AgroPatch AutoStrike for adversarial patch attacks on plant disease detection models

Area of Science:

• Agricultural AI
• Computer Vision
• Machine Learning Security

Background:

• Plant diseases pose significant threats to global food security and agricultural economies.
• Automated detection systems using artificial intelligence (AI), particularly vision transformers (ViTs), show promise but are susceptible to adversarial attacks.
• Existing adversarial patch attacks are not optimized for the unique characteristics of plant disease imagery.

Purpose of the Study:

• To introduce AgroPatch-AutoStrike (APA), a novel adversarial patch attack framework tailored for ViT-based plant disease detection.
• To evaluate the effectiveness of APA in compromising the performance of ViT models in agricultural applications.
• To underscore the need for domain-specific adversarial robustness testing in agricultural AI.

Main Methods:

• APA integrates the variance driven feature patch attack (VDFPA) for optimized patch placement.
• Adversarial focal loss (AFL) is employed to prioritize and emphasize challenging-to-attack categories.
• The framework was tested on a cassava leaf disease dataset using ViT models.

Main Results:

• APA significantly reduced the robust accuracy (RA) of most ViT models to near zero with minimal image perturbation (4 patches, 2% image coverage).
• APA demonstrated superior or comparable performance to existing methods like patch-fool and LaVAN in reducing RA.
• The attack achieved a 1-2% lower RA compared to other methods while maintaining similar inference times.

Conclusions:

• ViT-based plant disease detection models exhibit significant robustness gaps against targeted adversarial patch attacks.
• APA proves effective in demonstrating these vulnerabilities, even with limited adversarial perturbations.
• Domain-specific adversarial robustness evaluations are crucial for developing reliable AI systems in agriculture.