Abstract:
Steganography, or the concealment of information in digital media, is a credible cyber threat as it
allows concealed communication, data exfiltration, and cyber spying. As compared to
cryptographic methods for concealing information via encryption, steganographic methods
conceal information inside such a format that the digital carriers are viewed visually unscathed
and hence difficult to find and available for forensic examination. This paper solves two
fundamental issues of steganalysis: image quality assessment (IQA) and payload classification,
which are essential in enhancing cyber defense and forensic analysis of the digital world. The
Author introduces a mechanism that integrates transfer learning and multitask learning approaches
to solve these issues.
Dependence on the representational ability of pre-trained convolutional neural networks (CNNs),
our method trains these models on multitasks of payload classification and IQA. The model can
find various steganographic payloads, i.e., JMiPOD, and UERD, and evaluate the degree of
distortions introduced due to steganographic embedding. Multitask learning architecture enables
the model to achieve generalization of learned features across tasks but with reduced
computational complexity. Verification experiments confirm the efficiency of our solution. The
payload classifier module obtained an accuracy of 90%, strongly showing excellent discrimination
between various steganographic techniques.
