Year: 2019
Type: article
Abstract: Steganography, the process of hiding secret information within normal-looking data (a cover), lies at the intersection of computer science and security. Recently, these covers have come in the form of images. The steganographic task must pass two visual inspections: by humans and by machines. If a human examines the image and fails to spot the hidden data, the algorithm successfully passes the first test. The second task is preventing a machine from detecting patterns to reverse engineer the secret information. In 2017, researchers achieved some success [1], but there were two main issues: the steganography only worked with (a) fully saturated, (b) fixed-size (100 x 100) images. To curb these limitations, a new pipeline is explored to generate non-fixed size cover images with steganographic modification rather than embedding. This paper explores this new form of steganography with several key processes. First, the secret information is encrypted before combining it with the cover using neural cryptography. Second, the information hides in the stroke data of a person's handwriting on a white background, increasing task difficulty, forcing the steganographic approach to be robust to a plethora of data, including sparse images. In this sense, the strokes are directly modified, rather than inserted or embedded in-between pixels. The result is a toy problem utilizing realistic, generated, coordinate sequences of human handwriting modified with slight offsets dependent on the information combined with the coordinates in the sequence. With these slight offsets, the new generated coordinates are nearly identical to the original coordinates, preserving the primary structure of the handwriting, but shining light on a new avenue of steganography based on data modification rather than embedding.
Author: James Hahn
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2944220150