Adversarial Testing of Image Segmentation Models

Image segmentation is one of the most frequently addressed tasks in image processing. However, image segmentation models, like any other deep learning models, are susceptible to adversarial attacks — special data modifications at different stages of the standard machine learning pipeline that prevent the model from functioning correctly and pose a problematic issue for practical applications of deep learning models. This paper examines so-called evasion attacks, where input data is modified during the execution (inference) stage. The article presents an original tool — Segmentation Robustness Framework (SRF), designed for testing the robustness of segmentation models against digital adversarial attacks.