A Security-Oriented Architecture for Federated Learning in Cloud Environments

Federate architectures are able to provide a scalable and shared environment for the distributed training and testing of machine learning applications. However, federated learning architectures shows potential security vulnerabilities in particular to the data poisoning attacks, incoming from participants to the federation. In such a scenario, a malicious participant may inject bad data in order to sabotage the result of the training. Such attacks may provide both a downgrade of the general performance of the learned model and may compromise the fairness of the machine learning application. As such applications are growing in criticality, such learning models must face with security and privacy as well as with scalability issues. The aim of the paper is to improve federated models by providing an architecture base on cloud computing which grants these additional features. The paper also discusses technical details, in particular, the usage of blockchain schemes to provide integrity and homomorphic cryptography for guaranteeing privacy.