DETECTION OF STEGANOGRAPHY IN IMAGES USING LIGHTWEIGHT DEEP LEARNING MODELS
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
This paper presents a comprehensive experimental study on the applicability of lightweight convolutional neural networks (CNNs) for the task of digital image steganalysis under resource-constrained conditions. The main objective of the research is to evaluate whether compact architectures such as MobileNetV2 and EfficientNetV2S can achieve high detection accuracy when combined with appropriate high-frequency preprocessing, while significantly reducing computational and memory requirements compared to heavy specialized steganalytic networks. The study demonstrates that high-pass filtering (HPF) is a critical prerequisite for effective steganography detection. Without explicit suppression of image content, lightweight CNNs fail to distinguish weak stego-noise from natural image structures. Experimental results show that an insufficient number of HPF layers or filters leads to rapid overfitting and unstable classification performance. In contrast, the use of a sufficiently expressive and structured HPF preprocessing block enables stable convergence and high generalization ability. Special attention is paid to the influence of fixed versus trainable HPF filters. It is shown that allowing HPF kernels to adapt during training provides an additional performance gain by adjusting to the statistical characteristics of stego-noise. This hybrid approach effectively combines prior knowledge from classical steganalysis (SRM-based filters) with the adaptive capabilities of deep learning. The proposed detection framework was evaluated using images with data embedded by the least significant bit (LSB) method under various payload conditions. Both MobileNetV2 and EfficientNetV2S demonstrated high detection accuracy and near-optimal ROC characteristics while maintaining low computational complexity. These results confirm that lightweight CNN architectures, when properly augmented with high-frequency preprocessing, represent a viable solution for practical steganalysis, including deployment on embedded or mobile platforms. However, the current evaluation is limited to LSB-based embedding. Future work will focus on extending the proposed approach to more sophisticated adaptive steganographic algorithms and on improving robustness across different image domains.
Fields of Study & Subjects
Advanced Steganography and Watermarking Techniques
Digital Media Forensic Detection
Image and Video Quality Assessment
Publication details
- Journal / source
- Cybersecurity Education Science Technique
- Volume
- 4
- Issue
- 32
- Pages
- 336–336
- Publication date
- 26 March 2026
- Language
- EN