TrustFL: Robustness-Enhanced Privacy-Preserving Federated Learning Based on Blockchain

As a booming technology driving enterprise innovation, artificial intelligence (AI) has garnered significant interest in academia and industry. However, the use of AI raises a significant concern on “trust”. Trusted AI systems require the implementation of crucial safeguards such as privacy preservation, robustness, transparency, and decentration. To construct trusted AI systems, we propose a robustness-enhanced privacy-preserving federated learning scheme based on blockchain (TrustFL). Our TrustFL integrates Federated Learning (FL) with fully homomorphic encryption to protect the participants’ privacy for training models and employs the cosine similarity-based aggregation methods to achieve robustness for both Independently Identically Distribution (IID) and non-IID data without compromising privacy. Additionally, TrustFL leverages blockchain to record the trust levels of both the aggregation server and participating clients on a distributed ledger. It further incorporates the Algorand consensus protocol to securely generate and validate the aggregated results. We formally prove the privacy protection and robustness of TrustFL. Our experimental results demonstrate that TrustFL achieves robustness and efficiency in handling both IID and non-IID data, while outperforming existing approaches in terms of model accuracy and computational efficiency.