Blockchain-based Federated Learning Utilizing Zero-Knowledge Proofs for Verifiable Training and Aggregation
Elmira Ebrahimi, Michael Sober, Anh-Tu Hoang, Can Umut Ileri, William B. Sanders, Stefan Schulte
Abstract
Federated learning is a distributed learning technique that enables parties to train a model collaboratively without disclosing their local data. To this end, a centralized aggregator collects local models from participating parties and aggregates them to form a global model. However, if parties are malicious, this approach is susceptible to security threats such as model poisoning and global aggregation attacks. Blockchain-based federated learning has been introduced as an alternative to the centralized aggregator to aggregate local models directly on the blockchain. However, employing blockchain-based solutions introduces challenges, including high computational costs and privacy concerns. To mitigate such challenges, this paper leverages zero-knowledge proofs (i.e., zk-SNARKs) to guarantee the privacy and verifiability of local model training and global model aggregation. The proposed framework verifies the local and global models' computational correctness without storing and revealing them on the blockchain. We evaluate our framework by utilizing a real-life dataset, with experimental results demonstrating its viability regarding both computational costs and learning model performance.