Paper 2018/679

DeepChain: Auditable and Privacy-Preserving Deep Learning with Blockchain-based Incentive

Jiasi Weng and Jian Weng and Jilian Zhang and Ming Li and Yue Zhang and Weiqi Luo

Abstract

Deep learning technology has achieved the high-accuracy of state-of-the-art algorithms in a variety of AI tasks. Its popularity has drawn security researchers’ attention to the topic of privacy-preserving deep learning, in which neither training data nor model is expected to be exposed. Recently, federated learning becomes promising for the development of deep learning where multi-parties upload local gradients and a server updates parameters with collected gradients, the privacy issues of which have been discussed widely. In this paper, we explore additional security issues in this case, not merely the privacy. First, we consider that the general assumption of honest-but-curious server is problematic, and the malicious server may break privacy. Second, the malicious server or participants may damage the correctness of training, such as incorrect gradient collecting or parameter updating. Third, we discover that federated learning lacks an effective incentive mechanism for distrustful participants due to privacy and financial considerations. To address the aforementioned issues, we introduce a value-driven incentive mechanism based on Blockchain. Adapted to this incentive setting, we migrate the malicious threats from server and participants and guarantee the privacy and auditability. Thus, we propose to present DeepChain which gives mistrustful parties incentives to participate in privacy-preserving learning, share gradients and update parameters correctly, and eventually accomplish iterative learning with a win-win result. At last, we give an implementation prototype by integrating deep learning module with a Blockchain development platform (Corda V3.0). We evaluate it in terms of encryption performance and training accuracy, which demonstrates the feasibility of DeepChain.

Note: We revised the paper by adding a series of experiments for measuring the time cost was conducted.