Metasurface-based diffractive neural networks for security-enhanced image processing [Invited]

Metasurface-based diffractive computing enables efficient and broadband information processing at the speed of light. While such all-optical systems support high-throughput operation, they risk information security since the processing and the results outputs are typically performed in plaintext. As a step to overcome this issue, we numerically investigate a pluggable diffractive neural network (DNN) scheme for enhancing optical processing security. Here, DNNs are trained to yield correct outputs only when a unique authorized metasurface key layer is inserted at the specific position. We validate this approach in both single-task and multitask scenarios. Our approach demonstrates an enhanced processing security paradigm without compromising computational capacity and speed.