海南热带海洋学院学报2025，Vol.32Issue(5)：99-105,7.DOI:10.13307/j.issn.2096-3122.2025.05.11
基于残差图神经网络的伪造人脸图像多尺度检测方法
Multi-scale Detection Method for Forged Face Images Based on Residual Graph Neural Network
摘要
Abstract
The forged facial image,with its facial key point features in multi-scale dimensions,contains a large number of shallow high-frequency components.During the extraction process,the gradients quickly disappear,which re-sults in inaccurate detection of the fake image.Therefore,a multi-scale detection method for fake face images based on a residual graph neural network was proposed.The collected fake facial images were corrected and the positions of facial key points were preliminarily estimated.The facial key points were input into the residual graph neural network,which uses residual blocks to avoid gradient vanishing.At the same time,the multi-scale attention mechanism was combined to enhance the sensitivity to forged traces at different scales,extract the multi-scale features of the facial image key points,and construct the corresponding feature maps.After fusing the feature maps,complex features of facial key points were obtained.The loss function was utilized to calculate the similarity between the extracted features and the sample image features,so as to accurately detect the counterfeit areas in the image.The results showed that the proposed detection method has an average AUC value of 0.96 in multiple datasets,and can accurately detect the forged areas in facial im-ages.关键词
残差图神经网络/伪造人脸图像/图像检测/多尺度检测/特征融合/关键点特征Key words
residual graph neural network/fake facial images/image detection/multi-scale detection/feature fu-sion/key point features分类
计算机与自动化引用本文复制引用
俞弦,邓惠俊..基于残差图神经网络的伪造人脸图像多尺度检测方法[J].海南热带海洋学院学报,2025,32(5):99-105,7.基金项目
安徽省高等学校科学研究项目(自然科学类)重大项目(2024AH040213) （自然科学类）
安徽省学科(专业)带头人培育项目(DTR2024063) （专业）
安徽省高等学校质量工程项目(2023sx174) （2023sx174）
