Adversarial Exposure Attack on Diabetic Retinopathy Imagery Grading.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) is a primary cause of global vision impairment.
• Deep Neural Networks (DNNs) are increasingly used for automated DR grading using retinal fundus images (RFIs).
• Camera exposure variations in RFIs can lead to misclassification by DNNs, potentially worsening patient outcomes.

Purpose of the Study:

• To investigate the vulnerability of DNN-based DR grading systems to adversarial attacks targeting image exposure.
• To introduce a novel 'adversarial exposure attack' method for generating natural-looking images that mislead DNNs.
• To assess the effectiveness and image quality of the proposed attack on state-of-the-art DR grading models.

Main Methods:

• Development of a novel adversarial exposure attack technique.
• Validation on a public DR dataset using ResNet50, MobileNet, and EfficientNet models.
• Evaluation of attack success rate and the naturalness of generated images.

Main Results:

• The proposed adversarial exposure attack successfully misled state-of-the-art DNNs for DR grading.
• The generated images maintained high visual quality and natural appearance.
• The attack demonstrated significant transferability across different DNN architectures.

Conclusions:

• DNN-based automated DR grading systems are susceptible to adversarial exposure attacks.
• The developed attack method poses a potential threat to the reliability of current DR diagnostic tools.
• Findings underscore the need for developing exposure-robust DR grading methods to ensure patient safety.