A Formal Scheme of Fault Injection on Coherent Integrated Photonic Neural Networks

Abstract

Based on Mach-Zehnder interferometers (MZIs) coherent integrated photonic neural networks (PNNs) may provide a promising solution for the realization of deep learning with low power consumption, low latency, and ultra-high speed. Adversarial attacks have been widely confirmed to be a serious threat to deep learning. This has led to a large amount of studies in this direction of the electronic domain, including input attacks and inject faults for weights. In this paper, focusing on the phases in the linear operation unit of PNNs, a phase gradient attack (PGA) scheme based on the phase gradient sorting of the MZI-arrays and injecting disturbances along the gradient direction is proposed for the first time. The simulation results indicate that even with weak-intensity PGA, it is almost impossible for PNNs to perform the classification inference. Furthermore, taking into account the effects of fabrication-process variations (FPV) and thermal crosstalk in MZI-arrays that lead to tuning phase deviation in practical application, we systematically analyzed the validity of proposed scheme on the PNNs with phase uncertainties. Specifically, we tested the impact of injecting faults by compressing the number of attacked phase angles to 3, 5, and 7, respectively. The experiment results show that injection attack based using PGA on PNNs trained with Gaussian datasets would reduce classification accuracy to 27.97%, 15.47%, and 8.91% for corresponding cases.