Towards Trust Driven On-Demand Client Deployment in Federated Learning

Abstract

Federated Learning (FL) systems choose a certain number of clients from each round to take part in the learning. The ability to have more available clients in the learning areas is achieved using containerization technology. However, reliability concerns raise doubts about the trustworthiness of these devices as Docker containers are deployed on them to be able to serve as clients in FL scenarios. Moreover, the default random selection does not take into consideration trust values when selecting clients. Clients who are malicious may poison the learning process or the entire model if they are selected. In order to overcome these issues, we propose in this work that a trust factor must be considered while selecting these clients and deploying models in our architecture. We build a trust framework between the server and its available clients. The trust factor is continuously monitored and updated by checking clients that are not serving successfully the deployed jobs. In addition, it concludes relevant information about the local accuracy changes of each client by applying a two-step verification method to detect any label flipping and random updated weights. The simulations utilize the mobile data challenge dataset. In each round, clients with high trustworthiness are selected. The simulations are compared with the default random selection method, and the centralized model. By assigning trust values to each client, updating this factor, and crediting malicious clients with a low trust factor, the proposed architecture is able to detect these malicious clients while minimizing the number of normal and discarded rounds.