AdaptiveShield: Dynamic Defense Against Decentralized Federated Learning Poisoning Attacks

Federated learning allows decentralized devices to collaboratively train a shared model while keeping data local, enhancing the privacy and security of the training process. However, it is vulnerable to poisoning attacks, where malicious participants inject false data to corrupt the global model. To address this, we propose AdaptiveShield, a dynamic hybrid defense approach designed to protect decentralized federated learning against such attacks. AdaptiveShield employs dynamic detection strategies that consider multiple risk factors to assess the maliciousness index and dynamically adjust the detection thresholds, which is able to adapt to various attack scenarios. In addition to attack detections, AdaptiveShield minimizes the negative impact on the global model from missed attackers by dynamically adjusting hyperparameters, thereby enhancing the robustness of the defense. It also dissociates user identities from their uploaded local models through a hierarchical shuffle mechanism, providing an extra layer of privacy protection for both the users and their local models. We evaluate AdaptiveShield across various experimental environments, attack settings, and datasets, demonstrating that it outperforms state-of-the-art approaches by achieving over 0.1 improvement in training accuracy while incurring negligible time overhead.