基于噪声感知残差网络的JPEG隐写分析方法

为了进一步挖掘自适应JPEG隐写图像中隐写噪声信号特征,提出基于噪声感知残差网络的JPEG隐写分析方法.该方法由噪声感知、噪声分析和判断三部分组成.其中,噪声感知部分提取图像噪声,利用图像去噪网络,更加全面地捕获隐写引入的扰动;噪声分析部分获得噪声信息的统计特征;判断部分确定图像是否携带隐写信息.此外,网络中的残差连接有效融合多尺度特征,并防止训练中出现梯度消失和爆炸.多种条件下的对比实验结果表明,该方法相较于对比算法,能够提升针对JPEG自适应隐写的检测性能并具有更好的泛化能力.

JPEG steganalysis based on image noise-aware residual network

To further explore the characteristics of the steganographic noise signal in the adaptive JPEG steganographic images, this paper proposed a noise-aware residual network for JPEG image steganalysis. The proposed network included three components, noise-aware component, noise analysis component and judgement component. First, noise-aware component extracted the noise of an image, it used denoising network to offer more comprehensive information related to steganography. Then, noise analysis component got discriminative features from image noise related information. Last, judgement network gave the final detection result. Moreover, residual connections integrated high-level features and low-level features and prevented gradient vanishing and exploding. Experiments under various conditions show that the network has an advanced performance, and has better generalization ability.