GGPD: Defending Against Phase Gradient Attack on Photonic Neural Networks

Integrated photonic neural networks (PNNs) offer a novel accelerated framework for information processing systems and are gradually influencing the future direction of Artificial Intelligence (AI). With silicon-based photonic platforms increasingly deployed in security-sensitive areas, the security of PNNs has become more critical. Attacks targeting the gradient ordering of photonic components and injecting disturbances along the gradient direction have significantly threatened the security of PNNs. However, defense mechanisms against such attacks remain underexplored. In this paper, a defense strategy for PNNs based on Mach-Zehnder interferometers (MZIs)—the gradient-guided phase defense (GGPD) technique—is proposed for the first time. By filtering the vulnerable phases along the learning direction, this method effectively mitigates the impact of attacks targeting high-risk photonic components. Additionally, a multi-stage defense (MSD) framework is introduced, which increases the number of target phases for GGPD. Experimental results demonstrate that, under attacks targeting 40 vulnerable phases, the proposed method achieves 92.21% and 74.70% accuracy on the MNIST and Fashion-MNIST datasets, respectively.