An Efficient and Adaptive Watermark Detection System with Tile-based Error Correction

Xinrui Zhong, Xinze Feng, Jingwei Zuo, Fanjiang Ye, Yi Mu, Junfeng Guo, Heng Huang, Myungjin Lee, Yuke WangarXiv 2025 · DOI · arXiv

Abstract

As generative AI scales, verifying billions of daily uploads for content attribution becomes a significant computational challenge. Current state-of-the-art watermarking methods are often too slow for real-time pipelines. This paper introduces QRMark, a high-performance watermark detection system that speeds up verification by an average of 2.43×. QRMark achieves this by processing small image "tiles" rather than full frames, utilizing QR-code-inspired Reed-Solomon error correction to maintain accuracy, and employing an adaptive system that dynamically optimizes GPU workloads.

Introduction

Social media platforms face a daunting task in verifying the provenance of AI-generated content at scale. Watermarking—embedding a hidden digital signature into AI outputs—is a primary solution, but existing methods like Stable Signature are computationally expensive. For instance, processing one million images can take up to 30 minutes, which is insufficient for the real-time requirements of platforms like YouTube or TikTok.

A potential solution is tiling: analyzing a small $64 \times 64$ patch instead of a full $1024 \times 1024$ image. However, naive tiling typically leads to a significant degradation in detection accuracy and robustness. QRMark addresses this efficiency-accuracy trade-off by making tile-based detection reliable enough for production environments through algorithmic and system-level innovations.

Methodology

QRMark integrates algorithmic error correction with hardware-level optimizations to create a robust detection pipeline.

1. Reed-Solomon (RS) Integration

Borrowing from the mathematics that allow QR codes to remain readable even when damaged, QRMark encodes watermarks using Reed-Solomon error correction. This allows the system to reconstruct the original digital signature even if the data extracted from individual tiles is partially corrupted or noisy.

2. Tile-Specific Training and System Architecture

Unlike standard models trained on full-resolution images, QRMark’s extractor is specifically pre-trained and fine-tuned to recover messages from localized tiles. The system architecture, shown in Figure 1, consists of a trainer for the LDM decoder and a detection optimizer that manages the online workflow.

Overview of the QRMark system architecture, illustrating the offline training and online detection workflows.

3. System-Level Optimizations

  • Adaptive Multi-Channel Horizontal Fusion: The system dynamically assigns GPU streams to bottleneck stages (preprocessing, tiling, or decoding) based on real-time profiling.
  • Workload Interleaving: To maximize hardware utilization, QRMark overlaps CPU-side data loading for future batches with the GPU-side computation of the current batch.
  • RS Caching: To mitigate the CPU-intensive nature of RS decoding, the system uses a lookup table to cache results for frequently recurring raw bitstreams.

Experimental Evaluation

The authors evaluated QRMark using the MS-COCO and PartiPrompts datasets, comparing it against baselines like Stable Signature and AquaLoRA. Metrics included bit accuracy, robustness under common image attacks (cropping, resizing, JPEG compression), and system throughput.

Key Results and Insights

QRMark demonstrates that system-level orchestration is as critical as model architecture for achieving real-world speedups. The core insight is that a perfect tile-level extractor is unnecessary if a strong error-correction layer is present; by accepting a higher raw error rate from smaller, faster kernels, the overall system throughput increases significantly.

  • Throughput and Latency: QRMark achieves an average 2.43× throughput increase over sequential baselines. As shown in Figure 2, the performance gains are particularly pronounced at larger batch sizes, where the latency of the baseline can be nearly 4× higher than QRMark.
  • Accuracy: The system maintains >99% bit accuracy for tile sizes of $32 \times 32$ and larger.
  • Robustness: QRMark remains resilient to adversarial manipulations, including cropping down to 10% of the original image area and heavy JPEG compression.
  • Scalability: The system scales efficiently across multiple GPUs, reaching a 4.1× speedup on 8 GPUs, representing a 30% improvement over standard data-parallel methods.
Performance comparison of QRMark with and without tiling across various batch sizes, showing significant speedup (a) and reduced latency (b).

Limitations and Future Work

While highly effective, QRMark's accuracy drops significantly when tiles are smaller than $32 \times 32$ (e.g., $16 \times 16$ tiles result in ~75% bit accuracy). Additionally, Reed-Solomon decoding remains a CPU-bound process; if the cache hit rate is low, the CPU can become a bottleneck. Future research may explore advanced ML-based tile-size predictors and extend the framework to video watermarking, where temporal redundancy could offer further efficiency gains.

Summary updated

Abstract

Efficient and reliable detection of generated images is critical for the responsible deployment of generative models. Existing approaches primarily focus on improving detection accuracy and robustness under various image transformations and adversarial manipulations, yet they largely overlook the efficiency challenges of watermark detection across large-scale image collections. To address this gap, we propose QRMark, an efficient and adaptive end-to-end method for detecting embedded image watermarks. The core idea of QRMark is to combine QR Code-inspired error correction with tailored tiling techniques to improve detection efficiency while preserving accuracy and robustness. At the algorithmic level, QRMark employs a Reed-Solomon error correction mechanism to mitigate the accuracy degradation introduced by tiling. At the system level, QRMark implements a resource-aware multi-channel horizontal fusion policy that adaptively assigns more streams to GPU-intensive stages of the detection pipeline. It further employs a tile-based workload interleaving strategy to overlap data-loading overhead with computation and schedules kernels across stages to maximize efficiency. End-to-end evaluations show that QRMark achieves an average 2.43x inference speedup over the sequential baseline.
BibTeX
@misc{zhong2025efficient,
  author = {Xinrui Zhong and Xinze Feng and Jingwei Zuo and Fanjiang Ye and Yi Mu and Junfeng Guo and Heng Huang and Myungjin Lee and Yuke Wang},
  title = {{An Efficient and Adaptive Watermark Detection System with Tile-based Error Correction}},
  year = {2025},
  eprint = {2509.02447},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.02447},
}
Figure