Detection of Deepfake Medical Images Based on Spatial and Frequency Domain Analysis

The advent of deepfake technologies poses a significant threat to reliance on medical imaging. This can lead to misdiagnosis and threaten patient safety. To address this issue, this study introduces a novel deep-learning approach that integrates spatial and frequency–domain analyses to detect manipulated medical images. The proposed approach combines the pre-trained EfficientNet-B0 model for spatial feature extraction with the Discrete Wavelet Transform (DWT) for frequency-domain analysis. The model was validated using the CT-GAN dataset comprising real and GAN-generated fake CT scan images. Individually, the EfficientNet-B0 model achieved an accuracy of 97.03%, and the DWT-based analysis reached 97.84%. Combining both approaches, the hybrid model achieved an overall accuracy of 99.6%. Data augmentation and weighted cross-entropy loss were employed to enhance the training process and address class imbalance. Performance was assessed using standard metrics, including accuracy, precision, recall, F1-score, Receiver Operating Characteristic (ROC), and precision-recall curves. These results demonstrate the strengths of our model in detecting minor variations in an image and validating the power of combining frequency analysis with spatial evaluation. This method holds great promise for protecting the integrity of medical imaging against deepfake threats.