Enhancing the robustness of CNN-based lung cancer detection models against label-flipping poison attacks using defensive distillation

Deep learning has significantly advanced automated medical imaging, particularly in lung cancer detection, yet vulnerability to adversarial manipulation continues to limit clinical reliability. This study investigates the impact of a 30% random-uniform label-flipping poisoning attack on Convolutional Neural Network (CNN) models trained using the IQ-OTH/NCCD dataset comprising 1190 CT images. A multi-strategy defense pipeline is proposed, integrating defensive distillation, Isolation Forest data sanitization, and a noise-tolerant loss function to enhance robustness against training data corruption. To ensure a valid evaluation framework, Synthetic Minority Oversampling Technique (SMOTE) was applied only to the clean training subset before any poisoning was introduced, preventing the propagation of corrupted labels and establishing a balanced and uncontaminated foundation for teacher training. A high-accuracy teacher model trained on this clean SMOTE-balanced dataset produces temperature-scaled soft labels that guide the Student model. The Student is then trained on a sanitized dataset filtered to remove anomalous or inconsistent samples, using Symmetric Cross-Entropy loss to reduce sensitivity to mislabeled data. Experimental results show that the pipeline maintains strong performance, achieving 99% accuracy on clean data and 95 to 96% accuracy under poisoning, while preserving stable precision, recall, and confidence calibration across all classes. These findings demonstrate that the proposed strategy effectively mitigates label-flipping poisoning, offering a reproducible path toward secure and trustworthy AI systems for medical imaging applications.