Frequency-Aware Feature Learning with Single-Center Loss for Face Forgery Detection

With the rapid development of image generation techniques based on autoencoders and GANs, deepfake face forgeries have become both popular entertainment and serious security threats, making face forgery detection a hot research topic in computer vision.

Most existing detectors treat the problem as a binary classification task supervised by softmax loss, which lacks explicit constraints on intra‑class compactness and inter‑class separability, leading to insufficient discriminative features, especially because forged faces exhibit diverse intra‑class distributions.

To address this, the authors introduce a Single‑Center Loss (SCL) that aggregates natural faces around an adaptively updated center while pushing forged faces away, thereby improving feature discrimination without increasing optimization difficulty.

They also design an Adaptive Frequency Features Generation Module (AFFGM) that converts images to the frequency domain via JPEG‑style preprocessing and a learnable Adaptive Frequency Information Mining Block (AFIMB), enabling data‑driven extraction of discriminative frequency features.

The overall network combines spatial features from an RGB branch with frequency features from AFFGM, fuses them, and trains end‑to‑end under both softmax loss and SCL supervision.

Extensive experiments on the FF++ dataset compare SCL with center loss, triplet loss, and other baselines, showing superior performance; ablation studies evaluate the impact of different fusion modules, and the proposed method outperforms previous state‑of‑the‑art approaches.

In summary, the paper presents a novel loss function and a fully data‑driven frequency feature extractor that together enhance deepfake detection, especially under low‑quality conditions, and suggests future work on model generalization and applying SCL to other domains.