Decentralized Federated Learning Loop with Constrained Trust Mechanism.

Federated learning has made it possible to introduce parallel training of deep neural networks by multiple users. The use of model aggregation contributes to the generalization of it, although there is a possibility of attacks. An example of this is dataset poisoning. Hence, in this research paper, we propose the introduction of a constrained trust mechanism for individual clients. In addition, a decentralized approach makes it possible to increase the effectiveness of the training process by removing the server and reducing the risk of an attack on the transmitting model. The proposed modification of federated learning was subjected to performance tests and compared with other known solutions. The obtained results indicate an increase in safety and accuracy.