Evasion Attacks on Image Segmentation Models

Image segmentation (in all its variations) plays a crucial role in computer vision and scene understanding systems. Autonomous driving and robotics are examples of technical areas that are based on image segmentation. As in most other cases where analytical models are not available, machine (deep) learning techniques are used to solve segmentation problems. And image segmentation represents one of the most successful examples of using machine learning. However, all machine learning models are subject to adversarial attacks, where the data for them are subjected to specific modifications that aim to change the way machine (deep) learning models work. This paper discusses such modifications at the inference (usage) stage of segmentation models.