Deep Clustering Network for Steganographer Detection Using Latent Features Extracted from a Novel Convolutional Autoencoder

Steganography is typically used by law enforcement agencies to prevent unauthorized persons from becoming aware of the existence of a message communicated by military or other government organizations outside their network. Illegal uses of steganography such as fraud, gambling, criminal communications, hacking, electronic payments, harassment, offenses on intellectual property and viruses pose a great threat to society. Such illicit steganographers communicate with each other through stego files to exchange their plan without getting noticed by law enforcement. This paper presents a novel Deep Clustering Network for Steganographer Detection (DCNSD) based on convolutional autoencoders and a clustering model for identifying such steganographers since plenty of digital images are transferred over the internet that could carry hidden secret messages. Mostly the existing techniques involving deep learning networks for steganographer detection involves supervised learning approach which makes them unsuitable for real-world deployment. The foremost characteristic of this proposed network lies in its ability to segregate images transmitted from a steganographer from those innocent users’ images in an unsupervised approach. Thereby making the proposed DCNSD system plausible for real-world deployment. Based on the experimental results, it is shown that the proposed DCNSD framework excels in detecting steganographers who use content-adaptive steganography to embed secrets with 100% accuracy.