Robustness criterion for diffractive deep neural networks using normalized cutoff frequency.

Robustness of diffractive deep neural networks (D2NNs) as a key performance for the deployment in real-world remains insufficiently elucidated. We propose a criterion to evaluate the robustness of D2NN using the normalized cutoff frequency (NCF), and reveal that the robustness is determined by the spatial frequency propagation characteristics of the light field. We demonstrate that despite differences in D2NN architectures, the same robustness of D2NNs is attributed to equal NCF values in classification and regression tasks. This principle is physically interpretable and widely applicable to linear and nonlinear D2NNs. The relationship between the robustness of D2NN and the spatial frequency bandwidth of the optical field is discovered, enabling efficient and accurate prediction of the robustness of large-scale D2NN models.