Enhancing security and integrity in federated learning using blockchain and decentralized storage

Abstract

This work develops a Blockchain-Enabled Federated Learning (BCFL) framework to address some of the major concerns in federated learning: the use of a central server, the lack of transparency/traceability during the training process and the risk of a malicious participant leaking private information. By utilizing the blockchain as an intermediary to aggregate all the local models trained at each participants location, the BCFL framework ensures the integrity of the aggregated model, provides a transparent record of all transactions and prevents tampering. In addition, this framework utilizes the Interplanetary File System (IPFS) to allow nodes to store their own model parameters locally and provides a decentralized method of managing those parameters. The system uses a lightweight hash-based verification mechanism that allows model updates to be verified using their Content Identifiers (CIDs) on the blockchain, without sending the raw parameters to a central authority. Tests using the MNIST, Fashion-MNIST, and CIFAR-10 datasets show that the BCFL framework improves the accuracy of the Global model in the presence of a 20% label-flipping data poisoning attack by 31.65% to 42.40% compared to standard FL baselines without outlier defense. The integration of blockchain consensus and IPFS storage causes a predictable computational training latency overhead of around 7.5%-8.2%. Yet, the framework achieves baseline predictive accuracy (on average 88.49% on CIFAR-10), which indicates a practical and quantifiable trade-off between computing and decentralized security.