Anyone know how I can leverage NuPIC to perform adversarial patch detection? That is, how can anomalous adversarial patches be identified within images? Concretely, the goal is to draw bounding boxes around suspected adversarial patches. Thank you.
Hey @spatel, since this is image anomaly detection I don’t think NuPIC is equipped to help.
NuPIC is great for continuous learning and anomaly detection in sequential data. A touchstone example is the ‘Hot Gym’, where a stream of energy consumption values from the gym are fed into a NuPIC model. The model quickly learns the daily and weekly patterns, and find rare anomalies where these patterns are disrupted.
1 Like
Ok that gets it closer, though you’ll still need a way to encoder each sequential image into a binary vector for compatibility with the SP/TM. I haven’t done this myself, though there have been some discussions on the topic. By searching ‘image encoding’ on the forum I found this for instance: