Cycle-to-cycle Variation Enabled Energy Efficient Privacy Preserving Technology in ANN

Abstract

Differential privacy is emerging as an effective solution to achieve privacy protection for the Artificial Intelligence neural network (ANN). However, not only matrix calculations of a neural network but also random noise injection mechanisms for differential privacy consume large power and resources. Traditionally, most privacy protection technologies are software technologies using von Neumann architecture and hardware with extra noise generation circuit unit. In this paper, a memristor based crossbar in-memory computing system is proposed to enable energy efficient privacy preserving technology in ANN. We utilize inherent cycle-to-cycle variations of memristors and apply the proposed variation-based pulse pair method during the weight update process. As a result, the proposed methods realize a machine learning system with privacy protection and show up to 29.24% recognition accuracy improvement with various privacy budget ε.