Fuzzy analysis for consensus in federated learning with simulated heuristic attacks

A large number of objects participating in the voting can be an advantage as well as a disadvantage. In the case of decentralized federated learning, adding the model to the aggregation is preceded by a vote. The choice of voters and their results can be falsified through various attacks such as dataset poisoning. In this paper, we propose a fuzzy consensus analyzing the results of individual voters regarding the aggregation of a given model. The consensus is based on a fuzzy controller that selects the most reliable models for aggregation. For this reason, it uses image-modifying heuristics and quick evaluations of incoming results. If a decision is made that a selected client is unreliable several times, it is blocked to reduce the number of performed operations. The proposed system was tested on selected tasks related to image classification. The results were discussed and compared to evaluate the proposed system.