An Investigation into Steganography Algorithms

Abstract

This project aims to investigate the performance of the LSB algorithm with different replacement strategies (1-bit, 2-bit, 3-bit) and the performance of the BPCS algorithm at the maximum hiding capacity compared with the 1-bit, 2-bit and 3-bit LSB algorithms in terms of hiding capacity, imperceptibility and security. Moreover, it assesses the effectiveness of the improved BPCS method proposed by Shi and Li (2010) on color images in terms of hiding capacity, imperceptibility and security. We implemented these algorithms by using Java language and tested them on two different color images with a size of 512x512; These two images were not popular on the internet, which may make the investigation of performances of these algorithms deeper. The performance of each method was then evaluated. This study found that the hiding capacity of the 2-bit and 3-bit LSB algorithms is higher than that of the 1-bit LSB method, but they were ineffective, in terms of imperceptibility performance, with cover images that have a lot of flat colors, such as Sky. Further, the hiding capacity of the BPCS algorithm is dependent on the complexity of a cover image, therefore, the original BPCS method outperformed the 1-bit, 2-bit and 3-bit LSB algorithms in terms of embedding capacity that exceeded 50% of the size of cover image. However, in terms of visual imperceptibility performance, the BPCS algorithm at the maximum capacity was ineffective. Additionally, in terms of the impact of changing the value of threshold in the BPCS method, the embedding capacity increased when the threshold value decreased, while the visual quality of stego image was poor. Regarding security, neither the 1-bit, 2-bit, and 3-bit LSB algorithms nor the original BPCS algorithm can effectively resist histogram-based steganalysis attacks. Further, the improved BPCS method proposed by Shi and Li (2010) provides acceptable visual quality. However, the hiding capacity was lower than in the original method. Moreover, it is still vulnerable to histogram-based attacks. For the PSNR benchmark, we found that the PSNR measure is not effective for measuring the imperceptibility performance for information embedding algorithms that directly depend on the modification of the intensity values in images.