University Events

Description: EAS Doctoral Proposal Defense by Gaspard Baye

Date: Friday, November 17, 2023

Time: 10:00 a.m.

Topic: OpenLLM: A Generative Multi-Agent Network for Open World Intrusion Detection

Zoom Teleconference:
https://umassd.zoom.us/j/92608700069?pwd=cUdsaDFHMnd6YUpNakVqS0FmaXRYQT09

Location: Dion 307

Abstract:
Network intrusion detection is a pivotal element of cybersecurity, essential for identifying and mitigating threats to computer networks. Conventional intrusion detection systems excel at recognizing established attack patterns but often grapple with novel or previously unknown threats, including zero-day detections. This research introduces "OpenLLM," a generative Multi-Agent Network tailored for Open-World intrusion detection. OpenLLM harnesses the potency of generative machine learning techniques and employs a multi-agent architecture to enhance network intrusion detection capabilities. The framework identifies familiar intrusion patterns and potential unknowns, including zero-day detections. By modifying the generative transformers' neural network to identify unknown or novel inputs, OpenLLM extends the boundaries of intrusion detection, contributing to proactive network asset protection in dynamic, Open-World environments. This work delves into developing and evaluating OpenLLM, emphasizing its pioneering contributions to Open-World intrusion detection, particularly concerning zero-day detections. Extensive experimental results validate the framework's efficacy in detecting known and unknown intrusion activities, underscoring its potential to strengthen network security. The research further addresses practical considerations, including implementation and performance benchmarks, setting the stage for integrating OpenLLM into real-world network security systems. This dissertation proposal (i) presents an exploration of existing deep learning (DL) based Open Set Classifiers, and (ii) performs a comprehensive evaluation of the performance of existing DL-based Open Set classifiers. Our results indicate that existing Open Set Classifiers need to be significantly enhanced to be used in network intrusion domain. We also lay out the methodology for (i) creating an enhanced DL-based Open Set classifier, (ii) extension of the research by introducing LLM-based explainability for detected unknowns, (iii) modification of the transformer neural network to accommodate open-world recognition, and (iv) Implementation and performance benchmarking facilitate the seamless integration of OpenLLM into practical, real-world network security systems.


ADVISOR(S):
Dr. Gokhan Kul, Department of Computer and Information Science
(gkul@umassd.edu)

COMMITTEE MEMBERS:
Dr. Lance Fiondella, Department of Electrical and Computer Engineering
Dr. Jiawei Yuan, Department of Computer and Information Science
Dr. Ming Shao, Department of Computer and Information Science
Dr. Long Jiao, Department of Computer and Information Science

NOTE: All EAS Students are ENCOURAGED to attend.