A convolutional neural network-based linguistic steganalysis for synonym substitution steganography

Area of Science:

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Background:

• Steganalysis aims to detect hidden information within digital media.
• Detecting linguistic steganography, especially text modified by synonym substitutions, presents unique challenges.
• Existing methods may struggle with nuanced textual modifications.

Purpose of the Study:

• To propose a novel linguistic steganalysis method for detecting texts modified by synonym substitutions.
• To enhance the accuracy and robustness of steganographic text detection systems.
• To leverage deep learning, specifically convolutional neural networks (CNNs), for this task.

Main Methods:

• A two-level cascaded CNN architecture was developed.
• The first level employed a sentence-level CNN using pre-trained word embeddings to extract sentence features.
• The second level utilized a text-level CNN to classify texts as stego or cover based on sentence-level features.

Main Results:

• The sentence-level CNN effectively extracted features for steganalysis, achieving an average accuracy of 82.245%.
• The cascaded CNN approach significantly improved the detection performance for stego texts.
• The method demonstrated strong capabilities in distinguishing between modified (stego) and unmodified (cover) texts.

Conclusions:

• The proposed two-level cascaded CNN method is effective for linguistic steganalysis.
• The approach shows promise for robust detection of synonym substitution-based steganography.
• This work contributes to advancing the field of text-based steganography detection.