Applied Soft Computing 67 (2018) 505–518 Contents lists available at ScienceDirect Applied Soft Computing journal homepage: www.elsevier.com/locate/asoc Robust steganographic method based on unconventional approach of neural networks Robert Jarusek, Eva Volna, Martin Kotyrba ∗ University of Ostrava, Department of Informatics and Computers, 30. dubna 22, 70103, Ostrava, Czech Republic a r t i c l e i n f o Article history: Received 7 September 2017 Received in revised form 14 March 2018 Accepted 18 March 2018 Available online 26 March 2018 Keywords: Steganography Robustness Watermark Neural network Discrete cosine transform – DCT a b s t r a c t The article deals with the issue of using an apparatus of neural networks in the area of steganography. A new method called STEGONN was proposed. The proposed method is robust enough to an attack and the hidden message hard to be falsified. The core of our work lies in a design and implementation of a method for the use of neural networks as a native coder and decoder of a secret message (digital watermark) with an emphasis on the minimum necessary level of image data modification – covermedium. A covermedium is not perceived as a passive cover of a secret message, but we make active use of cover medium data, primarily its data markers (image markers) to insert a secret message. The advantage over other stegano- graphic methods is the fact that the method implicitly offer the possibility to detect corrupted parts of the stegomedium and inform the user about possible manipulation with the image. The characteristics of the proposed method have been experimentally verified and compared with commercially available steganographic applications. © 2018 Elsevier B.V. All rights reserved. 1. Introduction The aim of steganography is to hide a message (information, data) in a place where no one expects it while its presence is not detectable. Authors [11] mention three aspects of assessing systems for hiding information: • capacity • security • robustness Capacity is understood as the amount of information which can be hidden in the medium, security represents impossibility of an intruder to reveal the secret message, and robustness expresses the capability of the medium to hold hidden data even after its possible modification. The steganographic process is depicted in Fig. 1 The issue of hiding information in graphical data is complicated primarily due to its demands on sufficient robustness and capac- ity to insert a secret message into image data, where the data should be modified the least possible. An ideal state is when a secret information is inserted into image data so that the image data itself (covermedium) is not modified in any way, i.e. cover- medium = stegomedium. ∗ Corresponding author. E-mail addresses: robert.jarusek@osu.cz (R. Jarusek), eva.volna@osu.cz (E. Volna), martin.kotyrba@osu.cz (M. Kotyrba). 2. Current steganographic methods Currently, there are four basic steganographic methods, namely: End Of File (EOF) [3], Least significant bit (LSB) [9], and meth- ods based on the use of Discrete Cosine Transform (DCT) [1] or Discrete Wavelet Transform (DWT) [4]. Models to effectively cal- culate the degree of digital image’s trustworthiness are mentioned in [14]. Table 1 provides a comparison of selected Steganographic approaches. The levels where individual algorithms meet the requirements are defined as none, low, medium, and high. The comparative criteria are as follows: • Visibility: subjective visibility of a secret message in a ste- gomedium • PSNR (peak signal-to-noise ratio) expresses the ratio between the maximum possible energy of signal and the energy of noise defined by the relation (1) PSNR = 10 · log 10  MAX 2 MSE  = 20 · log 10  MAX √ MSE  (1) where MAX is the maximum possible value of a pixel in an image (i.e. 255 for 8 bits per channel). https://doi.org/10.1016/j.asoc.2018.03.023 1568-4946/© 2018 Elsevier B.V. All rights reserved.