针对低质量视频的双支流人脸伪造检测方法  

Double Branch Face Forgery Detection Method for Low Quality Video

作　　者：宋清华 吕东辉[1] 冯国瑞[1] 

机构地区：[1]上海大学通信与信息工程学院,上海200072

出　　处：《工业控制计算机》2024年第1期109-110,114,共3页Industrial Control Computer

摘　　要：随着人脸伪造技术不断的发展,如今经过伪造后的视频和图片的人脸伪造质量大幅度提升,这对人身安全、财产安全乃至公共安全存在一定的危害性。因此,迫切需要一种有效的检测方法来区分真假人脸。然而,现有的检测方法面对低质量的虚假人脸视频时存在一定的局限性,即面对压缩过后的低质量视频检测性能较差,此外,泛化性能较差,检测准确率有所下降。为了提升检测网络的准确性和泛化性,将语义信息和噪声信息相结合,提出一个双支流网络,在关注图像语义信息的同时通过高频噪声信息展示出伪造区域和真实区域的不一致性。利用高频噪声信息暴露出的不一致性,重点关注图像语义信息中的伪造痕迹。交互模块增进语义信息和高频信息之间的交互性与融合性。在FaceForensics++数据集进行了训练和测试,并在Celeb-DF数据集上评估该模型的跨数据集泛化性能。从实验结果中可以证明该模型的有效性和可靠性。Existing detection methods have certain limitations when dealing with low-quality fake facial videos,such as poor detection performance in compressed low-quality videos,poor generalization performance,and a decrease in detection accuracy.In order to improve the accuracy and generalization of the detection network,this paper proposes to combine semantic information and noise information,and proposes a dual tributary network.While focusing on the Semantic information of the image,this paper can show the inconsistency between the forged area and the real area through high-frequency noise information.The inconsistency exposed by high-frequency noise information focuses on the forgery traces in the image semantic information.It enhance the interactivity and integration between semantic information and high-frequency information through interaction modules.To evaluate,this paper traines and testes the model on the FaceForensics++dataset,and evaluated its cross dataset generalization performance on the Celeb DF dataset.

关 键 词：人脸伪造 双支流 高频噪声 图像语义 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]