A lightweight secret-sharing-based defense against model poisoning attacks in privacy-preserving federated learning

Abstract

As Artificial Intelligence of Things (AIoT) converges with Privacy-Preserving Federated Learning (PPFL), the challenge of defending against model poisoning attacks emerges as increasingly critical. Due to PPFL’s cryptographic protocols for protecting gradient exchanges, detecting poisoning attacks becomes challenging. Traditional defense mechanisms rely on plaintext gradient analysis and thus cannot be directly applied to encrypted gradients. Although homomorphic encryption-based defense schemes enable secure computations on encrypted data, their substantial computational overhead makes them impractical for resource-constrained Internet of Things (IoT) deployments. To address these challenges, we propose a Secret-Sharing-based Defense Framework (SSDF), a lightweight scheme that enables efficient similarity calculations on encrypted gradients under secure aggregation protocols. Our scheme facilitates robust aggregation of encrypted parameters in resource-constrained edge computing environments while protecting the privacy of local model updates. Extensive experiments on four datasets demonstrate that our proposed scheme provides robust defense capabilities against poisoning attacks for both Independent and Identically Distributed (IID) and non-IID data.