Abstract:
Steganographic algorithms are mainly evaluated by their security. Traditional
steganographic frameworks use different embedding algorithms to achieve this goal.
This means, embedding the secret message directly to a cover image. Nonetheless, with
the development of sophisticated machine learning based steganalysis algorithms even
the slightest modifications can be detected. This has triggered a lot of researchers to
pursue cover-less steganography. Image synthesis with generative adversarial networks
is one of the proposed solutions. However it has suffered drawbacks in convergence
speed, training stability and realistic image generation.
This research is about improving the generative steganography framework proposed
in (Hu et al., 2018) using a more powerful generative adversarial network called
“Boundary equilibrium generative adversarial network” as a generator, with the main
goal of improving the quality of the generated stego-images without compromising the
security. In addition to this, the recovery accuracy of the proposed Steganography
without embedding scheme is increased by incorporating a Reed-Solomon error
correction mechanism. Experimental results show that all of this goals have been
achieved. The images generated are more realistic and better represent the real dataset
having an inception score of 5.62 over celebA dataset, and on average an 8.2% increase
in recovery accuracy when compared to (Hu et al., 2018).