Attack-resistant federated learning with statistical propagation methods

Federated learning, an important method in the field of artificial intelligence, offers significant advantages in terms of personal data security by enabling training without directly obtaining data from participants. However, not receiving any data from participants in federated learning can leave the system vulnerable to various attacks. In the literature, methods developed against security vulnerabilities in federated learning have generally focused on low-density malicious participants. However, in cases where there is a high density of malicious participants, these methods may be insufficient and lead to security vulnerabilities. In the thesis study, a federated learning method resistant to attacks has been developed. Using the proposed statistical propagation methods, the power of our method has been emphasized, especially in scenarios where the rate of malicious participants is high. In the study, the performance of the method has been tested in Byzantine attacks and Label Translation attacks. Various experimental studies have clearly demonstrated the effectiveness of the proposed method and its superiority over existing methods. In addition, the proposed method has been compared with the methods in the literature and the improvements provided have been analyzed in detail.