Security-Aware Self-Supervised Adaptive DCNN for Rice Leaf Disease Detection

Abstract

Agriculture plays a critical role in global food security, yet crop production remains highly vulnerable to disease outbreaks that significantly reduce yield and crop quality. Deep learning-based models shows high performance under controlled environment; however, their reliability often deteriorates in real world environments due to variations in illumination, background complexity, and environmental noise, making the model vulnerable to the prediction error and adversarial perturbation. To address these challenges, this study proposes a novel Security-Aware Self-Supervised Adaptive Deep Convolutional Neural Network (SA-SSA-DCNN) for robust rice leaf disease detection. The proposed model uses the self-supervised learning to reduce the model’s reliance over the labelled data, attention mechanism to localize the features, and adversarial training to resilient against the environmental perturbation and noise. To evaluate the model’s performance the benchmarked rice leaf dataset was obtained from Kaggle. The proposed SA-SSA-DCNN achieved 99.25% accuracy and F1 score as 0.993, Consistently outperforms the standard SSL model and randomly initialized baseline models across all evaluation metrics. Furthermore, the model demonstrates strong robustness under adversarial and noisy conditions, significantly reducing attack success rate while maintaining high prediction reliability for real-world agricultural deployment.