A multi-branch convolutional neural network for detecting double JPEG compression

📝 Abstract

Detection of double JPEG compression is important to forensics analysis. A few methods were proposed based on convolutional neural networks (CNNs). These methods only accept inputs from pre-processed data, such as histogram features and/or decompressed images. In this paper, we present a CNN solution by using raw DCT (discrete cosine transformation) coefficients from JPEG images as input. Considering the DCT sub-band nature in JPEG, a multiple-branch CNN structure has been designed to reveal whether a JPEG format image has been doubly compressed. Comparing with previous methods, the proposed method provides end-to-end detection capability. Extensive experiments have been carried out to demonstrate the effectiveness of the proposed network.

💡 Analysis

Detection of double JPEG compression is important to forensics analysis. A few methods were proposed based on convolutional neural networks (CNNs). These methods only accept inputs from pre-processed data, such as histogram features and/or decompressed images. In this paper, we present a CNN solution by using raw DCT (discrete cosine transformation) coefficients from JPEG images as input. Considering the DCT sub-band nature in JPEG, a multiple-branch CNN structure has been designed to reveal whether a JPEG format image has been doubly compressed. Comparing with previous methods, the proposed method provides end-to-end detection capability. Extensive experiments have been carried out to demonstrate the effectiveness of the proposed network.

📄 Content

1   A multi-branch convolutional neural network for detecting double JPEG compression  

Bin Li, Hu Luo, Haoxin Zhang, Shunquan Tan, Zhongzhou Ji Shenzhen Key Lab of Media Information Security, Shenzhen University, P. R. China         ABSTRACT Detecting double JPEG compression is important to forensics analysis. A few methods were proposed based on convolutional neural networks (CNNs). These methods only accept inputs from pre-processed data, such as histogram features and/or decompressed images. In this paper, we present a CNN solution by using raw DCT (discrete cosine transformation) coefficients from JPEG images as input. Considering the DCT sub-band nature in JPEG, a multiple-branch CNN structure has been designed to reveal whether a JPEG format image has been doubly compressed. Comparing with previous methods, the proposed method provides end-to-end detection capability. Extensive experiments have been carried out to demonstrate the effectiveness of the proposed network.

Keywords: convolutional neural network; discrete cosine transformation; double JPEG compression; image forensics; multi-branch structure

INTRODUCTION With the tremendous evolution of Internet and online social media, multimedia data (especially images) have been pervasively used due to their abilities to show contents intuitively, exchange information conveniently, and entertain the public ubiquitously. Meanwhile, common people can edit images or videos with handy software; while human eyes can hardly notice the changes. Consequently, increasing concerns are drawn on the security issues regarding the originality and the authenticity of multimedia data, which is the main task of digital forensics.
JPEG (Wallace, 1992) images are the most common ones generated from cameras and widely spread on the Internet. Therefore, there is a high probability that a forger would choose JPEG images to edit, such as performing copy-move or splicing operations, and afterwards performing JPEG recompression for distribution. Detection of double JPEG compression may provide useful information to reveal the traces of image edition or fabricated manipulation.
Various kinds of approaches have been proposed to perform double JPEG detection based on extracting handcrafted features (Lukas et al., 2003; Popescu & Farid, 2004; Fu et al., 2007; Li et al., 2008; Amerini et al. 2014; Pasquini et al., 2014; Taimori et al., 2016). It is generally assumed that the AC (alternating current) DCT (discrete cosine transform) coefficients of a JPEG image follow Laplacian distribution, and their first significant digits conform Benford’s law. When the image has been JPEG compressed twice with a different quantization matrix, the 2   distribution of DCT coefficients is deviated from the statistical model. As a result, handcrafted features are often extracted based on the statistics of DCT coefficients. (Lukas et al., 2003) discovered that the histogram of DCT coefficients in a doubly compressed JPEG image fluctuate periodically. The phenomenon, called double quantization (DQ) effect, has been mathematically verified by (Popescu &Farid, 2004). Based on the observation in (Fu et al., 2007) that the DCT coefficients of singly JPEG compressed images follows the Benford’s law, while that of doubly compressed images violating it, (Li et al., 2008) designed mode-based first digit features (MBFDF) to effectively detect double JPEG compression. They extracted the first significant digits (0~9) of DCT coefficients in the first 20 AC sub-bands, and concatenated them as a 180-D feature vector. (Amerini et al. 2014) took the first digits of 2, 5, and 7 in the first 9 AC sub-bands to reduce the feature dimension to 27-D. (Pasquini et al., 2014) proposed a statistical model based on Benford-Fourier coefficients. (Taimori et al., 2016) developed a quantization table unaware method for double JPEG compression detection based on Benford’s law by using all non-zero and zero AC coefficients. It is tempting that data-driven approaches, such as those based on deep learned features (LeCun et al., 2015; Simonyan & Zisserman, 2014; Szegedy et al., 2015; He et al., 2016), may explore hierarchy data representation, and automatically expose statistical deviation without knowing much of the details of image manipulation. (Wang & Zhang, 2016; Amerini et al., 2017) have proposed effective CNNs for double JPEG compression. However, as shown in Fig. 1 with the overview of network architecture and detailedly described in the next section, the input of their CNNs is preprocessed by extracting the histogram of DCT coefficients, instead of directly using the raw DCT coefficients. As we know, histogram is a kind of first-order statistics. Such a preprocessing stage has already provided some handcrafted information for double JPEG compression, as implicitly shown in (Lukas et al., 2003). As a result, these CNNs may lack of end-to-end feature extract

This content is AI-processed based on ArXiv data.