IMPROVING TELECOMMUNICATION (TELECOM) FRAUD DETECTION ACCURACY THROUGH ADVANCED ARTIFICIAL INTELLIGENCE MODELS

Abstract

The criminal justice system addresses telecom fraud as a major concern. Artificial intelligence development has led to increasingly sophisticated and creative telecom fraud texts. Unfortunately, when it comes to detecting telecom fraud in real-time, current security techniques like cell number tracking and detection and conventional machine-learning-based text identification aren't exactly top choices. An LSTM network-based deep learning architecture is suggested for the aim of identifying fraudulent behavior in CDRs in this study. The preprocessing involved in the methodology is extensive, including data cleaning, elimination of irrelevant attributes, critical discovery of outliers, and random under-sampling to curb extreme cases of class imbalances. A robust set of input variables was constructed using feature extraction and engineering methods and performance estimation was performed using the stratified k-fold cross-validation technique. Temporal relationships and complex patterns in sequential telecom data were used to train the LSTM model. Its performance was evaluated in comparison to NN, QA, and LR, three more strategy-based models. With an F1-score of 99.53, a recall of 99.37, a precision of 99.68, and an area under the curve (AUC) of 1.0000, the results demonstrate that the suggested LSTM model performed better than the competition. Instead, NN, QA and LR provided lower and less consistent results. The results validate the strengths, elasticity, and real-time applicability of the presented approach in detecting telecom fraud on large scale.