基于深度学习的深度伪造检测方法  

Deep Forgery Detection Method Based on Deep Learning

作　　者：朱振 

机构地区：[1]东南大学自动化学院,江苏南京210096

出　　处：《工业控制计算机》2022年第6期14-16,共3页Industrial Control Computer

摘　　要：目前DeepFake这种深度伪造技术在网络上被滥用,来制作色情电影、虚假新闻,甚至可能将其用于政治人物来制造政治谣言,这对国家安全和社会稳定带来了潜在的威胁。针对这个问题,提出了一种基于深度学习的深度伪造检测方法。该模型是基于深度神经网络模型EfficientNet-B4,并结合了Vision Transformer技术。实验阶段将该方法在两个数据集FaceForensics++和Celeb-DF上进行了论证。实验结果表明,该方法较之前提出的方法不仅明显提高了识别的准确率,达到了96%,同时还具有很好的泛化性能,在跨数据集方面也取得了72%的识别准确率。该方法具有优良的性能。This paper proposes a deep forgery detection method based on deep learning.The model is based on the deep neural network model EfficientNet-B4 combined with Vision Transformer technology.In the experimental phase,the method is demonstrated on two frequent datasets,FaceForensics ++ and Celeb-DF.The experimental results show that this method improves the recognition accuracy significantly,reaching 96% compared with the previously proposed method.At the same time,it has good generalization performance, and has achieved a recognition accuracy of 72% across datasets.This method has excellent performance.

关 键 词：深度伪造 视觉变换器 高效网络 伪造检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]