Detecting chaos in adversarial examples

The puzzling phenomenon of adversarial examples continues to attract significant research within the machine learning community. The confirmation that adversarial examples can arise in natural real-life circumstances has but increased the interest. While several methods have been proposed for both generating adversarial examples and defending against them, in this work we focus on a previous serendipitous discovery indicating that they can be considered as chaotic signals. More specifically, it has been recently shown that measures of chaoticity in the input signal can be used to detect adversarial examples efficiently. In this work, we extend that approach in two aspects, leading to significant improvements in detection accuracy as demonstrated by results obtained in experiments with four datasets and using seven different attack methods.