On the robustness of mode decomposition deep neural network to adversarial perturbations

Mode decomposition (MD) is an important diagnostic tool in the design and development of novel high-power lasers, especially fiber laser systems. In recent years, deep neural networks (DNNs) have been utilized in MD and have achieved many remarkable results, demonstrating significant industrial potential. Nonetheless, research on adversarial attacks indicates that DNNs can be susceptible to very small input perturbations that are deliberately engineered, and the results in this work reveal that DNNs employed for MD are similarly affected. After the carefully designed adversarial perturbations are added to the input image of a typical DNN solving for MD problem, the accuracy of MD that represented by mean correlation coefficient of 1000 test samples decreases from 0.9965 to 0.9230, reflecting a reduction of 7.38%. It indicates that the robustness to adversarial perturbations should be meticulously considered before the DNN applied to practical scenes involved MD.