DeepFool: A Simple and Accurate Method to Fool Deep Neural Networks

The paper studies the well-known instability of state-of-the-art deep neural networks, which achieve strong image-classification accuracy but can be flipped by small, deliberately sought perturbations of their inputs. Observing that no effective technique existed to accurately quantify this robustness on large-scale datasets, the authors propose DeepFool, an algorithm that efficiently computes the perturbations required to fool a deep classifier, using the size of those perturbations as a reliable robustness measure.

Through extensive experiments, DeepFool is shown to compute adversarial perturbations more accurately and efficiently than prior methods, and the perturbations it finds can in turn be used to make classifiers more robust. By providing an accurate, efficient way to quantify robustness on large-scale datasets, it addresses a gap the authors identify in evaluating and hardening deep classifiers against adversarial perturbations.