DISTRIBUTED MACHINE LEARNING

Research on Federated Learning Node Selection Method Based on Reputation Mechanism

Federated learning enables collaborative model training without sharing raw data, but its deployment is hindered by node heterogeneity, inefficient client selection, and vulnerability to malicious attacks. Traditional selection strategies rely on single indicators and static rules, while existing defenses struggle with complex attack patterns, making it difficult to balance convergence efficiency and security. This thesis investigates node selection and malicious node defense in federated learning and proposes a collaborative optimization framework based on a reputation mechanism and anomaly detection. First, a multi-dimensional dynamic reputation model is constructed using local accuracy, response latency, participation frequency and security behavior, and a dynamic probabilistic node selection strategy based on information entropy is designed to prioritize high-reputation nodes while retaining exploration. Second, a hybrid anomaly detection model combining statistical distance, cosine similarity and DBSCAN clustering is developed, together with a progressive avoidance strategy and adaptive weighted aggregation to suppress malicious influence and enhance fault tolerance. Experiments on MNIST, Fashion-MNIST and CIFAR-10 under IID and Non-IID settings, with varying client scales and malicious node ratios, show that the proposed method improves accuracy, accelerates convergence and significantly strengthens robustness compared with baseline algorithms such as FedAvg and FedProx.