Vulnerabilities Of Machine Learning Algorithms To Adversarial Attacks In Medical Images

Machine learning (ML) techniques have gained widespread adoption in medical image diagnosis. However, their susceptibility to adversarial attacks raises concerns regarding their reliability in clinical applications. This study investigates the robustness of two convolutional neural network architectures, ResNet50 and VGG16, against adversarial perturbations introduced via the Fast Gradient Signed Method (FGSM) and DeepFool algorithms. An experimental evaluation was conducted using medical imaging data from the Lung Image Database Consortium (LIDC-IDRI), comprising computed tomography (CT) images annotated for lung lesions. Model performance was quantitatively assessed using metrics derived from confusion matrices, including accuracy, precision, sensitivity, specificity, and F1-score. The results demonstrate a significant vulnerability of both the ResNet50 and VGG16 networks to adversarial manipulations, resulting in considerable degradation of the classification accuracy, particularly under higher perturbation magnitudes. To mitigate these vulnerabil- ities, adversarial training employing FGSM-generated perturbations was implemented, notably enhancing model robustness and classification performance in adversarial settings. The findings confirm the efficacy of adversarial training as a defensive approach against adversarial attacks; however, further research into advanced adversarial defense mechanisms and novel model architectures remains essential to ensure the secure and reliable deployment of ML models in medical diagnostics.