A study on the Impact of Blockchain to Secure Federated Learning

Federated Learning (FL) enables collaborative model training without the need for centralizing data; however, traditional FL frameworks often suffer from significant security and trust chal lenges, including vulnerabilities to data poisoning, inference attacks, and depends on a central aggregator. To overcome these problems, this thesis proposes a Blockchain-Enabled Federated Learning (BCFL) framework that integrates blockchain technology to decentralize model update aggregation, thereby providing an immutable and transparent ledger for recording updates. The framework incorporates the InterPlanetary File System (IPFS) for decentralized storage of model weights and employs a simulation of Zero-Knowledge Proofs (ZKPs) to verify the integrity of submitted updates without compromising privacy. The proposed BCFL framework is evaluated on three image-classification tasks (MNIST, FashionMNIST, and CIFAR-10). Experimental results demonstrate that while the integration of blockchain operations introduces a moderate computational overhead—typically increasing training time by approximately 4% to 24%—the final model accuracy remains comparable to that of standard FL. Our primary goal was to maintain the accuracy even after introducing blockchain, which we have achieved here. Moreover, the blockchain-based approach enhances security by ensuring update traceability and robustness against tampering, with an effective outlier detection mechanism further mitigating potential malicious contributions. This research highlights a crucial trade-off between additional computational costs and the significant gains in transparency, security, and trust in distributed learning environments. The study also identifies limitations, including scalability challenges under real-world network conditions and the need for advanced cryptographic methods to replace the simplified ZKP simulation. Future research directions include the exploration of more efficient consensus mechanisms, the integration of advanced privacy-preserving techniques, and the development of interoperable BCFL systems capable of operating across heterogeneous blockchain networks.