Prosense - Defending Text Generation With Adversarial Feedback

Authors

• Anu Baluguri, The University of Southern Mississippi
• Vasudha Pasumarthy, The University of Southern Mississippi
• Yaswanth Raj Repakula, The University of Southern Mississippi
• Zhaoxian Zhou, The University of Southern Mississippi

Document Type

Conference Proceeding

Publication Date

1-1-2026

School

Computing Sciences and Computer Engineering

Abstract

Text generation models such as DeepSeek, Qwen, and ChatGPT have become indispensable tools for automation and data generation in the era of AI-powered creation. However, there are significant weaknesses that expose these models to attacks such as adversarial machine learning attacks. These attacks try to manipulate the input data to deceive the model into generating undesirable or unrelated results. In the proposed research, we examine how to strengthen the machine learning models’ defense against these kinds of attacks. In this paper, we propose optimizing the models on a combination of clean and adversarial data. Experimental results show that the implemented system is capable of producing meaningful and attack-resistant responses, even with manipulated input, making the system a more reliable and secure application.

Publication Title

Communications in Computer and Information Science

Volume

2720 CCIS

First Page

319

Last Page

333

Recommended Citation

Baluguri, A., Pasumarthy, V., Repakula, Y., Zhou, Z. (2026). Prosense - Defending Text Generation With Adversarial Feedback. Communications in Computer and Information Science, 2720 CCIS, 319-333.
Available at: https://aquila.usm.edu/fac_pubs/22098