection method that enhances data quality from two key perspectives. First, we introduce the Label Cleaning algorithm based on Clustering and Locally Sensitive Hashing (LCCLSH) to identify and correct mislabelled instances of suspicious electricity users. Second, we develop the Negative Sample Generation algorithm based on Roulette Wheel Selection strategy (NSG-RWS) to enrich and diversify theft samples. This approach not only increases the proportion of negative samples but also ensures their representativeness and reasonableness selecting high-confidence seed samples to generate new negative samples. At last, we conduct data enhancement by using real-world power consumption datasets. By leveraging enhanced datasets, we further conduct experiments to detect electricity thieves. The results demonstrate that the electricity theft detection accuracy is considerably improved, highlighting the effectiveness of our study.

Recently, federated learning (FL) has gained momentum because of its capability in preserving data privacy. To conduct model training by FL, multiple clients exchange model updates with a parameter server over the Internet. To accelerate the communication speed, it has been explored to deploy a programmable switch (PS) in lieu of the parameter server to coordinate clients. The challenge to deploy the PS in FL lies in its scarce memory space, prohibiting running memory consuming aggregation algorithms on the PS, like TopK. To overcome this challenge, we propose Federated Learning In-Network Aggregation with Global TopK Model Updates (Fed-GTopK) algorithm, consisting of two phases: voting and aggregating. In the voting phase, clients efficiently upload their votes for top model updates to the PS for selecting global top ones. Note that the voting phase consumes little memory or communication resources by only exploring the sparsity of top model updates without transmitting any model update values. In the aggregating phase, clients can reach the consensus to upload global top model updates such that the PS can swiftly aggregate global top model updates in a streamline manner without consuming much memory cost. Compared with existing works, our study is the first one accelerating in-network aggregation for FL by sparsifying model updates, and hence achieving the highest compression rate and the best learning performance. Finally, we conduct extensive experiments by using public datasets to demonstrate that Fed-GTopK remarkably surpasses the state-of-the-art baselines in terms of both model accuracy and communication traffic.

To address growing concerns about data privacy on mobile devices, the federated learning (FL) paradigm enables clients to collaboratively train models while sharing only local model updates. However, privacy risks remain in FL, as adversaries can still infer sensitive information from these updates. To enhance secure aggregation in FL, various protection mechanisms combining encryption and multi-party computation (MPC) have been proposed. These approaches, however, often introduce substantial communication and computational overhead, making secure aggregation impractical on resource-constrained devices, e.g., smart phones. To tackle these efficiency challenges, we are among the first to propose the integration of differential privacy (DP) with encryption and MPC for secure aggregation. Our proposed protocol, Federated Learning with Noise-based Secure Aggregation (FedNSA), injects noise through DP to obfuscate individual model updates. Encryption is employed to correlate the noise across different clients, while MPC ensures perfect noise cancellation at the server side. Finally, we theoretically analyze its advantages and conduct extensive experiments on public datasets to demonstrate the superiority of our approach across multiple dimensions in comparison with the state-of-the-art baselines.

Multi-modal large language model (MLLM) inference scheduling enables strong response quality under practical and heterogeneous budgets, beyond what a homogeneous single-backend setting can offer. Yet online MLLM task scheduling is nontrivial, as requests vary sharply in modality composition and latent reasoning difficulty, while execution backends incur distinct, time-varying costs due to system jitter and network variation. These coupled uncertainties pose two core challenges: deriving semantically faithful yet scheduling-relevant multi-modal task representations, and making low-overhead online decisions over irreversible multi-dimensional budgets. Accordingly, we propose M-CMAB (Multi-modal Multi-constraint Contextual Multi-Armed Bandit), a multi-adapter-enhanced MLLM inference scheduling framework with three components: (i) a CLS-attentive, frozen-backbone Predictor that extracts compact task representations and updates only lightweight adapters for action-specific estimation; (ii) a primal-dual Constrainer that maintains online Lagrange multipliers to enforce long-horizon constraints via per-round objectives; and (iii) a two-phase Scheduler that balances exploration and exploitation under irreversible budgets. We establish a regret guarantee under multi-dimensional knapsack constraints. On a composite multimodal benchmark with heterogeneous backends, M-CMAB consistently outperforms state-of-the-art baselines across budget regimes, achieving up to 14.18

Federated learning (FL) faces significant challenges from modality heterogeneity, which motivates multimodal federated learning (MFL) to leverage complementary modalities across decentralized clients for improved performance. However, modality imbalance introduces a new attack surface, making MFL more vulnerable to membership inference attacks (MIAs), an issue that remains largely unexplored. In this work, we present the first systematic study of MIAs against MFL and propose a modality-aware attack framework. We show that multimodal models are inherently more susceptible to MIAs due to heterogeneous modality contributions, and existing attacks are suboptimal as they treat multimodal parameters as a whole. By performing MIAs on individual modalities, we find that (i) attacking the dominant modality achieves comparable accuracy with lower overhead, and (ii) different modalities expose distinct membership patterns. To identify members with different patterns, we propose a modality-aware framework that exploits cross-modal performance gaps to adaptively select attack modalities and calibrate inference results. Experiments on three datasets show our approach outperforms baselines across multiple metrics.

