Enhanced Detection of Diffusion Model-Generated Deepfakes Using CNN Feature Maps and Forgery Traces

Recent advancements in generative artificial intelligence powered by deep learning have significantly improved image generation and manipulation, resulting in highly realistic images that pose substantial risks to multimedia security. The increasing similarity between authentic and deepfake images, especially those generated by Diffusion Models, highlights the pressing need for effective detection mechanisms. This study provides a comprehensive evaluation of counterfeit facial image detection, focusing on the generalizability and robustness of various detection methods. Using a dataset comprising real images from CelebA-HQ and synthetic images generated by five state-of-the-art models (StyleGAN2, VQGAN, DDPM, PNDM, and LDM), we benchmark four leading detection algorithms: Wang2020, Grag2021, Mandelli2022, and Ojha2023. The performance of these detectors was evaluated across different generative models and under various image perturbations, such as resizing, noise, blur, and compression. Additionally, we analyzed frequency-domain artifacts, revealing that GAN-generated images exhibit distinct frequency patterns, whereas DM-generated images closely resemble authentic ones. A novel hybrid approach combining spatial and frequency-domain features was proposed, yielding superior performance in detecting AI-generated human faces. Among the methods tested, Mandelli2022 achieved an AUC of 98.38%, while our ResNet-50+FFT model outperformed it slightly with an AUC of 98.42%. These results highlight the effectiveness of hybrid approaches in improving detection accuracy. However, detectors still face challenges in generalizing to diverse datasets, emphasizing the need for more adaptable and robust detection strategies.