Assessing the Impact of Dynamic Backdoor Attacks on Federated Learning Models

With rapid advances in data-driven techniques, Federated Learning (FL) has evolved as an attractive approach for collaborative model training without compromising the privacy of individual data sources. However, this decentralized structure poses substantial security concerns, especially in the form of backdoor attacks. Backdoor attacks, a subtype of data poisoning attacks, take advantage of the confined control of the server, allowing adversarial participants to inject trigger patterns into local models, compromising the integrity and performance of the global model. A stronger version, called a dynamic backdoor attack, presents an intensive threat, where the adversary uses dynamic trigger patterns across the data inputs, making recognition and prevention even more challenging. In this work, the sensitivity of FL models to such dynamic backdoor attacks and their impact on global model performance is evaluated. Using the two benchmark datasets, MNIST and CIFAR-10, the state-of-the-art methods are tested against distributed backdoor attackers under Independent and Identically Distributed (IID) and non-IID data scenarios to demonstrate the stealth and strength of dynamic triggers, constituting a substantial threat to current FL models.