Does the security of the software need to be updated when faced with the reality of distributed machine learning?

Abstract

Federated Learning (FL) is a promising distributed learning technique for training models on multiple devices without sharing data and maintaining privacy. However, it faces cybersecurity challenges such as data and model poisoning attacks. These attacks compromise data integrity and privacy. Mitigation strategies, such as data filtering and differential privacy, are essential for protecting ML applications. Integrating knowledge of these threats into the software security curricula and providing hands-on experience for students is key to prepare them for the job market. Tools like FedML and FADO facilitate experimenting and testing security mechanisms in the ML domain, thus contributing to a better understanding of vulnerabilities and security countermeasures.