Unapologetically Distributed: A Call for Decentralized Document Analysis

Adrià Molina; Oriol Ramos Terrades; Josep Lladós

cs.CVSep 30, 2026

Unapologetically Distributed: A Call for Decentralized Document Analysis

Authors: Adrià Molina, Oriol Ramos Terrades, Josep Lladós

Organizations: Centre de Visió per Computador Universitat Autònoma de Barcelona Bellaterra, Catalonia · Computer Science Department Universitat Autònoma de Barcelona, Bellaterra, Catalonia

Abstract

Privacy has become an increasingly important concern in the Document Analysis community, to the extent that in many environments such as archives, governmental institutions, and local businesses, the adoption of automation is restricted by legal and policy constraints. While federated learning has often been regarded as a ``necessary evil'', implying an unavoidable performance trade-off in exchange for decentralization and privacy, many prior works overlook its potential to improve robustness to out-of-distribution data. In this paper, we present Unapologetically Distributed, the first comprehensive study evaluating distributed learning in Document Analysis along three key axes simultaneously: the tasks addressed, the architectures employed, and the fine-tuning strategies applied. Specifically, we demonstrate how various distributed training approaches enhance generalization capabilities across diverse tasks such as Table Recognition, handwriting recognition, and Word Spotting, particularly during transfer learning stages. Our results provide strong evidence that decentralization is not merely a constraint, but a valuable opportunity to improve model robustness and adaptability in real-world Document Analysis scenarios.

Figures & tables

Panel of Figure 1 — Figure 1 : We show that distributed regimes (orange area, proposed) consistently outperform centralized pretraining (blue area) across a wide range of architectures and downstream tasks. We perform an extensive evaluation on 27 well-established Word Spotting and recognition datasets.

Figure 2 : Knowledge distillation scheme: multiple pretrained OCRs are combined into a frozen teacher, which guides an RNN to map text to images in unseen alphabets or domains.

Panel of Figure 3 — Figure 3 : Distributed models trained on Latin alphabets are evaluated on unseen scripts by averaging backbones and fine-tuning only the language head.

Panel of Figure 4 — Figure 4 : Distributed learning of message passing layers. Standard models (red/blue) are combined via distributed message passing (green), aggregating edge parameters ( $\theta$ ) while keeping node features ( $h_{j}$ ) fixed.

Dataset Name	Tasks	Finetuning	Pretraining	Alphabet
IAM [ Marti and Bunke(2002) ]	WS	KD	✓	Latin
Esposalles [ Romero et al.(2013)Romero, Fornés, Serrano, Sánchez, Toselli, Frinken, Vidal, and Lladós ]	WS	KD	✓	Latin
George Washington [ Fischer et al.(2012)Fischer, Keller, Frinken, and Bunke ]	WS	KD	✓	Latin
Parzival [ Fischer et al.(2012)Fischer, Keller, Frinken, and Bunke ]	WS	KD	✓	Latin
CoCoText [ Veit et al.(2016)Veit, Matera, Neumann, Matas, and Belongie ]	WS	KD	✓	Latin
MLT19 (Latin) [ Nayef et al.(2019)Nayef, Patel, Busta, Chowdhury, Karatzas, Khlif, Matas, Pal, Burie, Liu, et al. ]	WS	KD	✓	Latin

Table 1 : Overview of the 27 evaluation datasets. For each dataset, we indicate whether it is in-pretraining or out-of-domain.

Figure 5 : Mean Average Precision (mAP) for Query-by-String Word Spotting across 22 datasets using GRU, LSTM, and Transformer. Distributed (squares) excels in out-of-domain tasks, especially with shared alphabets.

mAP
Mean mAP (C / D)	.313 / .363
Recall
R@1 (C/D)	.256/ .300
R@5 (C/D)	.388/ .451
R@10(C/D)	.447/ .516

Table 2 : Centralized (C) vs. distributed (D) pre-training: average recall across word spotting datasets, including ID, OOD-L, and OOD-U scenarios.

	Vatican	Borg	Copiale	Arabic	Chinese	Japanese	Korean	Bangla	Hindi	$\times\Delta$
From Baseline	.549	.382	.825	.131	.020	.116	.282	.260	.470	-
Centr. (base)	.465	.000	.794	.175	.0103	.0951	.214	.115	.384	0.69
Dist. (base)	.591	.505	.838	.422	.073	.197	.374	.462	.514	1.62
Dist. (random)	.480	.272	.930	.410	.010	.114	.271	.266	.282	1.04

Table 3 : Transfer learning accuracy across languages. Distributed fine-tuning from a shared pretrained model (

Z

) outperforms centralized fine-tuning and task arithmetic from scratch, especially on low-resource alphabets.

	Multi-Lingual						Multi-Cipher
Accuracy $\uparrow$	Arabic	Bangla	Chinese	Hindi	Japanese	Korean	Borg	Copiale	Vatican
Distributed	.472	.469	.127	.539	.252	.435	.573	.840	.566
Centralized	.402	.385	.076	.468	.183	.324	.573	.819	.524

Table 4: Multi-lingual and multi-cipher training results.

