Autoencoder-based invisible watermarking: A lightweight deep learning approach for robust embedding and recovery

This paper presents a lightweight yet robust invisible watermarking framework based on deep autoencoders implemented in TensorFlow. Unlike traditional watermarking methods relying on handcrafted transformations such as DCT, DWT, or SVD, our approach uses a neural network to learn the embedding and extraction processes in an end-to-end manner. We demonstrate that a simple autoencoder trained on a single host image can achieve high imperceptibility and strong robustness against common image processing attacks, including JPEG compression and Gaussian noise. Experimental results show an SSIM of 0.9790 and PSNR of 27.28 dB between the original and watermarked images, with watermark recovery correlation reaching 0.9726 even after JPEG compression and 1.0000 after Gaussian noise, validating the method’s effectiveness for real-world applications.