Abstract:
Telecommunication fraud, especially through techniques like phishing calls, poses a
significant threat, resulting in substantial financial losses and compromising sensitive
personal data. While technological advancements have made life easier for consumers,
they have also introduced new vulnerabilities. Phishing through phone calls, wherein
individuals are tricked into revealing sensitive information, is a prevalent method
employed by fraudsters. Traditional approaches use blacklists of known fraudulent
numbers to detect phishing calls that are ineffective against new or previously unseen
numbers. To tackle the problem, I propose a system leveraging conversation transcript
analysis, combining an Automatic Speech Recognition (ASR) model to convert audio
to text and Natural Language Processing (NLP) techniques to analyze the text. Two
datasets are used in my experiment: a transcription dataset generated by a Bangla
ASR system and a clean conversational dataset. I demonstrate that fine-tuning deep
learning models to increase accuracy works well on transcription data provided by
ASR. The study utilizes four deep learning models: Artificial Neural Network (ANN),
Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and
Bidirectional Long Short-Term Memory (BiLSTM). The results demonstrate that the
RNN model attains the best accuracy of 97% on the clean conversational dataset. In
the transcription dataset, both LSTM and ANN models achieve highest accuracy of
83%. The results emphasize the capability of optimized deep learning models to
improve the identification of fraudulent phone calls, especially in situations where the
data is provided by ASR system. This provides a strong and effective approach to
detect telecommunication fraud.
