Digital images that are easy to edit, modify, and exploit can be widely shared and distributed via the Internet. Thanks to powerful image processing techniques, it is increasingly easier for everyone to perfectly edit digital images and create forgeries. Creating perfect forgeries can lead to the theft and misuse of intellectual property. Proving the origin of an image and its integrity is thus essential for an image owner. As a result, image authentication and integrity verification have become important issues in recent years.
In general, the authenticity of digital images can be guaranteed by using digital signatures. Digital signatures employ asymmetric cryptography to establish the authenticity and integrity of a digital image by attaching a hash of the image. One possible drawback of digital signatures is the fact that extra bandwidth is required to transmit the signatures. Moreover, authentication based on digital signatures cannot localize changes nor reconstruct tampered regions in the image, even if integrity protection is provided.
To address the above problem, fragile watermarking schemes have been proposed as a means of verifying image integrity [1
]. The visual redundancy of digital images makes it possible to embed invisible fragile watermarks into such images without modifying the essential features of the images. To further enhance the function of fragile watermarking schemes, researchers have designed self-embedding methods, such as that of a fragile watermark concurrently consisting of an authentication code and recovery information. With the hidden authentication code and recovery information, the modifications made to the watermarked image are expected to be localized, and tampered regions can later be restored. Different from robust watermarks, for both fragile watermarking and self-embedding fragile watermarking, embedded watermarks can be easily compromised by any kind of malicious attack [1
]. In other words, any attempt to alter the image content will also alter the embedded fragile watermark itself, which is thus capable of detecting every change that has occurred to the image [4
In 2011, Lee et al. designed a hierarchical fragile watermark based on VQ index recovery [5
]. In their scheme, with the hierarchical strategy and LSB substitution, the average image quality of the watermarked image was around 39.6 dB. In the next year, He et al. [6
] presented a self-recovery fragile watermarking scheme using block-neighborhood tamper characterization. They generated nonlinear block mapping to embed the watermark and used an optimized neighborhood characterization method to detect tampering. In the next year, Zhang et al. [7
] proposed a self-embedding fragile watermarking scheme. In their scheme, they first generated DCT coefficients for each 2 × 2 block. Next, they embedded the generated fragile watermark into another block according to the block mapping. The experiments confirm that the average PSNR of the watermarked images is around 42.6 dB, and the tampered regions can be successfully localized and exactly recovered for content-only tampering.
In 2014, Lin et al. [8
] proposed a high-quality image authentication scheme based on absolute moment block truncation coding (AMBTC). They used the parity of the bitmap to generate the authentication code for authenticating each compressed image block. The proposed hierarchical inspection structure was effective in resisting a collage attack. Unfortunately, Lin et al.’s scheme did not offer a recovery feature. In the same year, Yang et al. designed a fragile watermarking with a recovery function for halftone images [9
]. In the next year, Sarreshtedari and M. A. Akhaee [10
] compressed the whole image as recovery data and encoded it as a watermark by using Reed-Solomon codes (RS codes). The tampered regions can be restored by the error-correcting RS code. However, they cannot restore more than 𝑛-𝑘 erasures when using the RS(𝑛, 𝑘) code. In 2015, Li et al. designed a reference matrix-based watermark embedding strategy to conceal authentication codes into quantization levels of the compressed images generated by the block truncation coding (BTC) [11
]. In 2016, Qin et al. [12
] designed self-embedding watermarking based on a reference-data interleaving mechanism and adaptive selection of the embedding mode. In their scheme, the binary bits in the adopted MSB layers are scrambled and individually interleaved with different extension ratios and are then combined with authentication bits to form the watermark bits for LSB embedding. In Qin et al.’s scheme, the average image quality of the recovered image remained at 45.42 dB when the tamper rate was around 12%. In 2017, Cao et al. [13
] proposed a self-embedding fragile image watermarking scheme for tamper recovery. They also adopted MSB layers and LSB layers, but they applied a hierarchical recovery mechanism in their scheme. According to the contribution of the image quality, the binary bits in the MSB layers were first scrambled and then individually interleaved with different extension ratios. Later, the interleaved data served as recovery information and were embedded into LSB layers of non-overlapping blocks along with the authentication code. Even with the hierarchical recovery mechanism, the average image quality of the restored images was not significantly improved compared to Qin et al. [12
]. In the same year, Qin et al. designed a VQ-based self-embedding fragile watermarking [14
]. In their scheme, VQ indices derived from the original image served as recovery information. Later, hash values were computed from the combination of VQ indices, and the original image was computed and embedded into the image itself via LSB substitution. Different from Qin et al. [14
], Lin et al. designed a hybrid watermark hiding strategy for compressed images generated by [15
]. Although their scheme did not provide a recovery function, the image quality of the watermarked images had been significantly improved compared with existing schemes.
