A Cyber Defensive Model for Diabetic Retinopathy Using Bit-Plane Methods

Diabetic retinopathy (DR) remains a critical concern for individuals with diabetes, as it can progressively lead to vision impairment or even blindness if left untreated. Accurately identifying the stages of DR from retinal images poses a significant challenge, especially when considering the vulnerability of current predictive models to adversarial attacks, which can manipulate input data to deceive the model into making incorrect predictions. This study addresses this concern by introducing a robust cyber defensive model specifically designed to counter these attacks. The approach utilizes three transfer learning-based VGG16 models (M1, M2, and M3), each trained on different retinal image datasets representing various stages of DR. In addition, bit-plane slicing and bit-plane merging techniques were incorporated to enhance feature detection and bolster the models’ defenses against adversarial interference. These techniques break down images into distinct bit-planes, allowing for more refined analysis and the ability to counter subtle changes introduced by adversarial attacks. During testing, adversarial examples were generated using the FGSM method to assess the models’ resilience. M2 and M3 demonstrated superior robustness in comparison to M1, achieving AUC scores of 0.7222 and 0.77 against targeted adversarial examples, and 0.7848 and 0.7368 against untargeted attacks, respectively. These findings underscore the potential of combining transfer learning with advanced techniques like bit-plane slicing and merging to improve the security, accuracy, and reliability of diabetic retinopathy diagnosis systems in the face of adversarial threats.