基于多域时序特征挖掘的伪造人脸检测方法

Forgery face detection method based on multi-domain temporal features mining

随着计算机技术在金融服务行业中的不断发展,金融科技便利了人们的日常生活,与此同时,数字金融存在着危害性极大的安全问题。人脸生物信息作为人物身份信息的重要组成部分,广泛应用于金融行业中的支付系统、账号注册等方面;伪造人脸技术的出现不断冲击着数字金融安全体系,给国家资产安全和社会稳定造成了一定的威胁。为了应对伪造人脸带来的安全问题,提出了一种基于多域时序特征挖掘的伪造人脸检测方法。所提方法从视频在空域和频域中存在的时序特征出发,基于人脸统计特征数据分布的一致性以及时间上动作趋势的一致性,对篡改特征进行区分增强。在空域中,所提方法使用改进的长短记忆网络(LSTM)来挖掘帧间的时序特征;在频域中,利用3D卷积层来挖掘不同频段频谱的时序信息,并与主干网络提取到的篡改特征进行融合,进而有效地区分伪造人脸和真实人脸。所提方法在主流数据集中表现优越,证明了所提方法的有效性。

Financial technology has greatly facilitated people's daily life with the continuous development of computer technology in the financial services industry.However,digital finance is accompanied by security problems that can be extremely harmful.Face biometrics,as an important part of identity information,is widely used in payment systems,account registration,and many other aspects of the financial industry.The emergence of face forgery technology constantly impacts the digital financial security system,posing a threat to national asset security and social stability.To address the security problems caused by fake faces,a forgery face detection method based on multi-domain temporal features mining was proposed.The tampering features were distinguished and enhanced based on the consistency of statistical feature data distribution and temporal action trend in the temporal features of videos existing in the spatial domain and frequency domain.Temporal information was mined in the spatial domain using an improved LSTM,while in the frequency domain,temporal information existing in different frequency bands of the spectrum was mined using 3D convolution layers.The information was then fused with the tampering features extracted from the backbone network,thus effectively distinguishing forged faces from real ones.The effectiveness of the proposed method was demonstrated on mainstream datasets.