Abstract: Most existing deepfake face detection methods rely on specific forgery patterns, such as noise artifacts or local textures, making them highly dependent on known forged features and lacking generalization to unknown forgeries. To address this issue, this paper proposes a multiloss fusion detection framework based on the reconstructionclassification learning(RECCE) network. A prototype similarity mechanism is introduced to measure the distance between sample features and class prototypes, enhancing the model’s ability to detect unknown forgeries. A joint loss function combining prototype loss, binary crossentropy loss, reconstruction loss, and metric learning loss is designed to strengthen feature learning from multiple perspectives. Moreover, multilevel encoder features are fused, and a reconstructionguided attention mechanism focuses the model on forged regions rather than the entire face, improving robustness and accuracy. Experiments conducted on several benchmark datasets and compared with six stateoftheart methods demonstrate significant improvements: training time is reduced to 28% of the original model, and the AUC increases by 1.48% in crossdomain evaluation when trained on FaceForensics++ (c40) and tested on CelebDF. The results verify the superior performance and generalization ability of the proposed method.

Key words: deepfake face detection; prototype similarity; feature fusion; attention mechanism; generalization ability

摘要： 现有深度伪造人脸检测方法多依赖训练集中已有的伪造模式，如噪声或局部纹理，因此在应对复杂未知伪造样本时泛化能力不足.针对该问题，基于RECCE(reconstructionclassification learning)网络提出一种多损失融合的伪造人脸检测框架.该方法引入原型相似度机制，通过度量输入特征与类别原型的距离，提高对未知伪造的识别能力；并构建由原型损失、二元交叉熵损失、重建损失和度量学习损失组成的联合损失，以多角度增强伪造特征表达.对编码器多层级特征进行融合，引入重建差异引导的注意力机制，使模型聚焦伪造区域，提高检测准确率(Acc)与鲁棒性.实验在多个基准数据集上与6种方法比较表明，该方法训练效率提升显著，仅为RECCE的28%.在FaceForensics++(c40)上训练并在CelebDF上跨域测试时，曲线下面积(AUC)提升1.48%.较现有方法具有更优性能和泛化能力.

关键词: 人脸伪造检测；原型相似度；特征融合；注意力机制；泛化能力