ReliaGuard: Reconstruction-Aware Watermarking for Source Attribution and Blind Edit Localization

Hailong Wang, Wuzheng Tan, Zixin Zhang, Xujun Fang, Fengjuan GongarXiv 2026 · arXiv

TL;DR

ReliaGuard is a novel watermarking framework designed to track image provenance and pinpoint generative edits, even when a model reconstructs the entire image background. By making the watermark "reconstruction-aware" and using image context to interpret signals, it improves edit localization accuracy by over 45% compared to previous state-of-the-art methods.

What Problem Does This Paper Address?

Generative image editing (e.g., using Stable Diffusion to swap a background or add an object) creates a unique challenge for digital forensics. Most generative editors don't just "paste" new pixels; they pass the entire image through a Variational Autoencoder (VAE). This process, known as full-output regeneration (FR), introduces subtle reconstruction changes across the entire image, not just the edited area.

Existing proactive watermarks—signals embedded before editing to detect later changes—often fail here. They tend to flag these global reconstruction artifacts as "tampering," leading to a sea of false positives in the unedited background. The core challenge is distinguishing between a legitimate generative edit (the target) and the unavoidable "noise" of the generative model's reconstruction process.

Methods

ReliaGuard tackles this through a two-stage approach: Reconstruction-aware Evidence Writing (REW) and Evidence-guided Blind Reading (EBR).

1. The Writer (REW)

Instead of just embedding a watermark into the original pixels, REW optimizes the signal so it survives the VAE bottleneck. It ensures the keyed feature responses are stable in both the protected image and its VAE-reconstructed version. This is done under strict constraints to ensure the image remains visually identical to the original and the source identifier (the "who made this" bit) stays recoverable.

2. The Reader (EBR)

The reader is "blind," meaning it doesn't need the original image to work. It uses two clever sub-modules to interpret the watermark:

  • Background-contrast Evidence Refinement (BER): This module looks at the image content itself. If a watermark signal is weak, BER uses local image context to decide if that weakness is due to a real edit or just reconstruction artifacts in the background.
  • Native-guided Scale Reassembly (NSR): To ensure the predicted edit mask isn't just a blurry blob, NSR fuses high-resolution spatial details with lower-resolution context, allowing for sharp, pixel-level localization of complex shapes.
The overall ReliaGuard framework, showing how evidence is written to survive reconstruction and then read using image context.

Datasets

The authors evaluated ReliaGuard across three distinct image domains to ensure generalizability:

  • COCO: Common objects in real-world scenes.
  • ImageNet: A standard large-scale computer vision dataset.
  • Generated: Images created by AI models themselves.

The test set included 300 source images and 900 editing cases, covering various mask shapes and sizes (from 3% to 35% of the image area). They specifically compared Full-output Regeneration (FR)—where the whole image is re-processed—against Local Splicing (SP)—where only the edited pixels are changed.

Key Insights and Ideas

  • The VAE is the Enemy: The biggest hurdle for localization isn't the edit itself, but the VAE reconstruction. ReliaGuard’s "aha moment" is treating the VAE as a known degradation to be optimized against during the writing phase.
  • Context-Aware Reading: A watermark reader shouldn't just look at the watermark signal in a vacuum. By "looking" at the image (via the BER module), the reader can better distinguish between a modified object and a reconstructed background.
  • Dual-Path Protection: The framework successfully decouples source attribution (recovering a 96-bit ID) from localization. This ensures that even if an edit is hard to find, the owner of the original image can still be identified with 100% accuracy.

Key Results

ReliaGuard significantly outperforms existing baselines, particularly in the difficult "Full Regeneration" (FR) scenario:

  • Localization Accuracy: On FR images, ReliaGuard achieved an 85.61% Intersection-over-Union (IoU), beating the strongest baseline (APT) by a massive 45.02 percentage points*.
  • Source Attribution: It maintained 100% bit accuracy for the source identifier in both FR and spliced conditions.
  • Robustness: Even after common image processing (like JPEG compression or resizing), the method retained an IoU of over 81%.
Fig. 2. FR localization examples for active methods across three image domains and four target areas. Each row uses the same source and target mask across methods; each method receives its own protected and edited image. The first column shows our received image. IoUs are evaluated at the fixed thresholds in

Limitations and Future Work

While highly effective, ReliaGuard still faces challenges:

  • Small & Complex Targets: Very small edits (3% of the image) or shapes with high perimeter-to-area ratios remain the most difficult to localize perfectly.
  • Geometric Attacks: Large geometric transformations, such as 90-degree rotations or heavy cropping, can still degrade localization performance.
  • Embedding Distortion: The authors note that the stronger localization performance comes at the cost of slightly higher embedding distortion compared to "fragile" watermarks, though it remains within acceptable visual limits (PSNR ~32.7 dB).

Future work will likely focus on reducing this embedding distortion and improving the system's resilience to extreme geometric shifts.

Summary updated

Abstract

Generative image editing facilitates content creation but complicates source attribution and edit verification. Existing watermarking methods combine copyright protection with tamper localization, yet background reconstruction can weaken spatial evidence even when source identifiers remain recoverable. We propose ReliaGuard, a reconstruction-aware watermarking framework for source attribution and blind edit localization. The writer optimizes keyed feature responses in both the protected image and its variational autoencoder reconstruction under image-distortion and source-preservation constraints. Using only the received image and a secret key, the reader combines watermark responses with local image context through residual evidence refinement and predicts an edit mask through native-resolution-guided multiscale reading. It is trained on complete editing outputs, locally spliced counterparts, and processed versions of both. On an independent test set of 300 source images and 900 editing cases from three image domains, our method achieves edit-region intersection-over-union scores of 85.61% for complete outputs and 94.90% for locally spliced outputs at a threshold calibrated to a nominal 1% background false-positive rate on complete outputs. The complete-output result exceeds the strongest baseline by 45.02 percentage points, while decoded identifier-bit accuracy reaches 100% in both conditions. Across seven families of image processing, the corresponding localization scores are 81.01% and 86.49%, with identifier-bit accuracies of 92.76% and 92.79%, respectively. These results support source recovery and blind edit localization under background reconstruction and subsequent image processing.
BibTeX
@misc{wang2026reliaguard,
  author = {Hailong Wang and Wuzheng Tan and Zixin Zhang and Xujun Fang and Fengjuan Gong},
  title = {{ReliaGuard: Reconstruction-Aware Watermarking for Source Attribution and Blind Edit Localization}},
  year = {2026},
  eprint = {2611.16368},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.16368},
}
Figure