Testing the limits: exploring adversarial techniques in AI models

The rising adoption of artificial intelligence and machine learning in critical sectors underscores the pressing need for robust systems capable of withstanding adversarial threats. While deep learning architectures have revolutionized tasks such as image recognition, their susceptibility to adversarial techniques remains an open challenge. This article evaluates the impact of various adversarial methods, including the fast gradient sign method, projected gradient descent, DeepFool, and Carlini & Wagner, on five neural network models: a fully connected neural network, LeNet, Simple convolutional neural network (CNN), MobileNetV2, and VGG11. Using the EVAISION tool explicitly developed for this research, these attacks were implemented and analyzed based on accuracy, F1-score, and misclassification rate. The results revealed varying levels of vulnerability across the tested models, with simpler architectures occasionally outperforming more complex ones. These findings emphasize the importance of selecting the most appropriate adversarial technique for a given architecture and customizing the associated attack parameters to achieve optimal results in each scenario.