Defense strategies against adversarial attacks

In our previous article, we showed that neural networks are vulnerable to carefully crafted adversarial attacks , i.e. attacks based on minor modifications of the input. While those perturbations are often invisible to a human observer or perceived...