Sofistikerade och enkla fientliga attackers inverkan på spikande neuronnät

An adversarial attack aims to prevent a machine learning model from achieving its desired goal. It does so by generating adversarial examples by strategically modifying input data, deceiving the models whilst being imperceptible. This challenges the notion of interpretability and explainability of the models. This paper investigates the impact adversarial attacks of varying sophistication have on Spiking Neural Networks (SNNs), which have been trained using Backpropagation Through Time (BPTT). Three well–known adversarial attacks methods are compared—Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Carlini & Wagner (CW)—each representing a level of attack sophistication in ascending order. To explore the relation between model complexity and adversarial attack performance, two SNNs are trained to have different complexity, which in this context means that they are trained on a different amount of classes. The results reveal that although the most sophisticated CW attack had the best overall performance, the least sophisticated attacks, FGSM and PGD, had more consistent performance on both models, while the performance of CW diminished when used on the simpler model. This challenges the assumption that greater attack sophistication necessarily gives improved adversarial attack performance and suggests that, under certain conditions, less sophisticated attacks might have an advantage on less complex models.