Federated Learning is a distributed machine learning paradigm designed to protect data privacy and security while allowing multiple participants to collaboratively train a shared global model without uploading their data to a central server. Modern machine learning models often consist of millions of parameters, resulting in significant computational complexity and a heavy communication burden during the collection and distribution of training data. This becomes a key bottleneck that limits the effectiveness of federated learning. In order to reduce communication costs, we first propose the CS client selection strategy, which integrates model performance and fairness to construct scheduling priority metrics. Client selection is then carried out based on these metrics. And then, we propose a dynamic gradient sparsification algorithm that adjusts the level of sparsification applied to gradient updates for each client, ensuring that every client uploads a suitably compressed update. Finally, we conduct comprehensive experiments and the experimental results show that the method proposed in this paper can effectively reduce communication costs while ensuring model accuracy. © 2024 Copyright held by the owner/author(s).

This paper studies the difficult task of designing federated learning algorithms tailored for wireless ad hoc networks. Federated learning in such networks presents numerous challenges, including signal interference, decentralized infrastructure, dynamic changes in network topology, heterogeneous devices, and diverse data statistics. In response to these challenges, this paper proposes a fully distributed and personalized federated learning algorithm specifically designed for wireless ad hoc networks, named ADDPFed. ADDPFed tackles the issue of wireless interference by leveraging non-orthogonal multiple access technology and successive interference cancellation for enhancing overall communication efficiency. Another key strength of ADDPFed lies in its ability to enable direct local model exchanges among neighboring clients, eliminating the need for central server coordination for model aggregation. For the model aggregation, ADDPFed first calculates the distance between local models using the Tonimoto coefficient, and then assigns suited aggregation weights to these models. This approach emphasizes the contribution of similar models to the global model, effectively tackling the issue of data statistical heterogeneity. Extensive experiments validate the effectiveness of ADDPFed, paving the way for enhanced collaborative and distributed learning paradigms in wireless ad hoc networks. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

Federated learning enables users to train models without transmitting local data to central servers, preserving data privacy and reducing communication costs. However, federated learning also faces various challenges, such as single point of failure issue, byzantine attack, and a lack of incentive mechanism. In this paper, a trustworthy and incentivized federated learning framework is proposed. Firstly, a blockchain network is used as a substitute for the central server in federated learning, addressing the vulnerability of servers to attacks and dishonesty. Secondly, a multi-key homomorphic encryption mechanism is introduced to safeguard the privacy of local models. Furthermore, to address the challenge of detecting malicious models within encrypted models, this paper proposes a verification mechanism based on the cosine similarity of group models. Finally, this paper evaluates the contributions of multiple participants using Shapley Value. This process effectively coordinates and distributes the interests of multiple participants, ultimately incenting more users to engage in federated learning. Experimental results demonstrate that the framework proposed in this paper can still converge and maintain good training accuracy even in the presence of attacks. © IFIP International Federation for Information Processing 2025.

Ultra-reliable and low-latency communication (uRLLC) plays an important role in the development of 5G-advanced and 6G wireless networks. Combining unmanned aerial vehicles (UAVs) with non-orthogonal multiple access (NOMA) offers a promising solution to achieve improved reliability and lower latency. This is made possible by enabling line-of-sight (LoS) links and concurrent transmissions through the use of UAVs and NOMA, respectively. However, because of the inherent openness of wireless channel, uRLLC faces the security challenges against being eavesdropped. Physical Layer Security (PLS) has been proposed as an efficient method to secure uRLLC, since it uses only the properties of wireless channels (such as fading, interference, and noise). Although the potential benefits of NOMA-UAV provide a better coverage for ground users, it remains a significant challenge since it may provide a LoS link to eavesdroppers. Therefore, in this paper, we investigate the security and reliability performance of UAV and NOMA based uRLLC scenario, under which UAV serves two different types of users with different needs, i.e., secret users and public users. By using stochastic geometry tools, we derive the closed-form expression of the secrecy rate, an important metric in the study of PLS. Additionally, the secure performance is enhanced by maximizing the secrecy rate through optimizing the hovering height and power assignment of UAV. It should be noted that the hovering position is optimized via power allocation when there is only one secret user. Evaluations demonstrate the effectiveness and correctness of our theoretical analysis. © 2023 IEEE.

As one of the most popular cryptocurrencies, Bitcoin is essentially a decentralized ledger. Each node maintains the security of the blockchain through the workload proof mechanism, and the block that obtains the accounting right will receive a block reward in the form of Bitcoin. Because the Bitcoin system follows the "longest legal chain" principle, when a fork occurs, orphan blocks will inevitably be generated, and some miners’ computing power will be a waste. In recent years, researchers have discovered that miners can obtain profits disproportionate to their own computing power by deviating from Bitcoin’s honest mining. Selfish Mining (SM1) is a case in dishonest mining strategy, and dishonest miners (attackers) can obtain higher returns by retaining the blocks they create and selectively delaying their release. The stubborn mining strategy is a generalized form of selfish mining. It increases the revenue of the stubborn miner by adopting a wider range of parameters. Its three mining strategies are: Lead-Stubborn, Equal Fork stubborn and Trail stubborn. The mining problem can be formulated as a Markov Decision Process (MDP), which can be resolved to give the optimal mining strategy. This work describes the three mining strategies of stubborn mining as a Markov decision process, solves it and gives the lower bound of the highest return under the optimal stubborn mining strategy. Our experimental results demonstrate that the revenue of the optimal stubborn mining strategy is higher than SM1 under certain circumstances, and this strategy allows dishonest miners (stubborn miners) to obtain revenue that does not match the actual computing power paid. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.

