MBRS : Enhancing Robustness of DNN-based Watermarking by Mini-Batch of Real and Simulated JPEG Compression (2108.08211v1)

Abstract: Based on the powerful feature extraction ability of deep learning architecture, recently, deep-learning based watermarking algorithms have been widely studied. The basic framework of such algorithm is the auto-encoder like end-to-end architecture with an encoder, a noise layer and a decoder. The key to guarantee robustness is the adversarial training with the differential noise layer. However, we found that none of the existing framework can well ensure the robustness against JPEG compression, which is non-differential but is an essential and important image processing operation. To address such limitations, we proposed a novel end-to-end training architecture, which utilizes Mini-Batch of Real and Simulated JPEG compression (MBRS) to enhance the JPEG robustness. Precisely, for different mini-batches, we randomly choose one of real JPEG, simulated JPEG and noise-free layer as the noise layer. Besides, we suggest to utilize the Squeeze-and-Excitation blocks which can learn better feature in embedding and extracting stage, and propose a "message processor" to expand the message in a more appreciate way. Meanwhile, to improve the robustness against crop attack, we propose an additive diffusion block into the network. The extensive experimental results have demonstrated the superior performance of the proposed scheme compared with the state-of-the-art algorithms. Under the JPEG compression with quality factor Q=50, our models achieve a bit error rate less than 0.01% for extracted messages, with PSNR larger than 36 for the encoded images, which shows the well-enhanced robustness against JPEG attack. Besides, under many other distortions such as Gaussian filter, crop, cropout and dropout, the proposed framework also obtains strong robustness. The code implemented by PyTorch \cite{2011torch7} is avaiable in https://github.com/jzyustc/MBRS.

• The paper introduces MBRS to robustly embed watermarks by addressing JPEG compression's non-differentiability.
• It employs mixed noise layers with SE blocks, message processors, and diffusion blocks to optimize watermark integrity.
• Experiments show a BER below 0.01% and PSNR above 36 at JPEG quality 50, underscoring its practical effectiveness.

Enhancing DNN-based Watermarking Robustness Using Mini-Batch Real and Simulated JPEG Compression

The paper "MBRS: Enhancing Robustness of DNN-based Watermarking by Mini-Batch of Real and Simulated JPEG Compression" by Jia et al. presents a nuanced approach to digital watermarking in the context of JPEG compression, a common yet challenging perturbation for such algorithms. Contemporary digital watermarking methods based on deep neural networks (DNNs) have predominantly relied on autoencoder-like structures. These integrate an encoder, noise layer, and decoder to endeavor robust message embedding and extraction. The authors identify a critical shortcoming: existing methods do not adequately handle the non-differentiable nature of JPEG compression, thereby affecting the robustness of these watermarking strategies.

The authors propose the Mini-Batch of Real and Simulated JPEG Compression (MBRS) method. This innovative approach incorporates three types of noise layers into the network: real JPEG compression, simulated JPEG compression, and a noise-free (identity) layer, chosen at random for each mini-batch during training. The objective is to harness the strengths of both differential training mechanisms and real-world JPEG compression via a comprehensive training framework.

Several architectural enhancements accompany the MBRS method. The authors incorporate Squeeze-and-Excitation (SE) blocks into the encoder-decoder framework to exploit frequency domain features better. Additionally, they introduce a "message processor" to optimize the expansion and redundancy of embedded messages. An added diffusion block further supplements the robustness against cropping attacks, a common type of image distortion.

Empirically, the proposed architecture demonstrates substantial improvement over state-of-the-art models in both message integrity—quantified by a significantly lower bit error rate (BER)—and visual fidelity, indicated by higher PSNR values. Under JPEG compression with a quality factor of 50, the paper reports a BER of less than 0.01%, with a PSNR exceeding 36. This showcases notable resistance to JPEG distortions while maintaining high image quality. Furthermore, robustness tests against a spectrum of attacks, including Gaussian filtering, dropout, cropout, and other non-JPEG distortions, affirm the scheme's efficacy.

This paper's insights feature critical implications in practical and theoretical contexts. Practically, it bolsters the reliability of watermarking methodologies in digital media protection, a critical aspect amid increasing digital content proliferation. Theoretically, the successful integration of real and simulated JPEG environments navigates the challenge of non-differentiability, potentially inspiring future architectures in watermarking and related neural network applications.

The MBRS method's adaptability and robustness could spur further research into dynamic noise adaptation and real-world applicability across varied compression and distortion scenarios. Such developments could significantly influence the next generation of watermark embedded systems, extending their utility and effectiveness in protecting digital multimedia assets.

In conclusion, this paper bridges a crucial gap in DNN-based watermarking, tackling JPEG-specific vulnerabilities with an inventive and comprehensive approach. It sets a precedent for future watermarks that are adaptive and resilient across increasingly variable multimedia landscapes. The methodological approach and promising results open avenues for successive enhancements in the domain, suggesting further exploration and innovation in watermarking robustness technologies.