Image, audio, video, and many other kinds of data are nowadays mostly passed from person to person or from place to place in a digital form. It is often desirable to embed data into the digital contents for copyright control and authentication, or for secret data hiding.
Data-embedding techniques designed to take care of such tasks are commonly classi ed as watermarking or data hiding techniques in accordance with their functionalities. Watermarking techniques are often further divided into two groups: robust watermarking methods and fragile watermarking methods. In robust watermarking methods, the hidden information remains robust against manipulations from any possible sources including hostile ones. Hence such methods are usually developed to protect copyright.
On the other hand, fragile watermarking methods are usually designed to easily get broken so that common content processing operations, if there are any at all, can be found. Therefore, such methods are good for tampering detection and authentication. As for those classi ed as data hiding techniques, they are sometimes called steganographical methods, where the secret message blends in a common digital content, so that eavesdroppers will not have any idea that the secret message is there, and so they will not have the slightest 128 H. -W. Tseng, C. -C. Chang ntention of trying to break the protection. Under such circumstances, robustness seems to be less stringent, and the major issues here are the embedding capacity and invisibility.
In other words, a good data hiding method should be one that can embed as much data as possible, and the perceptual distortion of the digital content after the embedding procedure should be as little as possible. Current methods for the embedding of data into the cover image fall into two categories: spatial-based schemes (Adelson, 1990; van Schyndel et al. , 1994; Wang et al. , 2001) and transform-based schemes (Cox et al. 1997; Wolfgang et al. , 1999; Xia et al. , 1997).
Spatial-based schemes embed the data into the pixels of the cover image directly, while transform-based schemes embed the data into the cover image by modifying the coef cients in a transform domain, such as the Discrete-Cosine Transform (DCT). In this paper, we will focus upon data hiding in the DCT domain as well as quantized DCT coef cients. We shall embed the data into a JPEG (Pennebaker and Mitchell, 1993) compressed image, for most digital images are stored and transmitted in the JPEG compressed format.
Surprisingly, in the literature, data hiding techniques that deal with the JPEG compressed image (Kobayashi et al. , 1999; Noguchi et al. , 2000; Johnson and Jajodia, 1998; Chang et al. , 2002) are astonishingly few and far between. Kobayashi et al. (1999) presented a method to hide data into JPEG bitstreams. However, the embedding capacity is very limited. Jpeg–Jsteg (Johnson and Jajodia, 1998) is another famous hiding tool for embedding data into the JPEG compressed image.
The secret data is embedded into the LSB of the quantized DCT coef cients. In the scheme proposed by Chang et al. 2002), the secret data is embedded in the middle-frequency part of the quantized DCT coef cients. The scheme provides a larger embedding capacity than Jpeg–Jsteg, but the compression ratio of image is bounded. In general, for the purpose of avoiding too much distortion to the embedded image, the quantized DCT coef cients should be modi? ed as little as possible. Furthermore, the AC coef cients become zeros mostly after quantization. These zeros are usually incapable of embedding and the embedding capacity of the JPEG compressed image is thus limited.
To improve the embedding capacity, a high capacity hiding method based on the adaptive least-signi cant bit (LSB) substitution method and the human visual system (HVS) (Daly, 1994; Wandell, 1995) will be proposed. The number of bits embedded in the DCT components is computed through a prede? ned equation which is built up in accordance with the features of HVS to ensure the embedded image still preserves good image quality. Indeed, a high percentage modi? cation in DCT components will certainly lead to signi? cant distortion in the embedded image.
Therefore, how to nd the optimal balance between both ends (namely image degradation and embedding capability) is the topic of study in this paper. To keep consistency in this paper, we shall de ne some terms for later use. The data to be embedded is called the secret data, for it has usually been processed by such encryption methods as DES (DES Encryption Standard, 1977) under some security requirement. The image responsible for carrying the hidden secret data is called the cover image, and the image containing the hidden secret data is called the stego-image.
However, the compression ratio is quite restricted. It cannot be adjusted freely based on the choices of Q factor. The Proposed Method According to the descriptions and discussions in the previous section, an adaptive LSBs substitution data hiding method should be developed, and that is exactly what we have done.
In our new method, we do not embed the secret data in the high-frequency components in order not to expand the size of the stego-image. Besides, the LSB number in each DCT coef cient used for data hiding depends on the characteristics of the image High Capacity Data Hiding in JPEG-Compressed Images 33 according to HVS; as a result, the embedding capacity of the JPEG compressed image can be raised while avoiding signi? cant stego-image distortion. Meanwhile, the capacity formulas for the DC and AC components should be different due to the discrepancy between them.
To achieve our goal, each component in the JPEG default quantization table is chosen as a perceptual threshold, or “just noticeable difference”, for the visual contribution of its corresponding cosine-basis function. The table is tuned for most natural images according to the results of perceptual experiments. Since the design of the JPEG default quantization table is based on a simple HVS model, it can be easily applied to embedding capacity estimation. As we know, the image information (energy) is usually concentrated in the low frequency region after DCT transformation.
As we know, the magnitude of the DC component is positively proportional to the average pixel value in the original block, and the magnitude of the DC component is much larger than that of any AC omponent. Embedding data can be viewed as embedding a set of signals onto a larger set of background signals.
The embedded signals can only be detected by HVS when they surpass the detection threshold of HVS. To be more precise, the Weber–Fechner law states that the detection threshold of visibility for an embedded signal is proportional to the magnitude of the background signal. Thus, compared with AC components, DC components can be modi? ed by a larger quantity. However, even so, DC components still cannot be changed by a large percentage because signi? cant block artifacts are still very likely to happen in that case.