Federated learning (FL) has emerged as a popular paradigm for collaborative model training across decentralised data clients while preserving data privacy. However, FL is inherently vulnerable to poisoning attacks. As these attacks grow more sophisticated, various defence mechanisms are proposed to mitigate the threats. Most existing defences adopt a single perspective on Byzantine client detection resulting in both false positives and false negatives. We propose FLgym, a two-stage framework for Byzantine-resilient FL that integrates three components: a model similarity-based detection mechanism, a validation mechanism based on similarity estimation of clients’ local data, and a weight recovery mechanism for identified Byzantine clients. Extensive experiments show that FLgym consistently outperforms state-of-the-art baselines achieving the highest model accuracy, true positive rate of 90.95%, and lowest false positive rate of 6.3%.

We address the challenges in Federated Learning (FL) amid client mobility by proposing MOBILE, a framework that optimizes client selection and resource allocation while ensuring sufficient client interactions per round. MOBILE leverages historical mobility data instead of focusing on conventional energy constraints, formulating client selection as a regularized Mixed-Integer Quadratic Programming (MIQP) problem. By selecting clients with higher expected returns, MOBILE identifies reliable participants while minimizing wasted bandwidth. Our framework increases participation of clients likely to succeed in training tasks and is orthogonal to existing FL algorithms, introducing minimal overhead while preserving user incentives. Experiments on benchmark mobility and FL datasets demonstrate MOBILE’s superiority in model accuracy and resource utilization compared to baseline approaches.

Decentralized federated learning (DFL) has gained popularity due to its robustness and elimination of centralized coordination requirements. In this paradigm, clients actively participate in training by exchanging models with neighboring nodes in their network. However, DFL introduces significant overhead in both training and communication costs. While existing methods focus primarily on reducing communication costs, they often overlook training efficiency and the challenges of data heterogeneity. We address these limitations by introducing DA-DPFL, a novel sparse-to-sparser training scheme that initializes with a subset of model parameters which progressively decrease during training through dynamic aggregation. This approach substantially reduces energy consumption while preserving adequate information during critical learning periods. Our experimental results demonstrate that DA-DPFL significantly outperforms DFL baselines in test accuracy while achieving up to 5x reduction in energy costs. We provide theoretical convergence analysis that validates the applicability of our approach in decentralized and personalized learning contexts. The code is available at: https://github.com/EricLoong/da-dpfl.

Distributed Machine Learning (DML) at the edge of the network involves model learning and inference across networking nodes over distributed data. One type of model learning could be the delivery of predictive analytics services to formulate intelligent environments, however, those environments heavily rely on real-time inference and are significantly influenced by changes in the underlying data (concept drifts). Moreover, the quality of service and availability in DML environments are directly tied to each node’s reliability, since such environments are highly susceptible to the impact of node failures. Even if such challenges can be tackled with distributed resilience mechanisms, their effectiveness and efficiency, due to concept drifts, should be maintained to ensure continuous and sustained quality of service. DML systems operate in dynamic environments, thus, they require their models to be updated according to the novel trends embedded in the new data they encounter. We, therefore, introduce several model maintenance mechanisms to ensure resilient DML systems in the long term when concept drifts emerge. We provide a comprehensive experimental evaluation of our resilience maintenance mechanisms over synthetic and real data showcasing their importance and applicability in edge learning environments.