Defending Crop Recommendation Systems Against Poisoning and Evasion Attacks

Crop recommendation systems play a critical role in modern agriculture, providing personalized guidance for optimal crop selection and cultivation practices. However, these systems are vulnerable to adversarial attacks that can compromise their performance, integrity, and privacy. This research focuses on defending crop recommendation systems against poisoning and evasion attacks, which aim to manipulate the training data and deceive the model during prediction, respectively. The research investigates the vulnerabilities of crop recommendation systems to adversarial attacks and analyzes potential risks and consequences. To evaluate the effectiveness of the proposed defenses, extensive experiments are conducted using a real-world dataset of crop information. The results demonstrate the susceptibility of crop recommendation systems to adversarial attacks. Specifically, the accuracy of the model decreases to 0.64 and 0.65 under poisoned attacks, while the precision drops to 0.77 and 0.65. However, after implementing the defense mechanisms, the model's accuracy and precision improve to 0.73 and 0.75, respectively. Additionally, under evasion attacks, the model's recall decreases to 0.48 and the F1-score drops to 0.70. Nevertheless, the defense mechanisms significantly improve the model's recall to 0.86 and the F1-score to 0.80. The research also assesses the performance of the defense mechanisms by quantifying their ability to detect and mitigate adversarial attacks, demonstrating significant improvements achieved through their deployment.