Team04-AI Enabled Scam Call Detection

Indian Institute of Science Summer 2025 DA225o Submission20 Authors

Keywords: Vishing, ASR, NLU, LLM, Few-shot learning, DL, Human-AI Interaction, Fraud Detection

Abstract: The increasing number of frauds via phone calls and text messages has called for proactive and intelligent defense mechanisms. Our project proposes the design and development of an AI-powered fraud call detection application, whose concept is capable of operating on edge devices. The system leverages few-shot learning to analyze transcribed speech to detect patterns associated with scam calls, such as social engineering, urgency cues, and suspicious requests for sensitive information. The proposed application integrates automatic speech recognition (ASR), natural language understanding (NLU), and a decision engine to evaluate the risk of a conversation. When a potential fraud is detected, the application alerts the user with a concise explanation and suggests contextually appropriate questions to verify the caller’s authenticity. The entire inference process preserves user privacy without relying on cloud services. In future, it can be optimized for edge deployment. This work aims to demonstrate the feasibility of a privacy-aware, agentic AI system for fraud prevention, with potential applications in personal communication.

Submission Number: 20