📎 Breaking Linear Classifiers on ImageNet

We can take any arbitrary image (e.g. “panda”) and classify it as whatever class we want (e.g. “ostrich”) by adding tiny, imperceptible noise patterns… But in fact the core flaw extends to many other domains (e.g. speech recognition systems) and, most importantly, also to Linear Classifiers. It is in fact this linear nature that is problematic. And because Deep Learning models use linear functions to build up the architecture, they inherit their flaw.

Carles Bellver Torlà @carlesbellver