Fed-IGTD: A federated learning-based framework for tabular-to-image encoding

David Chushig-Muzo, Eva Milara, Francisco J. Lara Abelenda, Luis Carlos Estrada-Petrocelli, Diego H. Peluffo-OrdóñezarXiv 2026 · arXiv

TL;DR

Fed-IGTD is a new framework that brings tabular-to-image encoding into the privacy-preserving world of Federated Learning (FL). It allows decentralized clients to collaboratively build a shared image representation of their tabular data and train a Convolutional Neural Network (CNN) without ever sharing raw records, while offering formal differential privacy guarantees.

What Problem Does This Paper Address?

Tabular data—the kind found in spreadsheets and databases—is the backbone of industries like healthcare and finance. While Deep Learning, particularly CNNs, has revolutionized image processing, applying it to tabular data is tricky because tables lack the inherent spatial structure (like neighboring pixels) that CNNs exploit.

"Tabular-to-image" methods, such as IGTD (Image Generator for Tabular Data), solve this by mapping similar features to nearby pixels on a synthetic image grid. However, these methods traditionally require a "global view" of the data to calculate feature correlations, meaning all data must be centralized. In high-stakes fields, regulatory and privacy concerns often make centralizing raw data a non-starter. The authors address this by creating a federated version of IGTD that maintains privacy while capturing global data relationships.

Methods

The authors introduce Fed-IGTD, a horizontal FL framework. In this setup, multiple clients have the same types of data (features) but different individual records. The process happens in two main stages:

  1. Mapping Construction (One-Shot): Instead of sending raw data, each client computes local "sufficient statistics" (sums and Gram matrices). Because these statistics are additive, the server can sum them to perfectly reconstruct a global correlation matrix. The server then runs the IGTD optimization to create a feature-to-pixel mapping, which it broadcasts back to all clients.
  2. Federated Training (Iterative): Clients use the shared mapping to convert their local tables into images once. They then participate in standard federated training (like FedAvg) to train a CNN on these images.

To ensure formal privacy, the authors implemented two Differential Privacy (DP) mechanisms:

  • Gram-matrix DP: Adds noise to the initial statistics to protect the privacy of the image layout.
  • DP-FedAvg: Adds noise to the model updates during the CNN training phase to protect the final model from leaking information about specific clients.
Fig. 1: Fed-IGTD workflow.

Datasets

The framework was evaluated across five diverse domains using publicly available datasets:

  • Healthcare: Pima Indians Diabetes
  • Finance: German Credit
  • Chemistry/Agriculture: Wine Quality (White)
  • Signal Processing: Connectionist Bench (Sonar)
  • Physics/Radar: Ionosphere
TABLE 1: Datasets used in this study.

Key Insights and Ideas

  • Mathematical Elegance: The core "aha moment" is that the Pearson correlation matrix—the engine behind IGTD—can be reconstructed to numerical precision from simple additive statistics. This means the federated image layout is virtually identical to one created with centralized data.
  • Privacy is Not Free, but Some Parts are Cheaper: A major finding is that applying DP to the mapping step (the one-shot correlation reconstruction) has almost no impact on the model's final accuracy. However, applying DP to the training loop (DP-FedAvg) causes significant performance drops.
  • The "Macro-F1" Warning: The authors highlight that standard accuracy can be misleading in federated settings with skewed data (non-IID). Macro-F1 is a much more sensitive and honest metric for detecting when a model is failing on minority classes due to data heterogeneity.
  • Dataset Sensitivity: Smaller datasets (like Sonar) are much more volatile under privacy constraints and data skew, suggesting that "one size fits all" privacy budgets don't work in practice.

Key Results

  • Reconstruction Accuracy: The federated reconstruction of the correlation matrix was nearly perfect, with errors on the order of $10^{-15}$ (essentially floating-point rounding error).
  • Baseline Performance: Federated training without DP generally achieved 2% to 10% lower accuracy than the centralized equivalent, establishing a clear "federation tax."
  • Privacy Trade-offs: While Gram-matrix DP was "utility-neutral," DP-FedAvg significantly degraded performance at moderate privacy budgets ($\epsilon \approx 2-8$), particularly for the Ionosphere dataset, where Macro-F1 plummeted.
Fig. 3: Centralized and federated predictive performance under FedAvg with five clients (α = 1.0, without differ- ential privacy). (a) Test accuracy; and (b) Test macro-F1.

Limitations and Future Work

  • Small Cohort Size: The study used a fixed cohort of 5 clients. The authors note that DP-FedAvg typically performs better with much larger groups of clients, which would be a natural next step for evaluation.
  • Horizontal Only: The current framework is limited to horizontal FL. Extending this to Vertical FL (where different clients hold different features for the same individuals) would be highly valuable for cross-industry collaborations (e.g., a bank and a retail store).
  • Model Complexity: The study used a relatively compact CNN. Future work could explore how more complex architectures or different tabular-to-image encoding methods (like TabSOM) interact with the federated environment.

Summary updated

Abstract

Tabular-to-image encoding methods, which transform tabular data into image-like representations, have shown promising results in predictive tasks. Despite their benefits, these methods follow a centralized paradigm, in which the data are collected and models are trained on a central device. This raises privacy and security concerns, particularly the risk of data leakage, unauthorized data collection and access to users' private information. Federated Learning (FL) provides a decentralized framework in which multiple clients jointly optimize a shared model without directly sharing training data. We introduce Fed-IGTD, a horizontal FL-based approach in which clients share a common feature space and upload only local statistics, from which the server reconstructs the global correlation matrix and obtains a feature-to-pixel mapping to train a convolutional neural network via Federated Averaging (FedAvg). We studied two differential privacy (DP) mechanisms: (i) a one-shot Gaussian mechanism protecting the correlation-aggregation step; and (ii) a client-level DP-FedAvg variant protecting the training with privacy loss composed via a Renyi-DP accountant. We evaluated Fed-IGTD on five datasets under a systematic sweep of non-IID severity and privacy budgets. Privacy protection applied to the mapping-construction step has little impact on predictive performance. By contrast, DP-FedAvg substantially degrades utility at moderate privacy budgets with a small client cohort (K=5). Ablation study indicates that this degradation is driven mainly by the injected noise rather than by uniform averaging or update clipping. Macro-F1 exhibits sensitivity to performance degradation than accuracy under non-IID heterogeneity and class imbalance.
BibTeX
@misc{chushigmuzo2026fed,
  author = {David Chushig-Muzo and Eva Milara and Francisco J. Lara Abelenda and Luis Carlos Estrada-Petrocelli and Diego H. Peluffo-Ordóñez},
  title = {{Fed-IGTD: A federated learning-based framework for tabular-to-image encoding}},
  year = {2026},
  eprint = {2610.39008},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.39008},
}
Figure