Abstract: To address the difficulty of balancing detection accuracy and model complexity in lightweight face forgery detection networks, a lightweight spatial-frequency fusion network, termed SFGFNet (Spatial-Frequency Guided Fusion Network) was proposed. First, a spatial branch composed of the Local Artifact Enhancement Module (LAEM) and Texture Detail Feature Extraction Module (TDFE) was designed to enhance subtle boundary, texture, and local structural artifacts. Then, multi-level wavelet decomposition and frequency gating were used to extract and adaptively enhance multi-scale and multi-directional frequency responses. The Global Frequency Relation Context (GCTX) was introduced with phase information to model dependencies among frequency bands and capture global frequency anomalies. Finally, the Similarity-map Cross-Guided Fusion Module (SCGF) was built to reorganize spatial and frequency features through bidirectional similarity guidance and achieve complementary fusion. Experimental results show that SFGFNet has only 5.14 M parameters and requires 3.80 G FLOPs. Its AUC values on FF++ (HQ and LQ), DFDC, and WildDeepfake are 99.71%, 97.08%, 99.40%, and 94.56%, respectively. When trained on FF++ (C23), its cross-dataset AUC values on Celeb-DF and DFDC are 73.13% and 70.17%, respectively, representing improvements of 5.51 and 3.59 percentage points over LDSFNet (lightweight dynamic selection fusion network). These results show that SFGFNet improves detection in complex real-world scenarios and against unseen forgery types at low computational cost, providing a lightweight solution for resource-constrained face forgery detection.

Key words: deepfake detection, face forgery, lightweight network, spatial-frequency feature fusion, wavelet transform

摘要： 摘 要: 针对现有轻量级人脸伪造检测网络难以兼顾检测精度与模型复杂度的问题，提出一种面向人脸伪造检测的轻量级空频融合网络SFGFNet（Spatial-Frequency Guided Fusion Network）。首先，设计由局部伪影增强模块（LAEM）和纹理细节特征提取模块（TDFE）组成的空间纹理分支，以强化边缘过渡、纹理不连续和局部结构失衡等细微空间伪影；其次，利用多级小波分解和频率门控机制提取并自适应增强不同尺度和方向的频带响应，通过引入全局频率关系上下文（GCTX），结合相位参考信息建模不同尺度频带之间的依赖关系，以获得具有整体结构约束的全局频率异常表示；最后，构建交叉相似图引导融合模块（SCGF），通过空间相似图和频率相似图双向引导两分支特征重组，实现空间伪影与频率异常的互补融合。实验结果表明，SFGFNet仅含5.14 M参数量和3.80 G浮点运算量，在FF++（HQ、LQ）、DFDC和WildDeepfake数据集上的AUC分别达到99.71%、97.08%、99.40%和94.56%；以FF++（C23）为训练集时，在Celeb-DF和DFDC上的跨数据集AUC分别达到73.13%和70.17%，较LDSFNet(lightweight dynamic selection fusion network)分别提高5.51和3.59个百分点。结果表明，SFGFNet能够以较低的计算开销提高复杂真实场景及未知伪造类型下的检测能力，为资源受限场景中的人脸伪造检测提供轻量化解决方案。

关键词: 深度伪造检测, 人脸伪造, 轻量级网络, 空频特征融合, 小波变换