AMBER: Robust Federated Learning Based on Client Verification

The datasets used in this study serve as the experimental foundation to evaluate the effectiveness of the proposed AMBER framework in defending against Local Chained Attacks (LCAs) in Federated Learning (FL). Specifically, three image classification datasets are employed: MNIST (handwritten digit recognition), Car10 (10-class car classification), and Car100 (100-class car classification). These datasets are selected to cover diverse task complexities (from simple digit recognition to fine-grained car categorization) and are utilized to test AMBER\u2019s robustness across multiple scenarios, including varying attack types, model architectures, and Non-IID (non-independent and identically distributed) data distributions. By leveraging these datasets, the experiments validate that AMBER achieves superior defense performance with low overhead compared to existing approaches, ensuring comprehensive verification of data integrity, input consistency, and computational integrity in FL systems.