Deep learning models are being integrated into a wide range of high-impact,\nsecurity-critical systems, from self-driving cars to medical diagnosis.\nHowever, recent research has demonstrated that many of these deep learning\narchitectures are vulnerable to adversarial attacks--highlighting the vital\nneed for defensive techniques to detect and mitigate these attacks before they\noccur. To combat these adversarial attacks, we developed UnMask, an adversarial\ndetection and defense framework based on robust feature alignment. The core\nidea behind UnMask is to protect these models by verifying that an image's\npredicted class ("bird") contains the expected robust features (e.g., beak,\nwings, eyes). For example, if an image is classified as "bird", but the\nextracted features are wheel, saddle and frame, the model may be under attack.\nUnMask detects such attacks and defends the model by rectifying the\nmisclassification, re-classifying the image based on its robust features. Our\nextensive evaluation shows that UnMask (1) detects up to 96.75% of attacks, and\n(2) defends the model by correctly classifying up to 93% of adversarial images\nproduced by the current strongest attack, Projected Gradient Descent, in the\ngray-box setting. UnMask provides significantly better protection than\nadversarial training across 8 attack vectors, averaging 31.18% higher accuracy.\nWe open source the code repository and data with this paper:\nhttps://github.com/safreita1/unmask.\n