Date of Graduation

5-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Engineering (PhD)

Degree Level

Graduate

Department

Electrical Engineering and Computer Science

Advisor/Mentor

Wu, Xintao

Committee Member

Panda, Brajendra N.

Second Committee Member

Chen, Jiahui

Third Committee Member

Li, Qinghua

Keywords

Contrastive Learning; Fraud Detection; In Context Learning; Noisy Label Learning; Open Set Anomaly; Positive Unlabeled Learning

Abstract

Detecting fraud in computing platforms involves identifying malicious user sessions, often using deep learning models, but several challenges hinder effective deployment. Attackers can craft diverse malicious sessions that closely resemble normal ones, complicating the learning of robust decision boundaries. While supervised contrastive learning offers a promising solution through class-specific clustering, its potential remains underexplored. Real-world datasets typically contain few labeled malicious sessions and many normal ones, creating an open-set anomaly detection challenge. Costly expert annotation further limits labeled data, especially for smaller organizations, leading to Positive Unlabeled (PU) learning and noisy label learning issues. Organizations are increasingly turning to LLMs for their adaptability, minimal retraining needs, and the capability of In-Context Learning (ICL) to adapt to evolving fraud patterns, though this potential is not yet fully realized. This dissertation addresses these challenges using supervised contrastive learning to develop practical fraud detection frameworks, with key contributions outlined below.

We present a robust supervised contrastive learning based fraud detection framework which operates in the open-set anomaly detection setting;

We present a supervised contrastive learning based fraud detection framework which operates in the PU learning setting;

We present a supervised contrastive learning based fraud detection framework which operates in the noisy label setting.

We present an ICL based framework for coded fraud text and hate speech detection which leverages supervised contrastive learning for demonstration selection.

Download

Share

COinS