ItemConference Paper

Evaluating Generalization and Robustness of U-Net Based Image Steganography

Title
Evaluating Generalization and Robustness of U-Net Based Image Steganography
Creator
Adithiya, Dhanush; Jayabalan, Bhuvana; Beaulah Soundarabai, P.; Augustin, Peter
Description
This paper investigates the effectiveness and generalization ability of U-Net based image steganography models across multiple datasets, with comparisons to the classical Least Significant Bit (LSB) substitution method. Models were trained on STL-10, CIFAR-10, and Stanford Cars datasets and evaluated both in-distribution and on out-of-distribution internet images. Results show that the STL-10 model consistently achieved the best trade-off between imperceptibility and recovery quality, while the CIFAR-10 model failed to generalize due to its low resolution and limited diversity. Baseline experiments confirmed that LSB achieves extremely high PSNR and SSIM at low payloads, but suffers sharp increases in Bit Error Rate (BER) under higher payloads or even mild distortions such as JPEG compression and Gaussian noise. By contrast, the U-Net model provided more stable recovery and greater robustness, highlighting the advantages of learned feature embeddings over handcrafted substitution. These findings underscore the importance of dataset diversity and robustness testing in developing practical steganographic systems for real-world deployment. 2025 IEEE.
Source
Proceedings - 2025 International Conference on Transformative Computing Technologies, ICTCT 2025;pp.178-183
Date
01-01-2025
Publisher
Institute of Electrical and Electronics Engineers Inc.
Subject
Deep Learning; Generalization; Image Steganography; Robustness; U-Net
Coverage
Adithiya D., Christ (Deemed to be) University, Department of Computer Science, Bangalore, India; Jayabalan B., Christ (Deemed to be) University, Department of Computer Science, Bangalore, India; Beaulah Soundarabai P., Christ (Deemed to be) University, Department of Computer Science, Bangalore, India; Augustin P., Christ (Deemed to be) University, Department of Computer Science, Bangalore, India
Rights
Restricted Access; Hardcopy may be available in the library
Relation
ISBN: 979-833159195-3;
Format
online
Language
English
Type
Conference paper

Citation

Adithiya, Dhanush; Jayabalan, Bhuvana; Beaulah Soundarabai, P.; Augustin, Peter, “Evaluating Generalization and Robustness of U-Net Based Image Steganography,” CHRIST (Deemed To Be University) Institutional Repository, accessed June 18, 2026, https://archives.christuniversity.in/items/show/26132.

Output Formats