Federated learning (FL) is revolutionizing machine learning by enabling multiple decentralized clients to collaboratively train a shared model. In mobile scenarios, client devices exchange model parameters with a central server via wireless channels. However, designing efficient wireless FL (WFL) is challenging due to limited energy and channel capacity. To fully utilize communication resources, over-the-air (OTA) computation has been introduced, allowing direct aggregation of analog parameter signals. However, it conceals clients' information from the server, making advanced client selection strategies, e.g., cluster-based client selection, and sparsification compression algorithms like TopK inapplicable. To address these limitations, we propose the WFL with Voting-based Clustering (WFL-VC) algorithm, which can integrate advanced client selection and the TopK model sparsification algorithm with OTA computation. WFL-VC consists of two phases: 1) Phase 1: clients vote on significant parameters based on local models, allowing the server to select clients and identify the global Topk parameters; and 2) Phase 2: the selected clients upload model update parameters with globally aligned indices for over-the-air computation at the server. By combining OTA computation with cluster-based client selection and TopK sparsification, WFL-VC substantially reduces the energy consumption of WFL. Extensive experiments on real-world datasets show that WFL-VC outperforms competitive baselines while consuming considerably less energy.

Fine-tuning remains essential for adapting large models to diverse downstream tasks, yet doing so in a privacy-preserving and resource-efficient manner is challenging, particularly in federated learning (FL) on edge devices. Parameter tensor freezing is a promising solution. However, current methods face key limitations. Static tensor freezing struggles to adapt to the non-IID (non-independent and identically distributed) data distributions across FL clients, while localized dynamic freezing may lead to slow convergence or divergence across clients, harming overall accuracy. We propose FedFreeze, a communication-aware and dynamic tensor-freezing framework for federated fine-tuning over resource-constrained edge networks. FedFreeze delivers two key benefits: (1) it explicitly incorporates computation costs and bandwidth-dependent communication costs into freezingmask optimization, actively selecting which tensors are updated and transmitted to reduce computation and communication overheads; and (2) it improves convergence stability by coor-dinating freezing decisions based on sampled client statistics. We further analyze the memory usage patterns of FedFreeze and introduce the first kind of memory management strategy to minimize memory consumption of tensor-freezing based methods in FL. Experimental results based on real-world traces from NVIDIA Jetson hardware demonstrate that FedFreeze accelerates convergence by up to 5.46× and reduces peak memory usage by up to 53.9% without compromising model accuracy. Furthermore, evaluations under heterogeneous computation and communication environments confirm its robustness.

Differentially Private Federated Learning (DPFL) strengthens privacy protection by perturbing model gradients with noise, though at the cost of reduced accuracy. Although prior empirical studies indicate that initializing from pre-trained rather than random parameters can alleviate noise disturbance, the problem of optimally fine-tuning pre-trained models in DPFL remains unaddressed. In this paper, we propose Pretrain-DPFL, a framework that systematically evaluates three most representative fine-tuning strategies: full-tuning (FT), head-tuning (HT), and unified-tuning(UT) combining HT followed by FT. Through convergence analysis under smooth non-convex loss, we establish theoretical conditions for identifying the optimal fine-tuning strategy in Pretrain-DPFL, thereby maximizing the benefits of pre-trained models in mitigating noise disturbance. Extensive experiments across multiple datasets demonstrate Pretrain-DPFL's superiority, achieving 25.22% higher accuracy than scratch training and outperforming the second-best baseline by 8.19%, significantly improving the privacy-utility trade-off in DPFL.

Knowledge graphs (KGs) are widely used in recommender systems to alleviate data sparsity and cold-start problems, and improve explainability. Recent advances in graph contrastive learning enable better capture of higher-order structural information and help address the challenge of sparse supervisory signals. However, existing methods often treat contrastive views in isolation using graph neural networks, which can lead to noise accumulation and weaken user preference representations. To address these issues, we propose Knowledge-Flow Contrastive Learning (KFCL), a novel framework that reduces noise by facilitating the propagation of node representations across layers and leverages multi-layer information to guide contrastive learning. We enhance KG representations through heterogeneous relation-aware aggregators and update user and item representations within a collaborative knowledge graph. These representations are further refined through a user-item interaction graph and then undergo feature fusion to better capture user preferences. KFCL also incorporates adaptive layer aggregation and learnable interaction weights to further suppress noise and improve learning stability. Extensive experiments on three public benchmark datasets show that KFCL consistently outperforms state-of-the-art methods.

