Work

Deep Learning-Based Image Steganography with Latent Space Embedding and Smart Decoder Selection

Year: 2025

Type: article

Abstract: Image steganography is crucial for secure communication, enabling covert data embedding within cover images. While traditional methods such as LSB embedding are vulnerable to detection, deep learning techniques like GANs and autoencoders have improved performance, yet they still struggle with dynamic adaptation to diverse secret data types, limited training datasets, and resilience to distortions. To address these issues, we propose a flexible framework with adaptive multi-encoder-decoder pairs, extensive dataset training, and an optimized architecture with specialized components. Our model achieves significant improvements in Secret Recovery Accuracy (SRA), Stego-Image Quality (SSIM, PSNR), and robustness to noise, with SSIM of 0.99 and recovery accuracy over 98%. It also reduces the detection rate, with an AUC approaching 0.5 in steganalysis. These results set a new benchmark for secure image transmission and privacy-preserving communication.

Source: Entropy

DOI: 10.3390/e27121223

Authors: Yiqiao Zhou, Na Wang, Xiaolong Hong, Yanchun Peng, Shuo Shao

Institution: University of Shanghai for Science and Technology

Language: English

Topic: Advanced Steganography and Watermarking Techniques

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $3,137

Funders: National Natural Science Foundation of China, National Key Research and Development Program of China