SFSimNet: An Efficient Spatial-Frequency Multi-Scale Intra-Feature Similarity Measurement Network for Deepfake Detection

The rapid evolution of face synthesis technology, powered by AIGC (AI-generated content), has facilitated its widespread misuse, posing substantial security risks in consumer electronics applications. The core principle of deepfake technology involves manipulating or forgery of facial data using advanced AI/ML algorithms. However, such tampered or forged content inherently exhibits discrepancies from the original, creating opportunities for detection. In this paper, we exploit these discrepancies by employing intra-feature similarity measurements to distinguish between content for deepfake detection. We introduce a computationally efficient and low-complexity deepfake detection framework based on a multi-scale similarity measurement mechanism that operates in both the spatial and frequency domains. This approach effectively captures subtle traces of manipulation across diverse feature scales and domains. Comprehensive experimental results demonstrate our method’s exceptional performance, surpassing state-of-the-art deepfake detection techniques while requiring fewer parameters and lower computational costs (FLOPs), making it suitable for real-time consumer electronics applications. The code is available at.