Robustness Evaluation of Deep Learning-Based Plant Disease Classification under Adversarial and Physical Perturbations

Deep learning-based systems for plant disease identification have shown large amounts of classification accuracy when tested in a controlled setting; however, the successful operation of these technologies in the agricultural sector will require extra robustness to be able to withstand both adversarial perturbations and physical distortions. In this paper, we will examine how robust a CNN model trained to classify plants into disease categories would be under the FGSM, PGD, and against simulated physical perturbations such as changes in the illumination of the image, blur, and noise from the sensor. Additionally, we will examine where attention is focused, via grad-CAM, when the model is being attacked. Results from the experimentation indicate that there is considerable degradation of accuracy when the model is subjected to adversarial attacks (especially with PGD); however, there is only moderate degradation of accuracy when subjected to physical perturbations. These results demonstrate the importance of robust evaluation as a requirement for reliable use of AI in precision agriculture.Index Terms—Plant Disease Detection, Adversarial Attacks, FGSM, PGD, Grad-CAM, Robustness, Physical Perturbation