Vishing detection over call transcripts using Zero-shot Learning and machine learning
Loading...
Supplementary material
Other Title
Authors
Saradha, Anju
Author ORCID Profiles (clickable)
Degree
Master of Applied Technologies (Computing)
Grantor
Unitec, Te Pūkenga – New Zealand Institute of Skills and Technology
Date
2025
Supervisors
Ramirez-Prado, Guillermo
Barmada, Bashar
Barmada, Bashar
Type
Masters Thesis
Ngā Upoko Tukutuku (Māori subject headings)
Keyword
vishing
phishing
computer crimes
prevention
natural language processing
transfer learning (machine learning)
phishing
computer crimes
prevention
natural language processing
transfer learning (machine learning)
ANZSRC Field of Research Code (2020)
Citation
Saradha, A. (2025). Vishing detection over call transcripts using Zero-shot Learning and machine learning (Unpublished document submitted in partial fulfilment of the requirements for the degree of Master of Applied Technologies (Computing)). Unitec, Te Pūkenga - New Zealand Institute of Skills and Technology
https://hdl.handle.net/10652/6949
Abstract
RESEARCH QUESTIONS
• How can Zero-Shot Learning be utilized to detect fraud data without relying on task-specific labeled data?
• How does ZSL tackle the issue of limited vishing data for training models?
• How does ZSL perform compared to traditional ML and DL models?
ABSTRACT
Vishing assaults affect people and organizations by tricking victims into divulging personal information over the phone. Although deep learning models have been used in many fields, creating efficient models is difficult due to complicated datasets and insufficient data. Accuracy in voice-based phishing detection is frequently limited by additional challenges arising from data volume, accessibility, and privacy considerations. These issues are addressed by transfer learning, especially Zero-Shot Learning (ZSL), which uses semantic descriptions and pre-trained models to identify invisible categories.Financial loss, harm to one’s reputation, and data breaches can be avoided by being aware of social engineering techniques like mimicry, urgency, and authority.
In this thesis, a Zero-Shot Learning (ZSL) approach for identifying fraudulent vishing attempts from text-based voice conversation transcripts is proposed. In order to forecast fraud using attribute scores, the model takes textual elements and associates them with social engineering attributes, such as impersonation, urgency, authority, and sensitive information. ZSL is compared with traditional ML models (Naive Bayes, SVM, Logistic Regression, Decision Tree, Random Forest) and LSTM networks in experiments conducted on datasets with 1,000–5,000 entries. Without task-specific labeled data, the ZSL model shows its capacity to generalize and detect fraudulent material with high accuracy between 0.97 and 0.98 across all dataset sizes. By analyzing semantic context, ZSL also successfully manages ambiguous or invisible cases, demonstrating its promise as a scalable, label-efficient method for proactive fraud detection.
The evaluation underscores the ability of Zero-Shot Learning (ZSL) to identify deceptive content by interpreting the semantic context of a message rather than relying on previously labeled data. In one case, the ZSL model classified a message flagged as fraudulent in the dataset as normal due to its low alignment with predefined social engineering traits and low score of tactics, so this instance illustrates how ZSL uses the contextual information of the text compared to traditional models. This shows the ZSL’s capacity to handle unfamiliar or borderline cases with contextual awareness, addressing the research objective of exploring effective fraud detection without task-specific training data.
Publisher
Permanent link
Link to ePress publication
DOI
Copyright holder
Author
Copyright notice
All rights reserved