HIPP: Protecting Image Privacy via High-Quality Reversible Protected Version

Xi Ye, Lina Wang, Run Wang, Jiatong Liu, Geying YangIJCAI 2025 · DOI

Sharing personal photographs on Social Network Platforms (SNPs) has become a ubiquitous part of modern life. However, this practice exposes sensitive personal information—such as identity, location, and lifestyle habits—to potential misuse by malicious actors and automated surveillance systems. Traditional image privacy techniques generally fall into two categories: full encryption, which renders images as unintelligible noise and ruins the user experience, or partial obfuscation (e.g., blurring or pixelation), which is often visually jarring and fails to hide all sensitive features.

Recently, "thumbnail-preserving" methods have attempted to balance privacy and utility by maintaining a recognizable low-resolution version for browsing while hiding high-resolution details. However, these methods often introduce significant visual artifacts, signaling to attackers that an image is protected. To address these limitations, we present HIPP, a framework that generates natural-looking, high-quality protected images that are perfectly reversible for the original owner.

Comparison between traditional protection (a) and the proposed HIPP framework (b). HIPP generates natural-looking images that mislead attackers with fake data instead of triggering error messages.

Methodology: Latent Vector Decoupling

The core innovation of HIPP lies in its approach to image manipulation within the latent space of generative models rather than the pixel space. The framework utilizes pre-trained models such as Glow (for mathematical reversibility) and StyleGAN (for visual quality) to map images to latent vectors.

A custom-designed Detail Information Extractor (E) performs Latent Vector Decoupling, splitting an image's latent vector ($a$) into two distinct components:

  1. Contour Component ($a^{th}$): This captures the fundamental structural information, essentially the "thumbnail" or general layout of the image.
  2. Detail Component ($a^{cot}$): This contains high-frequency, private information such as specific facial features or identifying text.

To protect an image, HIPP retains the original contour vector but replaces the detail vector with one derived from a different image or a random sequence. The resulting "protected" image appears as a high-quality, natural photograph to the human eye and AI analyzers, but it contains "fake" details. Because the mapping process is reversible, the original detail vector can be securely recovered using a private key, allowing for the perfect reconstruction of the original high-fidelity image.

Experimental Evaluation

The HIPP framework was evaluated using standard benchmarks, including the CelebA and Helen face datasets, and the LSUN scene dataset. The experiments demonstrated that HIPP significantly outperforms state-of-the-art (SOTA) methods across multiple dimensions.

Visual results across different datasets. Rows show (from top to bottom): Original images, protected versions from SOTA methods, HIPP's natural-looking protected versions, and HIPP's restored images.

Key Results

  • Restoration Fidelity: The system achieved a Structural Similarity Index (SSIM) of 0.9899, indicating that restored images are virtually indistinguishable from the originals.
  • Visual Naturalness: Unlike previous methods that produce noisy or "broken" images, HIPP's protected versions look like authentic photos, effectively acting as a "honeypot" for malicious AI.
  • Efficiency: HIPP is highly performant, with a runtime of only 0.07 seconds per image. This is orders of magnitude faster than competing methods like BIPU (92.6s).
  • Storage Overhead: The file expansion rate was kept to a minimum of 1.1046, ensuring that protected images do not consume excessive storage or bandwidth.

Conclusion

HIPP introduces a "deception as defense" paradigm for image privacy. By decoupling structural and detailed information in the latent space, it provides a solution that is aesthetically pleasing for social media users while remaining highly secure against malicious analysis. Future work will focus on refining the decoupling process to eliminate any residual information leakage and extending the framework to support video privacy protection.

Summary updated

BibTeX
@inproceedings{ye2025hipp,
  author = {Xi Ye and Lina Wang and Run Wang and Jiatong Liu and Geying Yang},
  title = {{HIPP: Protecting Image Privacy via High-Quality Reversible Protected Version}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/253},
}
Figure