Graduation Semester and Year
Summer 2026
Language
English
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Computer Science
Department
Computer Science and Engineering
First Advisor
Dr. Shirin Nilizadeh
Second Advisor
Dr. Chengkai Li
Third Advisor
Dr. Jacob Luber
Fourth Advisor
Dr. VP Nguyen
Abstract
The proliferation of artificial intelligence (AI) across critical domains, including news summarization, privacy-policy analysis, and medical decision support, has raised growing concerns about the security and robustness of these systems against adversarial manipulation. This dissertation investigates adversarial robustness in generative AI by addressing three key research goals: (1) characterizing adversarial vulnerabilities across generative models, (2) developing systematic defenses to improve the robustness of generative models, and (3) designing deployment-time safeguards for securing LLM interactions.
Towards the first goal, we characterize adversarial vulnerabilities across text-based and multimodal systems. In abstractive text summarization, we show that inference-time perturbations can exploit lead bias to exclude targeted information from generated summaries. We further introduce influence-based data poisoning attacks that alter intended model behavior and induce collateral changes in generation style, including a shift toward more extractive summaries. We then examine automated privacy-policy analyzers under a reversed threat model in which organizations strategically modify their own policies to mislead automated analysis. In clinical pathology vision-language models, we demonstrate the first successful adversarial attack against PLIP, achieving 100% misclassification with visually imperceptible perturbations and redirecting model attention away from clinically relevant regions.
Building on our understanding of these vulnerabilities, we address the second goal by developing systematic defenses to improve the robustness of generative models, specifically for text summarization. For inference-time vulnerabilities, we introduce Bias-Amplified Task Vector Negation, which reduces lead bias by an average of 31% across 11 models and 6 datasets while preserving summarization quality and improving robustness to both lead-position and adaptive non-lead attacks. For training-time poisoning, we introduce a black-box detection method based on adversarial sensitivity and a white-box mitigation pipeline that combines influence-based sample localization, behavioral filtering, and gradient-ascent unlearning. The mitigation pipeline achieves an 85.3% behavioral recovery rate without requiring retraining from scratch.
Towards the third goal, we design deployment-time safeguards for securing LLM interactions through a role-specialized multi-agent framework for prompt-response safety verification. The framework jointly evaluates user intent and the harm of generated responses through separate task-analysis, safety-analysis, and adjudication agents. Across multiple safety datasets and language-model backbones, it outperforms single-agent and chain-of-thought baselines and generalizes across diverse malicious behaviors, including jailbreaks, prompt injection, phishing, unsafe code generation, and harmful-content requests.
Overall, this dissertation demonstrates that adversarial vulnerabilities in generative AI can arise across applications, modalities, and different stages of model use. By characterizing these vulnerabilities, developing systematic defenses and designing deployment-time safeguards, this work provides a foundation for improving the robustness, security, and trustworthiness of generative AI systems.
Keywords
Generative AI, Adversarial Robustness, AI Security, Adversarial Attacks, Data Poisoning, Large Language Models, Text Summarization, Multimodal AI, Machine Unlearning, AI Safety
Disciplines
Artificial Intelligence and Robotics | Computer Sciences | Cybersecurity | Data Science | Information Security
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Recommended Citation
Thota, Poojitha, "SYSTEMATIC APPROACHES TO CHARACTERIZING VULNERABILITIES AND ENHANCING ROBUSTNESS OF TEXT AND VISION-LANGUAGE MODELS" (2026). Computer Science and Engineering Dissertations. 24.
https://mavmatrix.uta.edu/cse_dissertations2/24
Included in
Artificial Intelligence and Robotics Commons, Cybersecurity Commons, Data Science Commons, Information Security Commons