Enhancing Robustness of Fruit and Vegetable Image Classification Against Adversarial Attacks Through Defensive Strategies

Typically, machine learning models allow number of applications including recognition and classification. The adversarial machine learning attack fools the machine learning and deep learning models by initiating a noise attack over the models. The adversarial attack targets the machine learning models to harm the models in a way that the models start classifying the images incorrectly. This leads to increasing the error rate and decreasing the accuracy rate of the machine learning models. The attack is not visible to the human eye to detect. Only the machine learning models can detect the changes. Before and after the attack, all images look identical for the human eye. In this study, we introduce a deep learning model that classifies fruit and vegetable images. Then, we conduct an experiment on the deep learning model. The experiment has two folds, in the first fold we apply three different adversarial attack methods on the model which are the Fast Gradient Sign Method (FGSM), Iterative Fast Gradient Sign Method (I-FGSM) and Momentum Iterative Fast Gradient Sign Method (MI-FGSM). We measure the model's performance in term of accuracy scores before and after applying each attack. The model's performance degrades, and the accuracy scores drops down because of the adversarial attack methods. The scientific experiment we conduct shows the decrease in the accuracy scores over different values of epsilon. The second fold of the study is implementing three different defense strategies into the model which are the distilled neural network, adversarial training, and binary thresholding. We measure the model's performance after applying the defense strategies individually. The model becomes more robust against the adversarial attack, and the accuracy scores of the model increases because of the defense lines incorporated into the model. The experimental results show that only the distilled neural network and the adversarial training strategies successfully protected the model against the adversarial attack.