UNT REU: Making Generative AI Responsible
Example Research Projects
The following examples illustrate research projects available for the upcoming summer program. Final project assignments are determined based on student interests and faculty availability.
Algorithmic Generation of Images with User-Defined Demographic Distributions
Description: This project develops algorithms using Large Language Models (LLMs) and text-to-image generators to provide precise control over the demographic characteristics represented in generated images. By refining text prompts to specify desired demographic attributes and their frequencies, the system aims to produce image outputs that accurately reflect specified statistical distributions. This approach incorporates model bias assessment and balanced training techniques to ensure the LLMs guide image generation toward achieving the intended demographic representation based on user input.
Assessing Performance Integrity in Multimodal Generative AI Systems
Description: This project investigates potential performance anomalies in advanced multimodal AI systems (such as ALIGN, BLIP-2, ImageBind, and Mini-GPT4), which integrate data like text and images. Combining different data types can inadvertently amplify inaccuracies or distortions present in the training data, potentially leading to generated outputs that do not align with objective representation or intended use. To address this technical challenge, the project proposes developing a framework using adversarial test samples generated by StyleGAN. This framework will be used to rigorously measure the representational accuracy and identify potential distortions within these multimodal models, aiming to enhance their reliability and operational integrity.
Optimizing AI Gesture-to-Speech Performance for Users with Motor Impairments
Description: This project aims to improve the performance of AI-based gesture recognition systems specifically for individuals with significant motor and speech limitations, as standard systems often prove inadequate for this user group. By employing advanced continual learning architectures and innovative user interfaces, the research seeks to create a more effective wearable gesture-to-speech communication technology. The work includes analyzing performance differences when models are trained on data from impaired versus non-impaired subjects to optimize system design and understand resource tradeoffs. The ultimate goal is to enhance communication speed and functional capability for users with these specific physical challenges by developing a more robust and user-friendly gesture vocabulary, validated through clinical application prototypes.
Hardware-Enabled Data Protection for GPUs in Generative AI Systems
Description: This project focuses on enhancing the security of sensitive data and Generative AI models through hardware-level techniques implemented within Graphics Processing Units (GPUs). Recognizing the critical role of GPUs in AI computation, the research investigates methods to integrate data protection directly into GPU hardware to safeguard information during processing and transmission, thereby supporting secure collaboration. The work involves developing distinct encryption strategies: robust cryptographic methods with message authentication for data moving between the CPU and GPU over PCIe buses, where stronger security can tolerate some performance overhead, and lightweight encryption techniques using efficient arithmetic operations for the high-speed, delay-sensitive on-chip network within the GPU itself.
Enhancing Security and Isolation in Multi-GPU Systems for Large AI Models
Description: This project addresses security vulnerabilities specific to multi-GPU systems used for large-scale Generative AI applications, where processes running on one GPU can potentially interfere with or access data from processes on other GPUs within the shared hardware. To mitigate risks such as cross-GPU cache contention or data leakage, the research focuses on developing techniques for enhanced data protection and process isolation in these environments. The approach involves adapting and extending existing security methods by investigating static and dynamic partitioning of shared resources, like memory, among different concurrent users or tasks. Performance, overhead, and scalability of these proposed partitioning solutions will be evaluated to provide effective security measures for large GPU clusters supporting demanding AI model training, enabling better security-performance tradeoffs.
AI-Driven Analysis for Consistent Application of Law in Judicial Decisions
Description: This project investigates methods to enhance the objectivity and consistency of judicial decisions by identifying and potentially neutralizing the influence of factors unrelated to legal merits. Utilizing Generative AI, the research aims to detect patterns where legally irrelevant variables may correlate with variations in judicial outcomes. An annotation scheme will be developed to categorize potential sources of decision anomalies, and large datasets will be constructed for analysis using machine learning. Generative AI will simulate alternative scenarios, adjusting variables to observe their potential impact, thereby identifying patterns inconsistent with purely fact-and-law-based reasoning. The project explores developing AI tools, potentially including alert systems for decision-makers or methods for generating decision drafts based strictly on relevant legal factors, employing techniques like prompt engineering and specialized algorithms to support consistent application of the law.
Prompt Engineering Methods for Assessing AI Detection of Code Accessibility Barriers
Description: While Generative AI models can rapidly generate code, concerns exist regarding the quality and adherence to technical standards in the output, including established accessibility guidelines like WCAG 2.2. Code failing to meet these technical standards can create functional barriers, impairing usability for individuals with various physical or cognitive limitations. This project evaluates the reliability and accuracy of specific Generative AI models (GPT, LLaMA, Gemini) in detecting code that violates WCAG 2.2 accessibility standards. Using WCAG 2.2 as a technical benchmark, the research involves developing systematic prompt engineering methods to assess the models' performance in identifying non-compliant code, aiming to ensure AI-assisted software development meets necessary requirements for quality and broad functional usability.