Detecting Fraudulent Activities In Telecom Networks Through Call Detail Analysis And Big Data Techniques

Abstract

Telecommunication networks have become prone to advanced fraud programs, resulting in a loss of more than USD 46.3 billion per year. This paper presents a hybrid fraud detection model that exploits both the properties of Call Detail Record (CDR), big data methods and machine learning techniques to boost real-time detection of anomalies in telecommunication networks. Taking advantage of Apache Spark as a distributed processing engine and TensorFlow as a deep learning framework, the system manages to process CDR data in terabytes. The results of both real-world datasets and simulated datasets experiments have shown a higher performance of Long Short-Term Memory (LSTM) networks with the accuracy of 98.1% and the ROC-AUC score of 99.0%, which was significantly higher than classical rule-based systems. The blockchain technology makes it impossible to alter the logging history, which makes it audit-friendly, whereas IoT integration will allow detecting anomalies at the edge. The hybrid solution minimised false positives to below 3 per cent and had a throughput of 50,000 CDRs every second, which is a good indication of its scalability and robustness. This framework fills the most important gaps in the existing fraud management systems as it unites high accuracy of the detection process with transparency in the operations. The future studies would consider federated learning to train models on privacy-sensitive data in a decentralised fashion and modify the framework to the 6G and AI-facilitated frauds. The findings mean that there is a massive possibility that the technology would be used in the development of a contemporary telecom infrastructure that would help protect revenues and win customer confidence.