
Modern deep learning models have driven significant performance breakthroughs in document key information extraction (KIE) systems. At the same time, however, their data-driven nature exposes these systems to a new class of privacy vulnerabilities. Since KIE systems directly interact with sensitive information in documents, they are potential targets for privacy attacks. Despite this, research on privacy-preserving multimodal document KIE remains scarce. To address this gap, we conduct the first systematic investigation into privacy-preserving training for multimodal document KIE models, comparing differential privacy (DP), federated learning (FL), and differentially private federated learning (DP-FL) approaches. Through extensive experimentation across six benchmark datasets, we demonstrate that while DP training typically causes substantial performance degradation, this can be effectively addressed through: (1) the use of pretrained document foundation models; (2) DP-suited hyperparameter tuning; and (3) task-specific pretraining that aligns target entity types with the private data distribution. In particular, this combination allows up to 77–93 ε∈ 8, 20). Notably, task-specific pretraining alone contributes major performance gains of 2–6.5
Extracting structured data from unstructured document images containing multiple form records remains a challenging task. Existing approaches rely on predefined templates or extensive labeled training data, limiting their applicability to diverse real-world scenarios. This paper proposes an end-to-end framework that leverages Large Language Models (LLMs) for automatic form extraction. We develop a three-stage pipeline comprising: (1) OCR-based document preprocessing that converts heterogeneous source documents into a normalized markdown representation while preserving table structures; (2) a Multi-Strategy Adaptive Segmentation (MSAS) algorithm that identifies record boundaries by combining equal-interval structural boundary snapping, token-frequency analysis with coefficient of variation (CV) filtering, and iterative boundary optimization guided by primary keys and values; and (3) LLM-based extraction and semantic mapping that performs zero-shot schema induction, field value extraction, and alignment to target database schemas. Experimental evaluation across 50 diverse application scenarios (including enrollment registration, job applications, hotel check-in, visa processing, etc.) with 2,488 form records demonstrates an overall cell-level extraction accuracy of 78.48
This paper presents a benchmark and a systematic empirical study for document layout analysis in administrative documents, motivated by the scarcity of region-level annotated datasets and the limited understanding of failure modes under real-world layout variability. We introduce a publicly available dataset comprising 2,720 page images from 1,513 PDF files across seven administrative document categories, annotated at the semantic region level to support fine-grained layout analysis. The dataset will be released upon acceptance. Information region extraction is formulated as an object detection problem, enabling a controlled comparison of representative detection paradigms, including an anchor-free single-stage detector (YOLOv12), a two-stage anchor-based detector (Faster R-CNN), and a one-stage anchor-based detector (RetinaNet). Experimental results show that YOLOv12 achieves the strongest overall detection performance while requiring substantially shorter training time than the evaluated baselines, whereas inference-time ensemble strategies provide only marginal improvements. A targeted error analysis reveals that detection failures are highly concentrated in structurally complex regions, particularly the documentTitle field, where multi-line layouts and semantic variability challenge conventional detection and Optical Character Recognition pipelines. Although the benchmark is constructed from Vietnamese administrative documents, the identified failure patterns are likely related to layout structure and typographic characteristics rather than language itself. Their applicability to other administrative domains should be validated in future studies.
Text line segmentation in Sanskrit manuscripts remains a difficult problem due to physical degradation, irregular layouts, and script-specific features such as the Shirorekha (the continuous headline) and dense conjuncts. We present ShirorekhaNet, a residual encoder-decoder network that performs pixel-accurate line-level segmentation. The architecture accepts a four-channel input-RGB combined with a binarised image-and integrates multi-scale Inception (MSI) modules with channel attention, residual connections with learnable gating, and a hybrid training curriculum that combines cosine learning-rate decay with a Proportional-Integral-Derivative (PID) controller. The model is trained on an augmented set of 8585 images derived from 85 annotated folios. On the 25 original, unaugmented test images, ShirorekhaNet (PID + cosine) achieves an IoU of 0.9075, a Dice of 0.9513, a Shirorekha accuracy of 95.68
Subgraph-level reasoning is increasingly important for analyzing localized structures within large heterogeneous graphs. However, existing benchmarks are mostly node-centric, edge-centric, or graph-level, with limited support for subgraph-level supervision, query semantics, or spotting objectives. Moreover, analyzing large heterogeneous networks can be computationally costly, especially in settings with limited resources. This work introduces the HSB-GDL benchmark dataset to address these gaps by supporting subgraph-level learning and detection in academic heterogeneous graphs with configurable complexity and evaluation depth, without requiring advanced computational resources. The benchmark is based on the OGBN-MAG dataset and includes a diverse set of 59,585 subgraphs spanning both homogeneous and heterogeneous structures. These include ten types of real-world patterns and six types of controlled synthetic patterns. The subgraphs are labeled, annotated, and organized for simple integration. Classification is evaluated under complete-standard and strict node-disjoint protocols. Query spotting is assessed through global retrieval over the benchmark catalog and conditional hard-negative ranking. The evaluation includes feature controls and GIN, RGCN, and TransformerConv baselines. We provide a detailed empirical evaluation that includes structural-bias diagnostics and stronger graph-learning baselines. The diagnostic results show that HSB-GDL contains both easily separable calibration patterns and more ambiguous real heterogeneous subgraphs. In particular, node/edge-type structure is a strong predictive signal for several classes. Under the complete-standard protocol, GIN achieves a Macro-F1 of 90.95± 0.53% , compared with 60.77± 0.96% for the node-type-only control. Under strict node disjointness, GIN obtains 90.45± 0.34% Macro-F1. For global query retrieval, designated-target Recall@100 is 61.25% , while valid-occurrence Recall@100 is bounded at 86.13– 97.88% . In conditional hard-negative ranking, the combined ranker achieves MRR 0.8710. These findings indicate that HSB-GDL serves as a subgraph-level benchmark with diagnostic capability, enabling researchers to distinguish shortcut-driven classification from topology-aware and relation-aware subgraph reasoning. The benchmark is open source, CPU-accessible, and extensible, making it useful for reproducible research on heterogeneous graph learning, subgraph-level pattern recognition, and query-oriented graph retrieval for research and educational purposes.