CoPiFL: A collusion-resistant and privacy-preserving federated learning crowdsourcing scheme using blockchain and homomorphic encryption

Federated learning (FL) is one of many tasks facilitated by crowdsourcing. Generally in such a setting, participating workers cooperate to train a comprehensive model by exchanging the trained parameters. While blockchain-based crowdsourcing approaches offer advantages such as data integrity and tamper-proof properties, platform designers must also address potential risks such as data leakage, de-anonymization, and collusion attacks. In this paper, we propose a collusion-resistant and privacy-protected FL crowdsourcing scheme implemented by smart contracts. Our proposed scheme supports fair reward distribution in FL, as well as ensuring data privacy and user privacy using homomorphic encryption and pseudo-identity techniques. The evaluation results show that the time cost and gas consumption of the proposed scheme are realistic in practice. Furthermore, we conduct a comparative evaluation of single worker and multi-workers (crowdsourcing) approaches using Alexnet, Resnet and VGG models, showing that FL with greedy mechanism can significantly accelerate the model training without compromising models' accuracy. Finally, we present a comprehensive security analysis of the proposed approach.