Invariant Representation Learning in LLMs for Model Attribution
Under review
Layer-wise analysis framework for identifying paraphrase-stable latent representations in LLMs. Supports semantic clustering and model attribution tasks.
Agnibh Dasgupta
AI Researcher · PhD (Artificial Intelligence)
Representation Learning · Watermarking · Multimodality
Building robust, invariant representations for AI models
I study how models encode semantic meaning that remains stable under perturbations. My work opertionalizes these representations for watermarking, , and identifying invariant latent features in LLMs for attribution and forensics.
At a glance
- PhD candidate at University of Nebraska - Omaha. Focus: Artificial Intelligence
- Research interests:
- Representation learning
- Watermarking (vision & language)
- Multimodality
- Explainability
- Other Interests:
- LLM forensics
- Reinforcement learning
- Applied AI
About
I'm a doctoral researcher in Information Science & Technology at the University of Nebraska Omaha. My dissertation centers on invariant representation learning and its applications to robust image watermarking and LLM robustness.
Broadly, I design systems that remain stable under content-preserving transformations: geometric/photometric augmentations for images and lexical/structural paraphrases for text. I care about what an AI model knows versus how it encodes it.
Selected Research
TIACam: Text-Anchored Invariant Feature Learning with Auto-Augmentation for Camera-Robust Zero-Watermarking
Conference on Computer Vision and Pattern Recognition (in press)
TIACam is a text-anchored invariant feature learning framework for camera-robust zero-watermarking that embeds messages in a distortion-invariant feature space. Using a learnable auto-augmentor and cross-modal adversarial training, it achieves state-of-the-art watermark recovery under synthetic and real camera captures.
Watermarking Language Models through Language Models
IEEE Transactions on Artificial Intelligence 2025
Prompt-based LLM watermarking framework that embeds detectable signals in model responses without modifying weights or data. Evaluated watermark generation and detection using instruction-tuned LLMs. The figure above is an overview of our prompting strategy.
Robust Image Watermarking via Cross-Attention & Invariant Domain Learning
International Conf. on Computational Science & Computational Intelligence 2023
Watermark embedding and extraction method resilient to geometric and photometric attacks. Utilizes ViT-based cross-attention to align invariant domain features for robust watermark decoding. The figure above shows an overview of our proposed franework.
Publications
-
Watermarking Language Models through Language Models — IEEE Transactions on Artificial Intelligence, 2025
Authors: A. Dasgupta, A. Tanvir, X. Zhong
IEEE Xplore -
Enhanced Image Watermarking Through Cross-Attention and Noise-Invariant Domain Learning — Imaging Science: Computer Vision, Image and Signal Processing Pattern Recognition, 2024
Authors: A. Dasgupta, X. Zhong
IEEE Xplore -
Leveraging Artificial Intelligence (AI) to Enhance CS Instruction — Frontiers in Education, 2024
Authors: H. Yoon, X. Zhong, A. Dasgupta, G. Nugent, G. Trainin
IEEE Xplore -
Robust Image Watermarking based on Cross-Attention and Invariant Domain Learning — International Conf. on Computational Science & Computational Intelligence, 2023
Authors: A. Dasgupta, X. Zhong
IEEE Xplore -
Perspective Transformation Layer — International Conf. on Computational Science & Computational Intelligence, 2022
Authors: N. Khatri; A. Dasgupta; Y. Shen; X Zhong; F. Y. Shih
IEEE Xplore -
Deep Q Learning Applied to Stock Trading — Master’s Thesis, Department of Computer Science, Utah State University, 2020
Author: A. Dasgupta (Advisor: N. Flann)
DigitalCommons@USU
Links to papers in press will be updated as they become available.
