Research work on Document Unlearning

This research advances document unlearning and secure document image understanding by addressing the critical challenge of selectively removing sensitive, proprietary, or confidential information from trained document AI models without requiring full retraining.

Our work focuses on developing novel model architectures and diverse unlearning approaches tailored to complex document layouts and text structures. We explore rigorous evaluation metrics to balance unlearning efficacy with preserved model utility, tackling core challenges such as catastrophic forgetting, spatial-textual dependency, and secure compliance in multi-modal document environments.

Standard deep learning vision-language models process documents by fusing spatial bounding boxes with textual embeddings. In document unlearning, our model architectures are explicitly designed to modularize information storage:

• Layout-Aware Decoupled Encoders: Separating visual layout representation layers from semantic text representation layers to target specific structural memory traces.
• Modular Sub-Network Partitioning: Designing compartmented neural paths so that confidential zones (such as patient records or legal contracts) can be surgically modified without disrupting general document parsing capabilities.

We investigate multiple paradigms for forgetting sensitive data fragments efficiently:

• Exact vs. Approximate Unlearning: Designing closed-form parameter updates versus optimized gradient-based scrubbing to eradicate specific document traces.
• Selective Gradient Ascent: Incrementally maximizing loss on targeted confidential document classes while minimizing loss on retention corpora.
• Influence Function Scrubbing: Estimating the impact of individual training documents on model parameters to pinpoint and reverse specific memorization paths.

Measuring the success of unlearning requires balancing two conflicting objectives: absolute privacy removal versus general model competence.

• Unlearning Efficacy: Evaluated using membership inference attacks (MIA), confidence score drops on forgotten samples, and text reconstruction failure rates.
• Utility Preservation: Ensuring accuracy on downstream document understanding tasks (e.g., QA, OCR, and layout parsing) remains uncompromised.
• Core Challenges: Tackling catastrophic forgetting, managing spatial-textual dependencies where text and layout context overlap, and guaranteeing legal compliance under data protection standards.