The payment channel networks (PCN) technique effectively improves the transaction efficiency of a blockchain system, further promotes its practical application. Atomic Multi-Path Payments (AMP) are usually used, in payment channel networks, to divide transactions to improve routing efficiency, thereby improving the transaction throughput. Improper transaction division, however, may increase the occurrence of routing failure. Therefore, how to perform efficient transaction partitioning is an urgent problem to be solved. In this work, we propose an improved transaction partition method, named Proportional Atomic Multi-Path Payments (PAMP), which can enhance the efficiency of transaction routing. The key insight of PAMP is that, when a transaction is executed, the trade share can be well divided by the remaining capacity in multiple channels, which can greatly improve the routing efficiency and maintain the balance of channel capacity in the network. Simulation results show that, in contrast to traditional routing algorithms, the transaction success rate is increased by 2.3%, and the average execution time is reduced by 75.09 ms. PAMP improves the transaction routing efficiency, and also promotes the balance of network channel capacity. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.

In this paper, we explore the challenges associated with federated learning, a distributed machine learning paradigm that promotes collaborative model training while preserving the privacy of local client data. One significant hurdle is the non-IID nature of clients' data, alongside limited communication resources between clients and the cloud server. These statistical heterogeneity and communication resource limitations pose practical obstacles to the implementation of federated learning. To address these challenges, we propose a communication-efficient framework called GCPFL for personalized federated learning. Our framework empowers individual clients to train personalized models while substantially reducing communication costs. Specifically, each client compresses the gradient before uploading it and handles the effects of gradient compression through an error correction process. By uploading only the compressed gradients, the communication costs are significantly diminished. On the cloud server side, the received gradients are recovered into models, and similarity aggregation is performed on these models to facilitate collaboration among clients. Once the aggregated models are received, clients conduct local updates to acquire personalized models. Extensive experimental results illustrate that the GCPFL algorithm not only achieves high model accuracy but also substantially reduces communication costs compared to existing methods. © 2023 IEEE.

PoS is a consensus mechanism where the probability of generating a block is directly proportional to a participant's stake, but block generation rights concentrated in rich participants can lead to unfairness. Existing solutions mainly use linear consensus mechanisms where block generation rights are linked to the stake, but since these solutions fail to extend block generation rights to the majority of participants, compounding may occur. To address the problem, we propose a block producer selection method that extends block generation rights, called Grouped Polynomial Elections-based PoS (GPE-PoS). Specifically, GPE-PoS groups participants with similar stakes into independent groups, and performs polynomial calculations on the nodes in each group. Then GPE-PoS calculates the probability of block generation for each participant based on the polynomial calculations results and selects the participant with the highest probability as the block producer. At the same time, the Gini coefficient is used to measure the fairness of the blockchain system, where a Gini coefficient closer to 0 indicates higher fairness. Simulation results show that compared to linear consensus mechanisms based on stake, GPE-PoS has a stable Gini coefficient of around 0.1, which is an 83.33% reduction compared to traditional PoS. © 2023 IEEE.

Generative Adversarial Networks (GANs) have been used in the field of speech enhancement due to their huge potentials in reducing the noise mixed in the signals. Most of existing GAN-based speech enhancement approaches either operate on time domain or exploit the magnitude spectra in time-frequency domain, but lack consideration of direct optimization of the phase. In this paper, we propose a GAN architecture for speech enhancement based on gated linear units (GLUs) and Dual-Path Transformers (DPTs), which simultaneously deals with the amplitude and phase information on the time-frequency domain. The generator of the proposed GAN architecture is designed following an autoencoder structure fed by the real and imaginary parts of the time-frequency frames. The encoder of the generator is constructed by multiple cascaded convolutional GLUs (ConvGLUs), while the decoder consists of two groups of cascaded deconvolutional GLUs (DeconvGLUs), one for the real part of the spectrogram and the other for the imaginary part. The GLUs are adopted since they are potential in avoiding the gradient vanishing issue dwelling in deep architectures by providing a linear path for the gradients while retaining non-linear capabilities. Aiming at capturing the long-range dependent features in speech, we place DPTs between the encoder and the decoder of the generator, which contains multi-head attention modules and Bi-directional Gated Recurrent Units (BiGRUs). Moreover, the DPT structure is also merged with multiple one-dimensional convolutional layers in the discriminator of the GAN. Such a design not only improves the speech enhancement performance of GAN by focusing on multiple features of speech, but also reducing the volume of model parameters of GAN. Experimental results suggest that the proposed GAN architecture outperforms the existing benchmark GANs in terms of both objective speech intelligibility and quality with less computational complexity. © 2022 IEEE.