Blockchain-Assisted Lightweight Secure Aggregation in Federated Learning via Trust-Aware Client Selection

Blockchain-enabled Federated Learning (BFL) enhances data privacy and auditability by integrating blockchain technology with the federated learning paradigm. However, prevailing BFL schemes often incur substantial computational overhead in resource-constrained edge computing environments, owing to their reliance on complex cryptographic operations and frequent on-chain interactions. Moreover, robustness issues may arise due to potential man-in-the-middle attacks and the involvement of malicious clients. To this end, this paper proposes a lightweight and trustworthy privacy-preserving secure aggregation framework, named BP-LiteFed. The framework features an efficient ciphertext aggregation scheme that employs Paillier homomorphic encryption to enable secure model parameter aggregation. Additionally, a batch-verifiable digital signature algorithm is incorporated to enhance verification efficiency and substantially reduce the computational overhead on client devices. BP-LiteFed further incorporates a multidimensional trust evaluation strategy to dynamically select high-quality and trustworthy clients prior to training, thereby improving the convergence performance of the global model. Theoretical analysis and experimental results demonstrate that the proposed BP-LiteFed framework offers substantial advantages in various aspects.