Structural Consistency for Face Forgery Detection via Frequency Domain Enhancement and Self-Predictive Learning

Existing face forgery detection methods often overfit to known forgery patterns, resulting in limited generalization to unseen manipulations. To address this issue, we propose a Structural Consistency method for Face Forgery Detection via Frequency Domain Enhancement and Self-predictive Learning, which embeds frequency-domain features into the spatial representation and leverages the structural consistency inherent in genuine facial contexts to provide more discriminative cues for forgery identification. Specifically, we design a data augmentation module to extract the frequency information through the Discrete Cosine Transform (DCT) and enhance it using a Frequency Domain Enhancement Module (FEM) to capture subtle forgery artifacts. Furthermore, we design a Self-Prediction Learning Module (SPLM) that reconstructs the occluded central region of a face by exploiting the structural consistency of real facial features. To further guide the learning process, we define a self-predictive reconstruction loss that minimizes the prediction error in the occluded region and helps reinforce structural consistency. Moreover, we propose a Reconstruction Difference Guidance (RDG) module, which explicitly emphasizes potential forgery regions by computing pixel-wise discrepancies between the reconstructed image and the original input. This process produces an attention map that guides the classifier to focus on semantically inconsistent or anomalous regions. Experimental results demonstrate that our method achieves superior generalization and robustness across diverse datasets.