Training Node Screening in Decentralized Trusted Federated Learning

The emergence of federated learning has to some extent solved the current problems of privacy protection of terminal data and the processing technology of massive data. However, its centralized architecture still has problems such as limited access and high establishment cost, so the trend of decentralization is inevitable. Although decentralized federated learning architecture circumvents the drawbacks of centralized structure, it also loses the convenience of third-party supervision. Therefore, to address the problem of missing supervision mechanisms for worker node training behavior in decentralized federated learning architecture, this paper proposes a backdoor-based supervision mechanism for arithmetic node training behavior. The mechanism can be applied to general classification tasks. Experiments revealed that this mechanism can accurately assess the training behavior of worker nodes while maintaining the accuracy of the original task. In addition, this paper proposes a rotation scheme for the watermarked datasets involved and gives a corresponding replacement prediction method, which further ensures that the training behavior of arithmetic nodes can be quantified completely by predicting and replacing the watermarked datasets, aiding the arithmetic party training operation of the behavior monitoring mechanism.