Vulnerability in MobileNet: Case of Diabetes Retinopathy Prediction

Abstract

Convolutional Neural Network (CNN) architectures have demonstrated significant potential in research endeavors, particularly in the realm of medical imaging for the detection of Diabetes Retinopathy (DR). The efficacy of CNN models in diagnosing DR has been validated through the analysis of diverse image modalities such as MRI, OCT, and X-ray images, yielding notably high classification accuracies. Consequently, assessing the robustness of CNN models utilized in medical applications against adversarial attacks is imperative. Ensuring the safety and reliability of diagnostic models in healthcare is crucial as critical healthcare decisions depend on them. However, little research has been conducted on this matter so far. In this study, we assessed the susceptibility of convolution neural networks, particularly MobileNet, to adversarial attacks using the Fast Gradient Sign Method (FGSM). To investigate the vulnerability of the model, we proposed an architecture to attack it and measured the performance metrics before and after the attack. Our experiments demonstrate that even slight perturbations can result in task failure and misclassification in the majority of cases.