Generating Self-Adaptive One-Pixel Attacks Against MedMNIST and COVIDx CXR-2

Abstract

One-pixel attacks can exploit vulnerabilities in convolutional neural networks to cause targeted misclassifications. However, most attacks, like the one-pixel, do not use an adaptive approach. This research compares the effectiveness of generating one-pixel attacks with and without self-adaptation. Several datasets were attacked to compare the two methods, including CIFAR-10, MedMNIST datasets such as PathMNIST, OCTMNIST, PneumoniaMNIST, BreastMNIST, and BloodMNIST, as well as COVIDx CXR-2. This research shows that self-adaptive differential evolution did not generate better solutions against the MedMNIST and COVIDx CXR-2 datasets compared to the non-adaptive algorithm. However, we show that PathMNIST, OCTMNIST, BreastMNIST, and BloodMNIST are vulnerable to the one-pixel attacks we produce, while PneumoniaMNIST and COVIDx CXR-2 are not.