Robust steganography: training and error-correction coding against active wardens

Adversarial training can produce steganographically secure models, but their practical utility is limited by a critical fragility to degradations common in real-world channels. This thesis addresses this gap by framing such degradations as the work of an active warden and develops a methodology to create a robust communication system. The research first quantifies the baseline model’s fragility against a suite of attacks. We then systematically investigate data augmentation strategies, demonstrating that training on simple geometric shifts fails to confer spatial invariance. A novel training approach using interpolation, specifically crop-and-resize augmentation, is then proposed and tested. Finally, to address residual errors, we evaluate a range of Error Correcting Codes (ECCs), from simple repetition codes to advanced LDPC codes, to characterize the channel and achieve reliability. Our findings demonstrate that interpolation-based training successfully instills generalized robustness, significantly reducing bit error rates against attacks that previously caused complete failure. The subsequent analysis of ECCs reveals that the steganographic channel under degradation behaves as a Binary Symmetric Channel with a diffuse error profile. Consequently, simple bit-level codes, like 3-repetition coding, are shown to be significantly more effective than complex, block-based codes like Reed-Solomon. This work presents a complete, two-stage methodology that transforms a fragile, theoretical model into a practical and resilient steganographic system by combining interpolation-based training for robustness with appropriate ECC selection for reliability.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Hong, Seunghwan

Contributors dc:contributor

• Moulin, Pierre

• Steganography
• Adversarial Training
• Error-correcting Codes
• Data Augmentation
• Active Warden

Statement dc:rights

• Copyright 2025 Seunghwan Hong

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129993

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129993