In federated news recommendation, representing each news with a single embedding entangles shared semantics with user-specific preference shifts, which weakens personalization under non-IID users and strains communication when catalogs are large. To address this problem, we introduce FedDEN, a Dual News-Embedding framework that cleanly disentangles the two roles of item representation: a server-maintained global embedding captures cross-user semantics, while a lightweight client residual captures user-dependent deviations only where needed. Based on this core and tailored to the peculiarities of news recommendation, we attach two targeted mechanisms: (i) a minimal cold-start bias that substitutes residuals when histories are scarce, enabling immediate participation without extra rounds; and (ii) an alignment-and-sparsity regularization that keeps residuals complementary to the global table while promoting a compact, communicable global news embedding via proximal updates. Experiments on three real world datasets show consistent gains over strong centralized and federated baselines on metrics, together with robust behavior under cold-start settings, and favorable per-round communication. By solving the conflation of shared semantics and user-specific shifts at its source, FedDEN achieves strong personalization with practical efficiency. Specifically, it outperforms centralized baseline NRMS by 13.61 % and FedRec by 15.59 %.

Occlusion presents a significant challenge in human pose estimation. The challenges posed by occlusion can be attributed to the following factors: 1) Data: The collection and annotation of occluded human pose samples are relatively challenging. 2) Feature: Occlusion can cause feature confusion due to the high similarity between the target person and interfering individuals. 3) Inference: Robust inference becomes challenging due to the loss of complete body structural information. The existing methods designed for occluded human pose estimation usually focus on addressing only one of these factors. In this paper, we propose a comprehensive framework DAG (Data, Attention, Graph) to address the performance degradation caused by occlusion. Specifically, we introduce the mask joints with instance paste data augmentation technique to simulate occlusion scenarios. Additionally, an Adaptive Discriminative Attention Module (ADAM) is proposed to effectively enhance the features of target individuals. Furthermore, we present the Feature-Guided Multi-Hop GCN (FGMP-GCN) to fully explore the prior knowledge of body structure and improve pose estimation results. Through extensive experiments conducted on three benchmark datasets for occluded human pose estimation, we demonstrate that the proposed method outperforms existing methods. Code and data will be publicly available.

Federated Learning enables multiple clients to collaboratively train a model without sharing their individual data, thereby protecting local data privacy. However, attackers, such as untrusted servers, can still compromise the privacy of clients’ local training data through various inference attacks. One feasible approach to protect client privacy during training is the incorporation of differential privacy. Nevertheless, achieving an ideal level of privacy protection with differential privacy often degrades the model’s performance, significantly reducing its accuracy. To enhance model accuracy while minimizing the additional client heterogeneity introduced by differential privacy, this paper proposes a method that integrates personalized layers and data transformations, FusTP-FL. The core of our FusTP-FL is the incorporation of personalized layers and personalized data transformations within the client’s local training model, which further reduces client heterogeneity and improves model accuracy. We evaluated the model’s accuracy on six common datasets; experimental results demonstrate that the proposed FusTP-FL effectively enhances model accuracy across two different differential privacy modes (CDP and LDP), increasing it by up to 45%. Furthermore, we show that compared to PRIVATEFL, our method achieves lower client heterogeneity.

Abstract Photos or videos taken by individuals often carry sensitive details such as facial identities, which has led to an escalating societal interest in privacy protection measures. We suggest an improved face identity transformer that offers password-protected anonymization and de-anonymization of photo-realistic facial images in visual data. Our face identity transformer is designed to (1) erase facial identity information after anonymization, (2) restore the original face when a correct password is provided and (3) generate an incorrect but realistic face when given an incorrect password. The processes of image anonymization and de-anonymization are facilitated through a password scheme, a multi-task learning objective and generative adversarial networks comprising InfoGAN and contrastive learning. In-depth experiments indicate that our methodology can execute anonymization and de-anonymization based on password conditions whilst reducing training time and enhancing image quality compared to existing anonymization procedures. Additionally, it maintains a recognition rate as low as 4.8% for anonymized images without sacrificing the face detection rate of the original method.

