基于区块链存证的Deepfake内容监管技术研究

Research on Deepfake Content Monitoring Technology Based on Blockchain Depository

随着人工智能技术的发展,基于深度学习的深度伪造技术日趋成熟,通过多媒体篡改工具可以对视频中的人脸进行随意的篡改,并且几乎无法被肉眼察觉。深度伪造人脸视频成为互联网内容监管中不可或缺的重要部分。以短视频社交平台为应用场景,提出了一种基于区块链存证技术的深度伪造人脸视频内容监管方法。方法针对实际场景中图像内容质量差异的特点设计出一种线性混合的检测方法,在边缘端采用基于轻量级微调神经网络的分类器架构,结合服务器端采取频谱特征进行分类。解决了传统独立检测方法面对数亿级短视频内容的上传与发布过程中检测效率与准确率之间难以平衡的问题。上述方法针对视频文件大且大量转发导致的数据冗余问题,采用超级账本与IPFS相结合的数据存储方式,可以对深度伪造人脸视频内容进行快速精准的追溯并对用户行为进行评价。实验结果表明,所提方法在两个公共深度伪造人脸视频数据集(DeepfakeDetection、Celeb-DF)中都表现出较好的效果,并且在针对内容监管平台的性能测试中表现较好。

With the development of artificial intelligence technology,Deepfake technology based on deep learning is increasingly mature,and through multimedia tampering tools that can be randomly manipulated on human faces in videos,it is almost impossible to be detected by the naked eye.Therefore,Deepfake face videos have become an essential part of Internet content monitoring.This paper proposes a Deepfake of face video content monitoring methodology based on blockchain depository technology with short video social platforms as an applicable field.The method is designed as a linear hybrid detection method for the difference of image content quality in actual cases,adopting a classifier architecture based on lightweight fine-tuned networks at the edge,combined with server-side spectral features for classification.It solves the problem of the balance between detection efficiency and accuracy in the process of uploading and posting several billion contents of short videos by traditional standalone detection methods.In addition,the method is designed to address the data redundancy problem caused by large video files and massive forwarding,using a combination of Hyperledger and IPFS data storage,which can quickly and accurately trace the content of Deep-fake face videos and evaluate user behavior.The experimental results show that the method performs well in both public Deepfake face video datasets(DeepfakeDetection,Celeb-DF),and it performs well in performance measurements for Deepfake content monitoring platforms.