Face Forgery Detection Method Based on Multi-source Features Fusion of GradGR
HAN Dongyu
GUO Ting
YANG Gaoming
ZHU Peng
SONG Yifan
Abstract：To address the issue that the existing face forgery detection methods performed well on specific forgery operations but had obvious deficiencies in cross-dataset generalization ability,a face forgery detection model based on image gradient-guided reconstruction(GradGR)was proposed.This model constructed an auxiliary gradient reconstruction branch on top of the original image reconstruction backbone.Through a feature transfer mechanism,the intermediate features of this branch were used to guide the backbone to focus on the forged region revealed by the gradient.In addition,a multi-source feature fusion(MSFF)module was included in the GradGR model,which effectively enhanced cross-layer feature interactions and markedly improved the efficiency of feature fusion by integrating codec intermediate features with two-way reconstruction differences.The results demonstrated that the GradGR model not only achieved the optimal level within the datasets but also improved the average cross-dataset accuracy by nearly 3.22%compared with the extreme inception(Xception)model.This model provided a new research idea for face forgery detection.
Keywords：facial forgery detectiondeep learningcomputer visionimage reconstruction learningimage gradientmulti-source feature fusion
Publication Date：2025-06-20
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：8( 217-223,258 )