The rapid development of current data centers and Industrial Internet of Things has brought about the explosive growth of information, which leads to the need for better performance of cluster communication systems. Coflow scheduling has the potential to enhance communication performance among applications in data parallel clusters. However, current coflow scheduling techniques that lack preliminary knowledge often depend on a fixed set of threshold parameters within a multilevel feedback queue (MLFQ), disregarding network variability. Though manually tweaking threshold settings may support network flexibility, it negatively impacts real-time performance and increases workload. Furthermore, manual threshold adjustments often fail to react promptly and adaptively to network environmental changes. To address the issues highlighted above, this paper presents D-MLFQ, a novel approach that leverages deep reinforcement learning to dynamically and autonomously regulate the threshold of MLFQ. As a result, D-MLFQ offers enhanced scheduling and communication optimization capabilities. Furthermore, the study performs trace-driven simulations to assess the efficacy of D-MLFQ. Empirical data indicate that D-MLFQ outperforms Aalo, which utilizes fixed threshold, by up to 1.39× in terms of coflow’s completion time. Compared to other common scheduling algorithms such as per-flow fairness, D-MLFQ achieves up to 2.13× faster completion time.

The swift advancement of distributed computing has enhanced the support for big data and massive-scale models. Yet, delivering superior services to manage large and intricate network flows in data center networks remains a formidable challenge. In this paper, we present ReQ-tank, an intricate flow scheduling approach based on a multi-level feedback queue (MLFQ) devised to achieve flow prioritization and efficient flow scheduling. ReQ-tank employs a two-tier scheduling strategy: On the flow scheduling layer, priority queues are segmented into two categories, and the flows within high-priority queues follow a strict priority scheduling, while those in low-priority queues adhere to differential weighted Round-robin scheduling; On the packet scheduling layer, ReQ-tank modifies the priority of initially high-priority re-transmitted packets to facilitate fine-grained data packet scheduling. We carry out simulation experiments on web search workloads and data mining workloads. Experimental results demonstrate that ReQ-tank can curtail packet wait time in the network, significantly truncate the flow completion time (FCT) of delay-sensitive flows and counteract the issue of flow starvation in traditional strict priority queues. Consequently, ReQ-tank is deemed more suitable for complex distributed network applications.

Industrial Internet of Things (IIoT) is the leading application scenario of the fifth generation wireless communication systems (5G) and beyond. Nonorthogonal multiple access (NOMA) has become a key technology for 5G due to its high spectrum efficiency. In this article, a joint channel and power resource allocation problem is investigated for cross-cell IIoT networks with aim of maximizing sum rate of NOMA-based machine-to-machine pairs and cellular Machine Devices (cMDs). Since joint channel and power resource allocation problem is an NP-hard problem, the original problem is transformed into a hypergraph model to optimize channel and power resource allocation. Then, a channel allocation algorithm based on hypergraph coloring theory is proposed, and an alternative power allocation algorithm is presented. Next, some properties of hypergraph coloring and complexities are analyzed. Finally, simulation results demonstrate that the proposed algorithm outperforms the graph-based algorithm in terms of sum rate, and also improves the spectrum efficiency significantly.

Federated learning enables local nodes to train a global model together by uploading only training updates to the parameter server without exchanging private data. However, as the complexity of the federated learning task increases, the communication volume of the training process becomes extremely large, hence the huge communication traffic becomes a serious bottleneck in current federated learning application. Existing methods reduce communication overhead from two aspects, the number of communications and the traffic per communication. But these methods usually lead to more consumption of computing resources or a decrease in model accuracy. To handle these problems, this paper proposes a data fitting based weight compression algorithm, FWC, which includes four sequential stages: sparsification, polynomial fitting, encoding, reconstruction and two mechanism: warm-up and accumulation. In particular, the warm-up mechanism can well address the problem of slow convergence in early training period. Experimental results on models with different scales show that FWC is able to provide more than 600x traffic compression at the cost of only millisecond-level computational time cost and less than 1% accuracy loss.