Abstract
The proliferation of AI-generated medical deepfakes, such as tumor insertions or removals in diagnostic scans, threatens patient safety and healthcare integrity. Existing detection methods often lack robustness against adversarial attacks or fail to integrate multimodal feature representations. To address these gaps, we propose AFFETDS (Adversarial Feature Fusion Enhanced Tumor Detection System), a novel ensemble framework combining adversarial training, feature fusion, and weighted voting. AFFETDS leverages adversarial attack methods (PGD, FGSM) to harden the model, fuses high-level ResNet50 features with handcrafted HOG descriptors, and employs an SVM-based ensemble classifier. Evaluated on a curated dataset of 1378 MRI scans (774 real, 604 manipulated) from TCIA and ADNI repositories, AFFETDS achieves state-of-the-art performance with 91.5% accuracy, 90.7% precision, and 91.2% recall, outperforming baseline models (SVM: 86.2%, CNN: 88.4%). The framework's ROC-AUC (0.80) and calibrated confidence scores demonstrate superior generalization across diverse imaging conditions. The ability of combining the adversarial techniques with multimodal feature fusion, our proposed AFFETDS framework improves the detection of subtle tumor manipulations, presenting an important safeguard to maintain the authenticity of medical images. The findings of research work underscore the urgent need of proactive defenses against growing deepfake threats in healthcare applications.