Research on Adaptive Watermarking and Secure Traceability Algorithm for Digital Media Images Based on Deep Learning

Abstract

Digital image copyright protection and traceability are critical due to widespread online dissemination. Conventional watermarking methods suffer from inherent trade-offs among robustness, imperceptibility, and security, and exhibit limited resilience against geometric distortions and deepfake-based adversarial manipulations. This paper proposes an adaptive framework built on an encoder-noise-decoder architecture, integrating channel-spatial attention and deformable convolutional networks (DCN) to enhance geometric robustness. Experiments demonstrate 98.7% extraction accuracy under combined attacks, 5.2 dB PSNR improvement over baselines, and effective defense against DLOVE overwriting attacks.