Asynchronous Deep Federated Learning Method Based on Local Trainer Data Pollution

The essence of federated learning is that each trainer cooperates with other trainers to achieve strong robustness and high prediction accuracy on the premise of ensuring their own private data security. When some local trainers’ data are polluted, it may cause systemic disasters. How to efficiently detect abnormal servers and dynamically manage abnormal local trainers is currently a challenging research direction. Based on this, this paper proposes the N batch RCF abnormal client federated detection algorithm based on normalized gradient. RCF detects the client abnormal gradient and records it in the abnormal gradient queue. If the N consecutive batch gradients uploaded by the client are abnormal, the communication with the global server is terminated. When the upload gradient does not meet the communication termination conditions, the communication is resumed; extracts sample features based on heterogeneous multi-model fusion to fully learn the spatial features of sample features.