This paper reports a machine learning study on detecting fraudulent banking transactions, carried out as an undergraduate project named Banking Intrusion Detection System. The dataset contains 1,000,000 transactions described by eight columns, of which 29,999 transactions (about 3 percent) are labelled as fraud. Timestamps were converted to date features, the merchant category was one-hot encoded, identifier columns were removed, and the data were split into 80 percent training and 20 percent test sets with stratification. A single Random Forest classifier with 100 trees and balanced class weights was trained and evaluated on 200,000 held-out transactions. The model reached an accuracy of 0.9457, but this is lower than the 0.97 obtained by labelling every transaction as non-fraud. For the fraud class, precision was 0.0293, recall was 0.0252 and the ROC-AUC was 0.4970, which is no better than random ranking. Lowering the decision threshold raised recall to at most 0.0713 in the tested range while accuracy fell to about 0.90 and precision stayed near the fraud share of the data. Supplementary checks show that the fraud rate is almost constant across merchant categories, months, weekdays and amount ranges. The study therefore does not demonstrate a usable detector. It documents a reproducible baseline, its limitations, and directions for improvement.

Mahaswetha K, Gopika T, Kaasini M, Bhuvan Kruthik, Dhanya K R,"Banking Intrusion Detection System: A Random Forest Approach to Fraudulent Transaction Classification" International Journal of Advanced Multidisciplinary Research and Educational Development, V2(5): Page(753-757) September - October 2026. ISSN: 3107-6513. www.ijamred.com. Published by Scientific and Academic Research Publishing.