Probabilistic Free Riding Attack in Federated Learning

Federated Learning (FL) has revolutionized how we approach machine learning by allowing distributed clients to train models collaboratively while maintaining privacy. However, this decentralized paradigm is vulnerable to free rider attacks, in which malicious clients take advantage of the system without contributing to the learning process. In this paper, we examine the impact of probabilistic free rider attacks on FL, which involve attackers injecting random parameters into the model depending on varying attack probabilities (α) and attacker percentages. We run a 100-round simulation with 40 clients in every round. We evaluate the system under attack probabilities of α = 0.5, 0.7, and 1.0, as well as attacker percentages of 10% and 30%. Our evaluation focuses on several essential parameters, such as average training accuracy and loss of the local models, as well as global model test accuracy, precision, recall, and F1 score. Our hypothesis that increasing attack probabilities and percentages lead to a considerable degradation in system performance is supported by the results. This study adds to the corpus of knowledge by quantifying the consequences of attack variability, paving the way for future defense mechanisms against such adversarial threats in FL systems.