Evaluating Adversarial Robustness of Detection-based Defenses against Adversarial Examples

Machine Learning algorithms provide astonishing performance in a wide range of tasks, including sensitive and critical applications. On the other hand, it has been shown that they are vulnerable to adversarial attacks, a set of techniques that violate the integrity, confidentiality, or availability of such systems. In particular, one of the most studied phenomena concerns adversarial examples, i.e., input samples that are carefully manipulated to alter the model output. In the last decade, the research community put a strong effort into this field, proposing new evasion attacks and methods to defend against them. With this thesis, we propose different approaches that can be applied to Deep Neural Networks to detect and reject adversarial examples that present an anomalous distribution with respect to training data. The first leverages the domain knowledge of the relationships among the considered classes integrated through a framework in which first-order logic knowledge is converted into constraints and injected into a semi-supervised learning problem. Within this setting, the classifier is able to reject samples that violate the domain knowledge constraints. This approach can be applied in both single and multi-label classification settings. The second one is based on a Deep Neural Rejection (DNR) mechanism to detect adversarial examples, based on the idea of rejecting samples that exhibit anomalous feature representations at different network layers. To this end, we exploit RBF SVM classifiers, which provide decreasing confidence values as samples move away from the training data distribution. Despite technical differences, this approach shares a common backbone structure with other proposed methods that we formalize in a unifying framework. As all of them require comparing input samples against an oversized number of reference prototypes, possibly at different representation layers, they suffer from the same drawback, i.e., high computational overhead and memory usage, that makes these approaches unusable in real applications. To overcome this limitation, we introduce FADER (Fast Adversarial Example Rejection), a technique for speeding up detection-based methods by employing RBF networks as detectors: by fixing the number of required prototypes, their runtime complexity can be controlled. All proposed methods are evaluated in both black-box and white-box settings, i.e., against an attacker unaware of the defense mechanism, and against an attacker who knows the defense and adapts the attack algorithm to bypass it, respectively. Our experimental evaluation shows that the proposed methods increase the robustness of the defended models and help detect adversarial examples effectively, especially when the attacker does not know the underlying detection system