Vision Transformers (ViTs) have recently achieved high performance in retinal Optical Coherence Tomography (OCT) classification studies. However, ViT models continue to face significant challenges, including high computational cost, vulnerability to adversarial attacks, and pronounced sensitivity to preprocessing techniques. This study introduces GreenShield, a unified framework designed to produce an efficient and robust ViT model, referred to as GreenShield-ViT, which outperforms existing lightweight ViT variants in terms of adversarial robustness for retinal OCT classification. The framework integrates a gradient-based block-importance pruning strategy to compress the ViT/B-16 architecture, and adversarial training with proper ImageNet normalization and anti-saturation techniques. The robustness was evaluated using FGSM, PGD, PGD-R3, Transfer-PGD, BIM, and the proposed hybrid attack (FGSM-PGD). The proposed approach achieves an approximately 50% reduction in Floating-Point Operations (FLOPs), inference time, and carbon footprint emissions, while preserving diagnostic accuracy. Experiments conducted using GPU P100 on the OCT-c8, OCTID, and UCSD-3 datasets achieved clean accuracies of 92.5%, 94.78%, and 89.20%, respectively, alongside a significant reduction in attack success rates and improved model calibration. GreenShield-ViT outperformed lightweight ViT variants (Mobile-ViT, ViT-Tiny, ViT-Small) in terms of robustness while offering competitive efficiency. These results suggest its applicability to similar ViT-based medical tasks.
Qasaimeh et al. (Thu,) studied this question.