MC-DQE FL: Robust Federated Learning Framework Based on Multi-Criteria Data Quality Evaluation

In the rapid development of artificial intelligence, data silos and strict privacy policies hinder cross-institutional data sharing. Federated Learning (FL) offers a solution by keeping data local while transmitting model parameters. However, evaluating the quality of non-IID and potentially malicious data remains challenging. This paper proposes a Federated Learning aggregation framework based on multi-criteria data quality evaluation (MC-DQE FL), which integrates similarity metrics such as Euclidean, cosine, and Manhattan distances to assess data quality. An adaptive benchmark update strategy further enhances evaluation robustness. In terms of privacy protection, MC-DQE FL incorporates differential privacy techniques by adding relatively small noise to the intermediate parameters, thereby protecting data privacy to some extent. Experiments on CIFAR-10 and FMNIST datasets show that MC-DQE FL outperforms existing methods in accuracy and attack resilience.