Transformer-Driven Contextual Analysis for Fraud Detection in Conversational Voice Data

Abstract

Voice-based fraud, commonly referred to as voice phishing or vishing, has become a rapidly growing cybersecurity threat in modern communication systems. Fraudsters frequently exploit real-time voice conversations to impersonate legitimate entities such as banking representatives, government officials, or service providers in order to manipulate victims into revealing sensitive information including passwords, banking credentials, and one-time passwords. Traditional fraud detection systems largely rely on rule-based filtering mechanisms and predefined keyword matching techniques. Although such methods are capable of identifying known fraud patterns, they often fail to detect emerging scam strategies that rely on contextual manipulation and dynamic conversational tactics. Recent advancements in Natural Language Processing have introduced Large Language Models capable of analyzing conversational context with high accuracy. However, these models often require substantial computational resources and are difficult to deploy efficiently in real-time environments due to high latency and infrastructure requirements. These limitations highlight the need for more efficient contextual analysis techniques capable of detecting fraudulent intent in conversational data while maintaining computational efficiency.This research proposes a transformer-driven contextual analysis framework for detecting fraud in conversational voice data. The proposed approach processes voice conversations by converting audio signals into textual transcripts using speech recognition techniques, followed by text preprocessing and contextual language analysis using a lightweight transformer model. By analyzing semantic relationships and conversational patterns, the system identifies linguistic cues associated with fraudulent behavior such as urgency-based requests, impersonation tactics, and attempts to obtain confidential information.Experimental evaluation demonstrates that transformer-based contextual analysis can achieve high detection accuracy while maintaining significantly lower computational overhead compared to large language model approaches. The proposed framework therefore provides an efficient and scalable solution for detecting voice-based fraud in real-time communication systems.