Massif: Interactive Interpretation of Adversarial Attacks on Deep Learning

Deep neural networks (DNNs) are increasingly powering high-stakes\napplications such as autonomous cars and healthcare; however, DNNs are often\ntreated as "black boxes" in such applications. Recent research has also\nrevealed that DNNs are highly vulnerable to adversarial attacks, raising\nserious concerns over deploying DNNs in the real world. To overcome these\ndeficiencies, we are developing Massif, an interactive tool for deciphering\nadversarial attacks. Massif identifies and interactively visualizes neurons and\ntheir connections inside a DNN that are strongly activated or suppressed by an\nadversarial attack. Massif provides both a high-level, interpretable overview\nof the effect of an attack on a DNN, and a low-level, detailed description of\nthe affected neurons. These tightly coupled views in Massif help people better\nunderstand which input features are most vulnerable or important for correct\npredictions.\n