How we might protect ourselves from malicious AI

New research could make deep-learning models much harder to manipulate in harmful ways, reports Karen Hao.

Hidden danger: Adversarial examples are a class of tiny changes that, when fed into a deep-learning model, cause it to misbehave. They can affect everything from image classifiers to cancer diagnosis systems. Despite their danger, however, they are poorly understood.

A classic example: By adding a little noise to an image of a panda, a system will classify it as a gibbon with nearly 100% confidence. That’s because of the way it detects patterns of pixels.

A possible solution: A new MIT paper suggests a potential fix, but it will involve fundamentally rethinking how AI models are trained in the first place. Read the full story here.