Abstract:Modern, state-of-the-art Convolutional Neural Networks (CNNs) in computer vision have millions of parameters. Thus, explaining the complex decisions of such networks to humans is challenging. A technical approach to reduce CNN complexity is network pruning, where less important parameters are deleted. The work presented in this paper investigates whether this technical complexity reduction also helps with perceived explainability. To do so, we conducted a pre-study and two human-grounded experiments, assessing the effects of different pruning ratios on CNN explainability. Overall, we evaluated four different compression rates (i.e., CPR 2, 4, 8, and 32) with 37 500 tasks on Mechanical Turk. Results indicate that lower compression rates have a positive influence on explainability, while higher compression rates show negative effects. Furthermore, we were able to identify sweet spots that increase both the perceived explainability and the model's performance.
Abstract:Adversarial examples are carefully crafted attack points that are supposed to fool machine learning classifiers. In the last years, the field of adversarial machine learning, especially the study of perturbation-based adversarial examples, in which a perturbation that is not perceptible for humans is added to the images, has been studied extensively. Adversarial training can be used to achieve robustness against such inputs. Another type of adversarial examples are invariance-based adversarial examples, where the images are semantically modified such that the predicted class of the model does not change, but the class that is determined by humans does. How to ensure robustness against this type of adversarial examples has not been explored yet. This work addresses the impact of adversarial training with invariance-based adversarial examples on a convolutional neural network (CNN). We show that when adversarial training with invariance-based and perturbation-based adversarial examples is applied, it should be conducted simultaneously and not consecutively. This procedure can achieve relatively high robustness against both types of adversarial examples. Additionally, we find that the algorithm used for generating invariance-based adversarial examples in prior work does not correctly determine the labels and therefore we use human-determined labels.
Abstract:The vulnerability of deep neural networks against adversarial examples - inputs with small imperceptible perturbations - has gained a lot of attention in the research community recently. Simultaneously, the number of parameters of state-of-the-art deep learning models has been growing massively, with implications on the memory and computational resources required to train and deploy such models. One approach to control the size of neural networks is retrospectively reducing the number of parameters, so-called neural network pruning. Available research on the impact of neural network pruning on the adversarial robustness is fragmentary and often does not adhere to established principles of robustness evaluation. We close this gap by evaluating the robustness of pruned models against L-0, L-2 and L-infinity attacks for a wide range of attack strengths, several architectures, data sets, pruning methods, and compression rates. Our results confirm that neural network pruning and adversarial robustness are not mutually exclusive. Instead, sweet spots can be found that are favorable in terms of model size and adversarial robustness. Furthermore, we extend our analysis to situations that incorporate additional assumptions on the adversarial scenario and show that depending on the situation, different strategies are optimal.
Abstract:Adversarial machine learning, i.e., increasing the robustness of machine learning algorithms against so-called adversarial examples, is now an established field. Yet, newly proposed methods are evaluated and compared under unrealistic scenarios where costs for adversary and defender are not considered and either all samples are attacked or no sample is attacked. We scrutinize these assumptions and propose the advanced adversarial classification game, which incorporates all relevant parameters of an adversary and a defender in adversarial classification. Especially, we take into account economic factors on both sides and the fact that all so far proposed countermeasures against adversarial examples reduce accuracy on benign samples. Analyzing the scenario in detail, where both players have two pure strategies, we identify all best responses and conclude that in practical settings, the most influential factor might be the maximum amount of adversarial examples.