Design of Robust Complex-Valued Feed-Forward Neural Networks
Ana Neacsu, Razvan Ciubotaru, Jean‐Christophe Pesquet, Corneliu Burileanu · 2022 30th European Signal Processing Conference (EUSIPCO) · 2022
This paper addresses the problem of designing robust complex-valued neural networks in order to reduce their sensitivity to adversarial perturbations. The robustness is guaranteed by imposing a bound on the Lipschitz constant of the network. We present a new architecture (RCFF-Net), for which we derive tight Lipschitz constant bounds. A constrained learning strategy is then developed to train the proposed structure, while controlling its global Lipschitz constant. The proposed approach is evaluated in an audio signal denoising task. The achieved results demonstrate the effectiveness of the aforementioned design method.