AgroPatch AutoStrike for adversarial patch attacks on plant disease detection models

Plant diseases threaten food security and cause substantial economic losses, motivating artificial intelligence for automated detection. Although vision transformers (ViTs) have shown promising performance in agriculture, their vulnerability to adversarial patch attacks remains underexplored. Existing patch attacks developed for generic images are less effective for plant disease images. We propose AgroPatch-AutoStrike (APA), an adversarial patch attack framework for ViT-based plant disease detection. APA combines the variance driven feature patch attack (VDFPA) with the adversarial focal loss (AFL) to improve patch placement and emphasize hard to attack categories. On the cassava leaf disease dataset, APA reduces the robust accuracy (RA) of most ViT models to nearly zero using four patches covering 2% of the image. Compared with patch-fool and LaVAN, APA achieves 1-2% lower RA with comparable inference time. These findings reveal robustness gaps in plant disease detection models and support domain specific adversarial robustness evaluation for agricultural AI.