Privacy-Preserving Backdoor Attacks Mitigation in Federated Learning Using Functional Encryption

Federated Learning (FL) revolutionizes collaborative machine learning by decentralizing model training across clients with locally generated data. This approach is particularly useful in scenarios involving private data, allowing organizations to collaborate on model training while preserving data privacy. However, recent research has exposed vulnerabilities in FL, particularly susceptibility to adversarial attacks, where malicious clients can manipulate their local training to compromise the global model. This study addresses the challenge of adversarial attacks in FL by proposing a novel defense framework. The framework includes an encryption mechanism for gradient updates to protect against inquisitive or malicious servers and a resilient filtering mechanism to precisely detect and mitigate malicious participants. Extensive experiments and systematic analysis demonstrate the superior efficacy of the proposed defense mechanism compared to contemporary state-of-the-art techniques in mitigating backdoor attacks in federated learning scenarios. The contributions of this work include an effective backdoor removal without compromising model performance, a gradient-specific encryption mechanism for data confidentiality, and a filtering approach enhancing anomaly detection in federated learning environments.