In 2018, Tai and Liao [16
] embedded the fragile watermark of one block into another block according to the embedding sequence generated by a chaotic map. To reduce the smooth blocking effect of the recovered images, they used a wavelet transform rather than the average as the recovery data to enhance the image contrast. This method can effectively resist a collage attack and constant-average attack. Hong et al. [17
] proposed an efficient authentication scheme for AMBTC compressed images. They protected the AMBTC codes by embedding the authentication codes into the least significant bits (LSBs) of two quantization levels to minimize the embedding distortion. In 2019, Chen et al. [18
] proposed a novel authentication scheme for the AMBTC of a compressed image using turtle-shell-based data hiding. Previous AMBTC-based schemes have the problem of having a high quantization level, which is lower or equal to a low quantization level caused by the hiding operation. Thus, they proposed an iterative embedding mechanism to solve the above issues and achieved high tamper detection accuracy. Su et al. [19
] presented an authentication scheme based on the matrix encoding for AMBTC-compressed images. The six-bit authentication code was embedded into two sub-bitmaps using matrix encoding. Their scheme offered an improved detection rate in the first hierarchical tampering detection. In 2020, Roy et al. [20
] designed a copyright protection mechanism with digital image watermarking. In their scheme, adaptive LSB replacement was adopted to embed the watermark. Moreover, to enhance the robustness of the hidden watermark, the higher bit-planes were also modified instead of only the LSB. In 2021, Hong et al. [21
] further improved the visual qualities of marked and recovered images by using matrix encoding and side match techniques. In the same year, Chang et al. used AMBTC compression results to first generate a watermark. Later, they applied the turtle shell data hiding method to conceal the watermark in the original image [22
]. Lin et al. [23
] presented a pixel pair-wise fragile image watermarking method. They used AMBTC to generate the authentication code and recovery information as watermarks. To reduce overhead information, the bitmap generated by AMBTC was further compressed by Huffman coding.
Although many self-embedding fragile watermarking schemes have been proposed in the last six years, it continues to be a challenge to enhance the tamper detection capacity and image quality of the restored images while maintaining an acceptable visual quality in the watermarked images. In particular, an increasing number of valuable digital images are being transmitted via the Internet and shared through social media platforms. Neither individual users nor companies are willing to compromise the quality of their personal images or digital image productions when they try to adopt integrity protection mechanisms. Thus, in this paper, we present a fragile watermarking scheme as an image assurance tool for integrity protection with tamper localization and self-recovery. Tamper detection performance is related to the size of the hidden authentication code, and recovery performance is related to the size of the recovery information and the correlation between the recovery information and the original image. However, the larger the amount of the hidden authentication code and recovery information, the lower the image quality of the watermarked image is. To maintain the tradeoff between the performance of tamper detection and recovery, and the image quality of the watermarked image, AMBTC compressed codes derived from the original image are used as the authentication code that can detect every possible change that has occurred in an image with a very high probability. To allow the tampered regions of the image to be partially repaired, we utilized vector quantization (VQ) compressed codes as the recovery information. The contributions of this paper are summarized as follows:
A novel fragile watermarking combining an AMBTC compression method is designed to enhance tamper detection performance. Using the AMBTC compression codes to generate authentication codes, the average TPR, and average FTP rate are around 97% and 0.12%, respectively, which outperforms other existing schemes.
Utilizing VQ indices as the recovery information, the image quality of the restored image ranges from 31.26 dB to 46.05 dB. Even in the worst case, the PSNR is still above 30 dB.
With the increased concealment of the authentication codes, our scheme provides better tamper detection performance against a copy-paste attack and cutting attack where different tamper ratios are encountered.
The experimental results confirm that the proposed fragile watermarking scheme exceeds the performance of most existing work with respect to balancing tamper detection and the image quality of watermarked images and restored images.
reviews the basic concepts of AMBTC and VQ that are needed for the fragile watermarking scheme. Section 3
explains the proposed watermarking scheme, including the embedding, detection, and self-recovery algorithms. Experimental results and their analysis appear in Section 4
. Finally, the paper is concluded in Section 5