Proof of Privacy-Preserving Machine Learning: A Blockchain Consensus Mechanism with Secure Deep Learning Process

As the most popular consensus mechanism in blockchain applications, Proof of Work (PoW) requires significant computational resources. Except establishing the miners' rights for ledger maintenance, these huge amounts of computational power are used in nothing but a meaningless way. To mitigate the resource waste in PoW, the Proof of Deep Learning (PoDL) is proposed. This consensus approach requires miners to train deep learning models instead of performing meaningless hash computations. However, most existing work can not ensure the confidentiality of models and data, which negatively affects the interests of model requesters. To address this issue, this paper proposes a blockchain consensus called Proof of Privacy-Preserving Machine Learning (PPPML). This mechanism preserves the confidentiality of datasets by facilitating privacy-preserving model training and validation, thus safeguarding the data privacy and model ownership for the model requester. To balance security and efficiency, we employ a hybrid training approach to cope with the high computational demands in encrypted deep learning. The simulation results show that the PPPML is feasible to be used in the blockchain consensus while protecting the privacy of data and models, and achieving comparable accuracy in the non-privacy-preserving setting.