Security of Approximate Neural Networks against Power Side-channel Attacks

Emerging low-energy computing technologies, in particular approximate computing, are becoming increasingly relevant in key applications. A significant use case for these technologies is reduced energy consumption in Artificial Neural Networks (ANNs), an increasingly pressing concern with the rapid growth of AI deployments. It is essential we understand the security implications of approximate computing in an ANN context before this practice becomes commonplace. In this work, we examine the test case of approximate ANN processing elements (PE) in terms of information leakage via the power side channel. We perform a weight extraction correlation Power Analysis (CPA) attack under three approximation scenarios: overclocking, voltage scaling, and circuit level bitwise approximation. We demonstrate that as the degree of approximation increases the Signal to Noise Ratio (SNR) of power traces rapidly degrades. We show that the Measurement to Disclosure (MTD) increases for all approximate techniques. An MTD of 48 under precise computing is increased to at minimum 200 (bitwise approximate circuit at $\mathbf{2 5 \%}$ approximation), and under some approximation scenarios $\gt1024$. i.e. an increase in attack difficulty of at least x4 and potentially over x20. A relative Security-Power-Delay (SPD) analysis reveals that, in addition to the across the board improvement vs precise computing, voltage and clock scaling both significantly outperform approximate circuits with voltage scaling as the highest performing technique.