Diffusion-Aware Deepfake Detection via Kernel PCA on Noise Residuals for Semantic Communication

Deepfake creation relies on advanced deep learning models enabling hyper-realistic synthetic media with applications in industrial and social semantic communication. However, these advancements pose severe threats to digital security and semantic communication, where preserving content integrity is critical. Traditional detection methods fail against diffusion-generated content due to diminishing distributional gaps between real and synthetic samples. We propose a novel Deepfake Detection Pipeline that leverages the inherent noise-prediction mechanism of Denoising Diffusion Probabilistic Models (DDPMs) to extract discriminative distributional features. By fine-tuning a DDPM noise predictor, we capture subtle artifacts in synthetic images that are imperceptible to conventional detectors. These features are then analyzed through a kernel-based normative compliance detector, combining Kernel PCA (KPCA) with out-of-distribution detection to isolate deepfakes by measuring deviations from natural image manifolds. The proposed approach addresses the key challenge of detecting fine-grained distributional differences in high-quality synthetic media. Experiments demonstrate state-of-the-art performance on diffusion-generated deepfakes, outperforming existing methods on CREMA-D and Remote Sensing DeepFake Detection Dataset. The pipeline allows integration with existing detection frameworks, offering a scalable solution for evolving threats.