Zhichang Guo, Professor, Harbin Institute of Technology

陈冲 副研究员

Adversarial Transferability and Hybrid Geometric Regularization Models in lmage Restoration

November20 (Thursday),15:00-16:00 PM 腾讯会议: 868-294-519

The first part of this talk focuses on the transferability of adversarial attacks for deep image denoising networks. Through systematic experiments on several representative denoising models, we observe that adversarial perturbations crafted for one model often remain effective across different architectures, exhibiting strong adversarial transferability. Building on the geometry of typical sets for high-dimensional Gaussian noise, we interpret—via differential entropy and typical set theory—how “artificial noise injection combined with similar training strategies” induces similar typical set structures in the noise space of different models, which in turn leads to highly transferable adversarial perturbations. Furthermore, we propose a training strategy based on expanded typical sets, termed Typical Set Sampling (TSS), which significantly enhances robustness against various adversarial attacks while maintaining standard Gaussian denoising performance.

The second part centers on image restoration under combined multiplicative noise and blur. To address the staircase artifacts, speckle residues, and edge oversmoothing often observed in traditional total variation and single-geometry regularization, we develop a hybrid geometric regularization model that couples minimal surface information with higher-order curvature information, and provide it with a clear surface-geometric interpretation. This enables the model to better preserve edges and fine structures while effectively smoothing noise. On the numerical side, we design first- and second-order unconditionally energy-stable time discretization schemes within an anisotropic diffusion and RSAV framework. Extensive experiments on images degraded by multiplicative noise and blur demonstrate the superior overall performance of the proposed approach in terms of PSNR, SSIM, and visual quality. Overall, the talk presents recent advances in image restoration models from the dual perspectives of adversarial robustness and geometric regularization, spanning theoretical analysis and numerical algorithms.