PIRNet: Privacy-Preserving Image Restoration Network via Wavelet Lifting

Xin Deng, Chao Gao, Mai XuICCV 2023 · DOI

Abstract

Cloud-based image restoration offers significant computational advantages but introduces critical privacy concerns. Traditional solutions relying on homomorphic encryption are often computationally prohibitive and visually conspicuous. This paper presents PIRNet, a novel privacy-preserving image restoration network that operates in the steganographic domain. By leveraging a Lifting-based Invertible Hiding (LIH) network and a Lifting-based Secure Restoration (LSR) network, PIRNet conceals sensitive images within natural-looking "stego" images. This allows for secure denoising, deblurring, and super-resolution on public servers. Experimental results on the DIV2K dataset demonstrate that PIRNet significantly outperforms existing state-of-the-art pipelines in both restoration quality and security efficiency.

Introduction

The proliferation of cloud-based image processing services has provided users with access to powerful restoration tools. However, sending private images to third-party servers exposes sensitive data to potential breaches by malicious providers or external hackers. While image encryption is a common defense, encrypted data is often visually distinct (appearing as random noise), which can flag it for targeted attacks. Furthermore, advanced cryptographic techniques like homomorphic encryption, while secure, are too computationally expensive for real-time restoration tasks.

PIRNet addresses these challenges by moving image restoration into the steganographic domain. Instead of encrypting data, PIRNet hides a secret low-quality image within a natural-looking "cover" image. The resulting "stego" image can be processed by a public server, and the restored secret image can then be recovered by the client.

Methodology

The PIRNet framework is built upon two primary components: the Lifting-based Invertible Hiding (LIH) network and the Lifting-based Secure Restoration (LSR) network.

Overview of the PIRNet framework showing the interaction between the private client and public server.

Lifting-based Invertible Hiding (LIH)

The LIH network, located on the client side, is responsible for concealing the secret image within a cover image. By utilizing an invertible wavelet lifting scheme, the network ensures that the hiding process is information-lossless and highly stable. This reversibility is crucial for the subsequent "Reveal" step, where the restored secret image is extracted from the processed stego image.

Lifting-based Secure Restoration (LSR)

The LSR network resides on the public server. It is designed to perform restoration tasks—such as denoising, deblurring, or super-resolution—directly on the stego image. The authors discovered a synergistic relationship between steganography and restoration: improvements in the quality of the stego container consistently translate to quality improvements in the hidden secret image.

Detailed architecture of the Lifting-based Invertible Hiding (LIH) and Secure Restoration (LSR) networks.

The LSR network employs a multi-scale approach, utilizing Wavelet Lifting to decompose images into low and high-frequency bands. This is particularly effective for restoration, as tasks like deblurring focus heavily on recovering high-frequency details. The restoration modules themselves utilize a U-Net style architecture with Non-linear Activation Free Blocks (NAFBlocks) for state-of-the-art efficiency.

Experimental Results

PIRNet was evaluated on the DIV2K dataset across three restoration tasks: denoising, deblurring, and 4x super-resolution. To establish a baseline, the authors combined top-tier hiding methods (e.g., HiDDeN, DeepMIH) with leading restoration models (e.g., NAFNet, MIRNetv2).

PIRNet demonstrated superior performance across all benchmarks:

  • Denoising: Achieved a PSNR improvement of 2.11dB over the best competing pipeline.
  • Deblurring: Improved PSNR by 1.56dB.
  • Super-resolution: Improved PSNR by 0.79dB.

Beyond quantitative metrics, the stego images produced by PIRNet exhibited higher visual fidelity to the original cover images, avoiding common artifacts like texture-copying or color distortion found in other hiding methods.

Conclusion

PIRNet represents a significant shift in privacy-preserving image processing by demonstrating the viability of restoration in the steganographic domain. By prioritizing visual "invisibility" and computational efficiency, it offers a practical alternative to encryption-based methods. Future work will focus on enhancing resistance to statistical steganalysis and developing more generalized restoration modules that can handle multiple degradation types within a single network.

Summary updated

BibTeX
@inproceedings{deng2023pirnet,
  author = {Xin Deng and Chao Gao and Mai Xu},
  title = {{PIRNet: Privacy-Preserving Image Restoration Network via Wavelet Lifting}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2023},
  doi = {10.1109/iccv51070.2023.02044},
}
Figure