Adversary-Robust Graph-Based Learning of WSIs

Abstract

Enhancing the robustness of deep learning (DL) models against adversarial attacks is crucial, particularly in critical domains such as healthcare, where significant financial interests increase the risk of such attacks. Whole slide images (WSIs) are high-resolution, digitized tissue samples used for cancer diagnosis; however, their large size and multiresolution nature make them vulnerable to both image- and graph-level attacks. In this work, we propose a graph-based defense model that utilizes a graph neural network (GNN) to extract features from graph representations of WSIs while integrating a denoising module based on untrained neural network prior (UNNP) and graph signal processing principles, along with a vision transformer (ViT) for adversarial defense. The denoising module leverages overparameterized networks and early stopping to filter out adversarial perturbations, ensuring improved robustness by optimizing the GNN parameters for each input graph independently rather than relying on a pretrained model. This reduces the impact of adversarial noise at both the image level, addressing attacks on WSIs stored in the cloud, and the graph level, countering perturbations introduced during GNN training. Additionally, the ViT module enhances prostate cancer Gleason grading by capturing complex dependencies in the data. We evaluated the model on clean WSI samples, with and without the denoiser, to establish a baseline and assess its effectiveness. The model was then tested against various adversarial attacks, demonstrating a significant improvement in cancer diagnosis accuracy. These results highlight the robustness and efficiency of the proposed approach in mitigating adversarial challenges in medical imaging.