Hierarchical Adversarial Policy Learning for Robust and Explainable Defence in Ophthalmic Artificial Intelligence Systems
Year: 2025
Type: article
Abstract: Background: The increasing reliance on deep learning for ophthalmic diagnostics has raised concerns about the vulnerability of artificial intelligence (AI) systems to adversarial attacks, particularly in retinal imaging. Conventional defense strategies often lack adaptability to evolving threats, highlighting the need for more robust and dynamic protection mechanisms in medical AI. We propose hierarchical adversarial policy learning (HAPL), a novel reinforcement learning framework designed to secure ophthalmic AI systems against adversarial attacks in retinal scans. Methods: Our approach introduces a hierarchical policy architecture comprising a meta-policy controller with explainable attention mechanisms and domain-specific subpolicies that collaboratively mitigate adversarial perturbations. At the top, a meta-policy controller analyses each image and decides how to allocate defense resources. It then activates specialised subpolicies such as adaptive denoising and spectral filtering to apply the right defensive actions based on context. To avoid overfitting to known attacks, we use a regulariser that encourages diversity in strategies. The framework integrates with current ophthalmic AI systems by replacing raw images with purified ones and refining its defences through feedback from diagnostic tools. Results: We implement HAPL using a hierarchical deep neural network trained with adversarial perturbations, showing compatibility with standard workflows and outperforming static defence methods. Experiments on retinal scan datasets confirm its effectiveness in countering both spatial and frequency-domain attacks, setting a new benchmark for medical imaging defence. Conclusion: The proposed method addresses key gaps in medical AI security by combining hierarchical decision-making with domain-specific robustness, providing a scalable, interpretable solution to protect diagnostic accuracy in ophthalmic systems.
Institutions: Taibah University, King Abdulaziz University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4413280607