Enhancing tumor deepfake detection in MRI scans using adversarial feature fusion ensembles

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cybersecurity

Background:

• AI-generated medical deepfakes, like altered tumor scans, threaten patient safety and healthcare integrity.
• Current deepfake detection methods often lack robustness against adversarial attacks and fail to integrate multimodal features.

Purpose of the Study:

• To propose AFFETDS (Adversarial Feature Fusion Enhanced Tumor Detection System), an ensemble framework to enhance the detection of AI-generated medical deepfakes.
• To improve the robustness and accuracy of medical image authenticity verification.

Main Methods:

• Developed AFFETDS, an ensemble framework combining adversarial training (PGD, FGSM) and multimodal feature fusion (ResNet50 + HOG).
• Employed an SVM-based ensemble classifier for weighted voting.
• Evaluated on 1378 MRI scans from TCIA and ADNI repositories.

Main Results:

• AFFETDS achieved state-of-the-art performance: 91.5% accuracy, 90.7% precision, 91.2% recall.
• Outperformed baseline models (SVM: 86.2%, CNN: 88.4%).
• Demonstrated superior generalization with ROC-AUC of 0.80 and calibrated confidence scores.

Conclusions:

• AFFETDS effectively detects subtle tumor manipulations in medical images by combining adversarial techniques and multimodal feature fusion.
• The framework serves as a crucial safeguard for maintaining the authenticity of medical images.
• Highlights the urgent need for proactive defenses against deepfake threats in healthcare.