Efficient Privacy-Preserving Machine Learning for Blockchain Network

Abstract

A blockchain as a trustworthy and secure decentralized and distributed network has been emerged for many applications such as in banking, finance, insurance, healthcare and business. Recently, many communities in blockchain networks want to deploy machine learning models to get meaningful knowledge from geographically distributed large-scale data owned by each participant. To run a learning model without data centralization, distributed machine learning (DML) for blockchain networks has been studied. While several works have been proposed, privacy and security have not been sufficiently addressed, and as we show later, there are vulnerabilities in the architecture and limitations in terms of efficiency. In this paper, we propose a privacy-preserving DML model for a permissioned blockchain to resolve the privacy, security, and performance issues in a systematic way. We develop a differentially private stochastic gradient descent method and an error-based aggregation rule as core primitives. Our model can treat any type of differentially private learning algorithm where non-deterministic functions should be defined. The proposed error-based aggregation rule is effective to prevent attacks by an adversarial node that tries to deteriorate the accuracy of DML models. Our experiment results show that our proposed model provides stronger resilience against adversarial attacks than other aggregation rules under a differentially private scenario. Finally, we show that our proposed model has high usability because it has low computational complexity and low transaction latency.

Blockchain-Empowered Mobile Edge Intelligence, Machine Learning and Secure Data Sharing

“…In LearningChain, they developed a differential privacy mechanism for the local gradient computing process to protect the privacy of individual data providers, and a l-nearest aggregation scheme to defend against Byzantine attacks in the global gradient's aggregation process. In the next year, Kim et al [72] pointed out that the LearningChain system has several limitations such as low computation efficiency, zero support on non-deterministic function computations, and weak privacy preservation. To revolve those issues in a systematic way, they developed an improved distributed machine learning model for permissioned blockchains.…”

Synergy of Edge Intelligence and Blockchain: A Comprehensive Survey

“…As such, the DPPDL further ensures fairness during the data download and upload processes of the model training. Differently, H. Kim et al in [132] take a new error-based aggregation rule for preventing attacks by adversarial nodes in the aggregation process. This rule can aggregate the nearest learning results utilizing low errors saved in the immutable ledger as a log.…”

How to Democratise and Protect AI: Fair and Differentially Private Decentralised Deep Learning

“…However, each access to the training data causes information leakage of the training data and thus incurs privacy loss from the overall privacy budget . To apply DPGAN and DPSGD to the distributed/decentralized settings, we follow recent work [30], [31], [32], [33] to conduct local gradient computation and calculate privacy on a per party-basis, where each party individually applies moments accountant [18] to keep track of the spent privacy budget. Each party repeats the local training process until the allocated privacy budget is used up.…”