Experimental Evaluation of a Convolutional Neural Network Classifier for Image Recognition Under Adversarial Attacks

Within Artificial Intelligence, Convolutional Neural Networks (CNNs) are widely used for computer vision tasks including image classification, object detection, medical diagnosis, among others. The main objective of this research is to evaluate CNN-based image classification models against adversarial attacks by comparing a proposed model architecture with existing ones from the scientific literature. We conducted an analytical empirical investigation using a quasi-experimental method to develop a prototype. The methodology incorporated: (1) systematic observation of relevant scientific articles, and (2) quantitative evaluation of prototype performance. The study focused on developing a vehicle classification prototype for identifying cars and trucks, specifically evaluating its robustness against adversarial attacks. Three attack methods (FGSM, PGD, and BIM) were implemented to generate perturbed examples and analyze their impact on model performance. Comparative testing between clean and adversarial data revealed the model’s vulnerability to such threats. The results demonstrate these attacks’ ability to generate effective adversarial noise without significantly altering the images’ semantic structure.