面向AI安全的深度伪造视频检测技术

Video deepfake detection technology for AI security

针对实际应用伪造类型多样化,而检测过程中只检测单一伪造数据类型等问题,提出了一种新的数据预处理方法,解决了视频中运动图像容易连续重叠的问题;针对人脸模糊伪造,侧面人脸伪造和遮挡人脸伪造问题,利用小卷积核代替大卷积核的方法训练出自己的Deeper网络;针对帧插入、帧复制、帧修改、帧内篡改四种常见的深度视频伪造,尤其是在测试领域差距较大的情况下,容易产生泛化性不足的问题,设计了一种改进的基于双流分析的深度伪造视频检测模型。

Introduces a novel data preprocessing technique aimed at addressing the challenge of persistent image overlap in video sequences.To combat issues such as face blur forgery,side face forgery,and occluding face forgery,we employ a Deeper network trained using small convolution kernels,in lieu of large ones.Furthermore,we propose an improved deepfake detection model that leverages a two-stream analysis approach to enhance generalization.This model effectively tackles common video forgery types,including frame insertion,frame copying,frame modification,and intra-frame tampering.Particularly,our approach exhibits robust performance even in scenarios where there is a substantial divergence in the test data.