Hybrid Telecom Fraud Detection with Machine Learning, Large Language Models, and Blockchain

Abstract

This paper presents a first-of-its-kind modular AI framework for telecom fraud detection, integrating machine learning (ML), large language models (LLMs), and blockchain smart contracts to unify statistical classification, semantic reasoning, and decentralized enforcement. A synthetic dataset of 100 users across 300 sessions in Birmingham, UK, simulated telecom usage with$\mathbf{1 \% - 5 \%}$injected fraud, including GPS spoofing, excessive transmission power, and prolonged usage. Seven ML models were trained, with Random Forest optimized using a precision-recall threshold of$\mathbf{0. 7 2 1 7}$. Six configurations varied the decision logic between ML and GPT-4o-based LLMs, with LLMs performing context-aware reasoning via behavioral prompts. Solidity smart contracts on a local Ethereum network enforced decisions, mapping users to blockchain identities with a Proof-of-Stake-style validation mechanism. The ML-only configuration achieved 92.25 % accuracy with perfect user-level precision and recall, while LLM variants enhanced behavioral and temporal reasoning. This framework advances robust and explainable fraud detection for future telecom infrastructures.