	M96		RLV		ICDAR		CON-Anonym
	Node Acc	Edge Accuracy	Node Acc	Edge Accuracy	Node Acc	Edge Accuracy	Node Acc	Edge Accuracy
Centr. (zero-shot)	87.50%	85.70%	76.70%	72.90%	68.10%	78.00%	75.50%	70.10%
Dist. (zero-shot)	88.00%	87.00%	78.10%	78.00%	79.00%	84.70%	73.00%	77.20%
Centr. (partial)	87.60%	85.20%	76.90%	69.40%	82.60%	81.70%	89.30%	73.80%
Dist. (partial)	88.00%	86.50%	79.70%	72.30%	83.40%	78.70%	89.10%	75.40%
Centr. (E2E)	86.10%	69.90%	85.30%	81.50%	92.00%	95.50%	97.80%	90.60%

Table 5 : Table Recognition results: Node (cell) and Edge (link) accuracy. Models are trained on three datasets and fine-tuned on the fourth.

Regime	Low-Data	ZS	OOD	Param.-Efficient	E2E	ID
Distributed	Strong	Strong	Strong	Strong	Moderate	Low
Centralized	Moderate	Low	Moderate	Moderate	Strong	Strong

Table 6 : Summary of the conclusions and general performance trends.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

	M96		RLV		ICDAR		CON
	Node	Edge	Node	Edge	Node	Edge	Node	Edge
ZS – r2	88.6	87.3	76.9	79.6	80.6	76.9	72.4	75.7
ZS – r3	88.1	86.7	75.0	78.6	83.5	82.2	74.5	76.7
ZS – r4	88.2	86.8	75.2	79.0	82.2	78.7	72.5	79.4
ZS – r5	87.6	86.7	74.3	78.2	82.3	81.8	68.7	78.9
Part – r2	88.6	87.0	80.9	72.3	83.5	81.4	88.5	76.4

Appendix

Table 7 : Multi-round FedAvg on Table Recognition, rounds 2–5 (round 1 and centralized in submitted Tab. 5). No consistent gain beyond round 1.

Explore similar work

Federated Learning of AnDE Classifiers
Sep 23, 2026Pablo Torrijos, Juan C. Alfaro, José A. Gámez +1Federated Learning Secure Aggregation
Departamento de Sistemas Informáticos. Universidad de Castilla-La Mancha, 02071 Albacete, Spain. · Instituto de Investigación en Informática de Albacete. Universidad de Castilla-La Mancha, 02071 Albacete, Spain.
Privacy-Preserving and Verifiable Approximate Distributed Coded Computing
Jul 2, 2026Xavier Martínez-Luaña, Alba Gude-Santos, Manuel Fernández-Veiga +1Privacy-Preserving Machine Learning Federated Learning
Galician Research and Development Center in Advanced Telecommunications (GRADIANT) Estrada do Vilar, 56-58, Vigo, 36214, Spain · atlanTTic, Information & Computing Lab (ICLAB), Telecommunication Engineering School, Universidade de Vigo Vigo, 36310, Spain
Byzantine-Robust Aggregation for Securing Decentralized Federated Learning
Sep 26, 2024Diego Cajaraville-Aboy, Ana Fernández-Vilas, Rebeca P. Díaz-Redondo +1Federated Learning Byzantine Attacks
atlanTTic Research Center – ICLAB – Universidade de Vigo, Vigo, 36310, Spain

Sep 23, 2026, cs.LGJ/K move · Enter open · S save

Federated Learning of AnDE Classifiers

Pablo Torrijos, Juan C. Alfaro, José A. Gámez, José M. Puerta

Departamento de Sistemas Informáticos. Universidad de Castilla-La Mancha, 02071 Albacete, Spain. · Instituto de Investigación en Informática de Albacete. Universidad de Castilla-La Mancha, 02071 Albacete, Spain.

Federated Learning Secure Aggregation Discriminative

Open paper AskPDF arXiv

This work presents a federated framework for training Averaged

n

-Dependence Estimators (AnDE) in distributed environments. The proposed method focuses on the discriminative setting, where model weights are learned locally and aggregated globally, supporting any dependency order

n

. This design allows federated training without transmitting semantically meaningful parameters, improving privacy. Additionally, generative AnDE models are federated to provide a comparative baseline, with optional differential privacy applied to the aggregation of probability tables. Experiments on 12 discrete datasets show that discriminative models with

n \geq 1

consistently outperform federated Naive Bayes (NB,

n=0

), and that privacy-preserving aggregation is effective with limited accuracy loss. These results establish federated AnDE as a viable and privacy-preserving framework, showing that probabilistic models remain applicable in modern federated learning settings.

Abstract

Explore similar work

Federated Learning of AnDE Classifiers

Privacy-Preserving and Verifiable Approximate Distributed Coded Computing

Byzantine-Robust Aggregation for Securing Decentralized Federated Learning