Blockchain-based Monitoring for Poison Attack Detection in Decentralized Federated Learning

To be able to train a model in a way where access to datasets remains private, Federated Learning (FL) was proposed as a machine learning technique to address the privacy challenges by enabling the model to be trained across nodes that hold their datasets locally. Moreover, blockchain-based FL was recently proposed as a distributed FL architecture to achieve decentralized FL. In this setting, the chief that was responsible of the aggregation in FL is eliminated from the learning process. Instead, the workers collaborate amongst themselves to train the global model. This decentralization incurs additional delays and FL-based applications relying on it need to account for it in their deployments. In this work, we study the end-to-end learning completion latency particularly when this system is under attack. Blockchain-based FL systems need to detect targeted/untargeted poisoning attacks, thus, we investigate a realistic decentralized FL process protected against them. The technique that we propose consists in decoupling the mechanisms used in the defenses proposed against these attacks. The operations of the monitoring and detection phases to assess the behavior of the workers are conducted in parallel for an improved decentralized FL deployment. We demonstrate that our technique improved time efficiency, robustness and network scalability. Parallelizing the operations resulted in a decrease in latency over the communication network and a decrease in computation, and consensus delays during the blockchain and FL process.