Decentralized Role Rotation Privacy Protection in Federated Learning

In decentralized federated learning frameworks, the absence of a logically centralized trusted authority makes it challenging to verify the trustworthiness of each participating node. This limitation makes the system vulnerable to privacy leakage caused by a single attacker orchestrating multiple colluding identities. This study proposes a blockchain-based role rotation privacy protection method for decentralized federated learning. A dynamic role assignment mechanism is designed to adjust the roles of miners, validators, and clients in real time based on their reputation and contributions. Meanwhile, blockchain technology is utilized to record the global model and validation processes, constructing a tamper-resistant and traceable model optimization and validation mechanism. During the training process, a reward and punishment system is implemented to incentivize nodes to actively participate in collaboration while gradually marginalizing malicious nodes. Experimental results demonstrate that the proposed method can accurately identify malicious nodes and prevent them from continuing to participate in critical role assignments within the federated learning process. This ensures a secure and efficient distributed training process, thereby enhancing privacy protection.