Increasing Adversarial Robustness Around Uncertain Boundary Regions with Amodal Segmentation

Adversarial perturbations in object recognition and image classification tasks impact a learning model's ability to perform accurately and increase the safety risks of deployed machine learning models. During the adversarial example generation process, adversaries approach areas most prone to model uncertainty. Identifying partially occluded items, especially without understanding general object shapes, contributes to significant model uncertainty since object boundaries are not inherently at the forefront of the feature generalization process in deep learning models. Thus, this work aims to reduce model uncertainty surrounding partially occluded boundaries and increase adversarial robustness by augmenting the training dataset with amodal segmentation boundary masks. By observing performance degradation, robust sensitivity, and loss sensitivity, we show how including these masks during training impacts an adversary's ability to generate effective adversarial examples on the versatile MS COCO dataset. Lastly, we observe how including these masks during training influences the performance of adversarial training.