Encrypted deep robust reversible image watermarking with preset private key

Abstract

Robust reversible watermarking (RRW) is an important branch of digital watermarking, which supports lossless recovery of original watermarks in lossless channels, while ensuring reliable watermark extraction against various distortions and attacks in lossy channels. However, existing deep learning-based RRW methods still suffer from insufficient security, irreversible quantization errors, severe cumulative feature errors, and low stego image quality. To alleviate these issues, this paper proposes an encrypted deep robust reversible image watermarking (EDRRW) model with a deeply integrated preset private key. The proposed model introduces four core innovative designs. First, an attenuation factor mechanism is developed to flexibly regulate the influence of watermark feature propagation on the cover image. Second, the preset private key is embedded into the entire invertible network to strengthen the security of watermark embedding and extraction. Third, a lightweight 4-layer invertible architecture is constructed to balance error suppression and watermark robustness. Fourth, wavelet constraints are introduced into reversible watermarking for the first time to maintain the overall hidden quality within an acceptable range. Extensive experimental results show that the proposed EDRRW model achieves an average PSNR of 41.49 dB on public test datasets, outperforming state-of-the-art methods. Meanwhile, the model exhibits strong robustness against common image distortions and attacks, and enables secure and controllable watermark extraction relying on the correct private key. These results fully validate the effectiveness and superiority of the proposed method.