Research map / Adversarial robustness and security / Attacks on text and graphs

RoDLA: Benchmarking the Robustness of Document Layout Analysis Models

CVPR 2024 · Adversarial robustness and security › Attacks on text and graphs · approach: Biologically inspired robust vision · 14 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces RoDLA, a robustness benchmark for document layout analysis with 12 document perturbations at 3 severity levels across PubLayNet-P, DocLayNet-P and M6Doc-P, plus the mPE and mRD metrics and a robustness-enhanced DINO-style model with channel attention and average pooling.

Problem

Document layout analysis models are evaluated on clean documents, and their robustness to realistic document perturbations (scanning, photographing artifacts) is underexplored; existing degradation metrics depend on a chosen baseline model.

Approach

Build a hierarchical taxonomy of 12 realistic document perturbations in 5 groups with 3 severity levels, propose mPE (combining IQA scores and model degradation) and mRD to measure perturbation effect and model robustness independent of a single baseline, and propose RoDLA, which adds channel attention and dilated average pooling to a DINO-based encoder to suppress attention on perturbed tokens.

Key results

• RoDLA achieves mRD of 116.0 on PubLayNet-P with 70.0% P-Avg mAP and 96.0% clean mAP, versus 120.7 for the next best DINO-InternImage in mRD.
• On DocLayNet-P, RoDLA reaches 80.5% clean mAP, 65.6% P-Avg and mRD 135.7, versus SwinDocSegmenter at mRD 282.7.
• On M6Doc-P, RoDLA obtains 70.0% clean mAP, 61.7% P-Avg and mRD 147.6, the lowest among compared methods.

Measured effects reported in the paper

Surprising findings

• SwinDocSegmenter drops from 93.7% to 37.3% and 29.2% mAP under blur, showing large fragility of some DLA models.
• The multimodal LayoutLMv3 does not show better robustness than the unimodal DiT on PubLayNet-P.

Limitations

• Benchmark covers only vision-based single-modal DLA models in its evaluation scope
• No human-in-the-loop testing of robustness
• Degradation component of mRD still relies on a baseline model (Faster R-CNN) for mPE

Experimental setup

Datasets: PubLayNet DocLayNet M6Doc ImageNet

Benchmarks: PubLayNet-P DocLayNet-P M6Doc-P PubLayNet DocLayNet M6Doc

Metrics: mAP P-Avg mRD mPE Degradation

Baselines: Faster R-CNN Mask R-CNN Cascade R-CNN DocSegTr SwinDocSegmenter DiT LayoutLMv3 Co-DINO DINO LayoutParser

Builds on: DINO InternImage ImageNet-C Token-aware Average Pooling

Closest papers by meaning

• M2Doc: A Multi-Modal Fusion Approach for Document Layout Analysis (AAAI 2024 · similarity 0.83)
• Graph-based Document Structure Analysis (ICLR 2025 · similarity 0.83)
• MMLONGBENCH-DOC: Benchmarking Long-context Document Understanding with Visualizations (NeurIPS 2024 · similarity 0.82)
• Towards Real-World Document Parsing via Realistic Scene Synthesis and Document-Aware Training (CVPR 2026 · similarity 0.82)
• SDEval: Safety Dynamic Evaluation for Multimodal Large Language Models (AAAI 2026 · similarity 0.82)
• RoDA: Robust Domain Alignment for Cross-Domain Retrieval Against Label Noise (AAAI 2025 · similarity 0.81)