Abstract

This study aims to evaluate the robustness and security of deep vision-based crop–weed segmentation models against targeted adversarial evasion attacks and to assess the effectiveness of defense mechanisms for improving reliability in precision agriculture systems. Experiments were conducted using the SorghumWeedDataset_Segmentation dataset containing 5,555 annotated segments from 252 high-resolution RGB images of sorghum, grasses, and broad-leaf weeds. Convolutional neural networks were trained using state-of-the-art methods, and they were used to perform pixel-based segmentation and classification. Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) were used to produce adversarial examples at different magnitudes of perturbations (ε). The measures of model performance were accuracy, precision, recall, F1-score, Intersection over Union (IoU), Attack Success Rate (ASR), and a robustness measure. Adversarial training was introduced in order to increase resilience. The model had 94.8% classification and a 0.887 mean IoU under clean conditions. Nevertheless, it deteriorated considerably during adversarial perturbations. However, performance degraded significantly under adversarial perturbations. At ε = 0.03, FGSM reduced accuracy to 68.4% and mean IoU to 0.612 (ASR: 31.6%), while PGD caused a sharper decline to 54.7% accuracy and 0.482 mean IoU (ASR: 45.3%). After adversarial training, adversarial accuracy improved from 68.4% to 81.2%, and mean IoU increased from 0.612 to 0.743, with only a marginal reduction in clean accuracy (94.8% to 92.6%). The results indicate that deep agricultural vision models are very susceptible to gradient-based adversarial attacks, especially iterative ones, such as PGD. The Introduction of adversarial training makes it much more robust and trades off a little clean performance. Robustness evaluation is therefore essential for the secure deployment of AI-driven weed management systems in real-world agricultural environments.