Research projects and publications

DocDjinn generates controllable synthetic documents by combining vision-language models with handwriting diffusion.

DocVCE generates diffusion-based visual counterfactual explanations for document image classification, revealing minimal edits that change a model's decision.

DP-DocLDM generates synthetic document images under differential privacy guarantees using latent diffusion models.

A robust and interpretable deep neural network for document image classification that couples strong accuracy with built-in attention-based explanations.

Examines the trade-off between privacy and explainability, measuring how differential privacy and federated learning affect attribution-based explanation methods.

A comprehensive benchmark for privacy-preserving document image classification, evaluating differentially private training across models and datasets.

A case study questioning the real-world reliability of high-performing document image classification models under realistic conditions.

StylusAI adapts handwriting style for robust German handwritten text generation using generative models.

Uses latent diffusion models for guided generation of realistic document tables.

DocXplain is a model-agnostic explainability method that produces feature-importance explanations for document image classifiers.

Investigates active learning strategies to reduce annotation cost for document image classification, analyzing when and how much labeled data can be saved.

ColDBin applies cold diffusion to document image binarization, restoring degraded and historical documents without relying on Gaussian noise.

PrIeD-KIE studies privacy-preserving key information extraction from documents using differential privacy and federated learning.

A large-scale study benchmarking the robustness of deep document image classifiers against real-world distortions and corruptions, introducing a corruption benchmark for document images.