INVARIANT FEATURE LEARNING IN AI MODELS: EXPLORING REPRESENTATION SPACES FOR ROBUST WATERMARKING AND BEYOND

Author

• Agnibh Dasgupta, University of Nebraska at OmahaFollow

Date of Award

5-7-2026

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Dr. Xin Zhong

Abstract

This dissertation investigates invariant feature learning in modern AI models, with a focus on how representation spaces can be trained or analyzed to preserve semantic identity under perturbation while enabling robust watermarking. The central argument is that grounding invariant representations in semantic content produces more reliable feature spaces than those in which invariance arises implicitly as a by-product of task-driven learning, and that watermarking provides a principled framework for measuring this invariance across modalities.

In the visual domain, two contributions are presented. The first directly trains an invariant feature space for watermarking, establishing that a directly trained invariant domain improves robustness over borrowed pretrained representations. The second introduces a learned adversarial auto-augmentor and cross-modal text anchoring, grounding the invariant space in semantic content and achieving state-of-the-art robustness under real-world camera distortions. In the language domain, two contributions are presented. The first demonstrates that prompt-guided watermarking signals persist across LLM architectures and model transformations without access to model internals. The second, an ongoing investigation, examines whether transformer-based LLMs contain geometrically structured invariant subspaces in their deeper layers that are stable under paraphrasing and exploitable for robust downstream applications.

Together, these contributions advance a representation-space perspective in which semantic grounding is the organizing principle for invariance across both vision and language.

Comments

The author holds the copyright to this work, any reuse or permissions must be obtained from them directly. 

PDF passed Adobe's accessibility checker prior to upload.

Recommended Citation

Dasgupta, Agnibh, "INVARIANT FEATURE LEARNING IN AI MODELS: EXPLORING REPRESENTATION SPACES FOR ROBUST WATERMARKING AND BEYOND" (2026). Computer Science Theses, Dissertations, and Student Creative Activity. 7.
https://digitalcommons.unomaha.edu/compscistudent/7