基于生成对抗网络的解压位图反取证方法
Anti-forensics of Decompressed Bitmap Based on Generative Adversarial Network
冯翔 1毕成龙1
作者信息
- 1. 安徽理工大学 计算机科学与工程学院,安徽 淮南 232001
- 折叠
摘要
针对图像解压缩过程中引入色度上采样痕迹,导致未压缩和解压缩图像色度平面上奇偶和偶奇像素对差异分布的不同,提出了一种基于生成对抗网络的解压位图反取证方法,将JPEG解压缩反取证工作建模为图像到图像的转换.在此基础上,针对解压缩过程中所引入的色度上采样痕迹设计了损失函数,经过迭代训练后,模型可以生成具有极高视觉质量和合理统计特征的重建图像.实验结果表明:提出的反取证方法生成的修改图像能够欺骗现有的检测器,并且具有出色的视觉质量;在客观评价指标上,在降低检测器准确率方面取得了显著成效;相比于其他基于生成对抗网络的反取证方法,在压缩质量为50 的情况下,生成图像的峰值信噪比和结构相似度均呈现了一定的提升.
Abstract
In view of the traces of chroma upsampling introduced during the image decompression process,which lead to different distributions of differences between odd-even and even-odd pixel pairs on the chroma plane of uncompressed and decompressed images,a decompressed bitmap anti-forensics method based on generative ad-versarial networks is proposed.The approach models JPEG decompression anti-forensic work as image-to-image conversion.On this basis,a loss function is designed for the chroma upsampling traces introduced during the de-compression process.After iterative training,the model can generate reconstructed images with extremely high vis-ual quality and reasonable statistical characteristics.Experimental results show that the modified images generated by this anti-forensics method are able to deceive existing detectors and have excellent visual quality.In terms of objective evaluation indicators,this anti-forensics method has achieved remarkable results in reducing the accura-cy of the detector.Compared with other anti-forensics methods based on generative adversarial networks,when the compression quality is 50,the peak signal-to-noise ratio and structural similarity of the generated images have shown a certain improvement.
关键词
反取证/生成对抗网络/色度上采样/JPEG解压缩Key words
anti-forensics/generative adversarial network/chroma upsampling/JPEG decompression引用本文复制引用
基金项目
安徽省质量工程项目(2020mooc188)
出版年
2024
NETL
NSTL
万方数据