AttenHideNet: A novel deep learning-based image steganography method using a lightweight U-net with soft attention

Image-to-image steganography, embedding secret information within images while preserving visual quality, has become essential due to growing demands for secure and efficient digital communication. Traditional methods often struggle to achieve high embedding capacity without sacrificing imperceptibility. Recent advancements in deep learning have offered promising solutions by enabling more complex data embedding strategies. In this paper, we propose AttenHideNet, a novel deep learning-based steganography method leveraging a lightweight U-Net architecture (<1.2 million parameters) combined with soft attention mechanisms. By utilizing the YUV color space instead of RGB, our method significantly improves embedding efficiency, capacity, and visual imperceptibility. AttenHideNet achieves an embedding capacity of up to 24 bits per pixel (bpp) while maintaining high visual quality. The soft attention mechanism dynamically identifies and prioritizes embedding in less perceptually sensitive image regions. Experimental results on benchmark datasets demonstrate that AttenHideNet achieves superior visual quality (PSNR up to 52.67 dB) compared to state-of-the-art methods, with low latency (18 ms/image) and minimal memory usage (4.11 MB), making it suitable for real-time applications. Despite these advantages, the method shows limited robustness under firm JPEG compression and geometric transformations, highlighting essential directions for future research.