With the advent of the Industry 4.0 era, Federated Learning (FL) provides robust data privacy protection for smart manufacturing and supply chain optimization, while facilitating collaborative intelligent optimization across enterprises and devices. However, the complex and overparameterized deep neural networks used in FL result in significant computational overhead for Industrial Internet of Things (IIoT) devices, leading to low energy efficiency and hindering the practical deployment of FL on IIoT devices. Moreover, the widespread data and device heterogeneity in the IIoT exacerbates the decrease in energy efficiency caused by inconsistent computational efficiency across nodes. This article proposes an energy-efficient dynamic model pruning method for FL, named EDPrune-FL, to address the aforementioned challenges. Compared to existing methods, this approach offers greater flexibility and efficiency by utilizing a dynamic pruning rate allocation mechanism. This mechanism updates the pruning rate for each participating client in every communication round, allowing the pruning upper bound to adapt to the varying importance of different learning stages in FL. EDPrune-FL ensures the global model’s performance while reducing the training energy consumption of clients in heterogeneous environments. To guarantee that dynamic pruning maintains the stability and effectiveness of the model in heterogeneous environments, we also demonstrated the convergence of EDPrune-FL and discussed the relationship between pruning rates and convergence, providing a qualitative analysis. Experimental results demonstrate that our method outperforms the state-of-the-art technique across four real-world datasets. With tests conducted on 100 clients, our approach reduces energy consumption by 10% while maintaining comparable accuracy.

Ethereum 2.0 has adopted Proof-of-Stake (PoS) consensus to replace the energy-intensive Proof-of-Work (PoW) consensus in Ethereum 1.0. Under PoS, each block must be processed within a fixed 12-second slot, with a strict 4-second propagation window to ensure security and stability. Fast block propagation is therefore critical not only to meet this requirement but also to maximize throughput by allowing more transactions per block. However, our investigation on Ethereum 2.0 MainNet reveals that approximately 9% of blocks exceed the 4-second propagation threshold, leading to forks that undermine network reliability. This propagation delay is primarily caused by two factors: (1) serial transaction execution, which prolongs validation time; and (2) large transaction payloads, which increase transmission time. To address these challenges, we propose FBP-Eth2.0 (Fast Block Propagation in Ethereum 2.0), a novel approach that integrates two core innovations: (1) a practical parallel transaction execution scheme by classifying transactions into independent and dependent sets to reduce validation latency; and (2) a proactive compact block protocol that leverages the fixed 12-second slot to proactively exchange transaction pool snapshots among nodes, precisely identifying and replacing redundant transactions with ultra-short hashes. This proactive compaction design compresses the block while eliminating the additional communication overhead typically associated with compact block protocols, significantly improving transmission efficiency. Extensive experiments on Ethereum 2.0 MainNet demonstrate that FBP-Eth2.0 accelerates block propagation by up to 3× over existing schemes. Importantly, it reduces the ratio of blocks exceeding the 4-second threshold to below 1%.

Accurate disease diagnosis increasingly depends on integrating diverse clinical data such as health records, medical images, and clinical notes. However, real-world medical datasets are often incomplete and distributed across hospitals under strict privacy regulations, which limiting the development of robust Machine Learning (ML) models. This study investigates multimodal Federated Learning (FL) as a privacy-preserving framework for collaborative model training without sharing raw patient data. Using a selected subset of 312 patients from the MIMIC-IV database containing complete EHR, medical imaging, and radiology note modalities, neural encoders were trained for each modality and distributed across simulated hospital nodes to simulate federated training. The predictive objective was to classify in-hospital mortality. Generative AI tools, including ChatGPT and Claude, were employed to synthesize missing radiology notes to address incomplete modalities. The results indicate that multimodal inputs yield higher predictive accuracy than unimodal input. Employing generative AI to synthesize missing modalities problem effectively restores model performance, while federated learning preserves patient privacy without compromising predictive accuracy.

Deploying Large Language Models (LLMs) on memory-constrained edge servers to serve requests from mobile devices is challenging due to their substantial resource demands. The Key-Value (KV) cache and Feed-Forward Network (FFN) parameters consume the majority of available memory. However, existing methods typically rely on static memory partitioning for these components. This rigidity leads to critical performance bottlenecks: (1) An insufficient region for neurons causes neuron swapping or direct computation on the CPU, introducing high latency. (2) As the KV cache grows until it exhausts the allocated region, the system must resort to high-latency fallbacks (e.g., KV cache offloading, recomputation, or neuron swapping). Contextual sparsity of ReLU-based or sparsity-friendly LLMs, where only a small subset of neurons is active during inference, is a promising solution. By leveraging the skewness in neuron activation frequency, we observe two opportunities. (1) We can utilize memory fragmentation to store hot neurons. (2) When memory is exhausted, we prioritize evicting cold neurons from the GPU. Both opportunities improve inference throughput. Based on these insights, we propose ElasticMem, a dynamic memory management framework. It splits KV cache and neuron parameters into blocks to enable flexible neuron placement and eviction. We further design several mapping tables to enable logical-to-physical mapping. Moreover, specialized FFN operators and a CPU-GPU hybrid scheduling pipeline support efficient execution of ElasticMem. Finally, experiments on real-world edge platforms show that, compared with PowerInfer, the state-of-the-art sparsity-aware CPU-GPU hybrid execution baseline, ElasticMem improves throughput by up to $5.97\times$ across multiple LLMs and deployment scenarios. In addition, compared with dense llama.cpp under memory pressure, ElasticMem achieves up to $54.7\times$ higher throughput.