Machine Learning-Based Detection of Fraudulent UPI Transactions Using Behavioural Features: A Prototype Study

UPI now carries a very large share of retail paymentsin India, and payment scams have grown alongside it: fake collectrequests, impersonation calls and takeover of accounts are allcommon. Fixed rules catch the obvious cases but are slow tochange when fraudsters change tactics. This paper describes asmall, reproducible Python prototype that labels a UPItransaction as legitimate or fraudulent from behavioural signalssuch as the amount compared with the user’s usual spending, thehour of the day, the age of the payee and whether the device haschanged. Real UPI records are not public, so a synthetic set of20,000 transactions (5.04% labelled fraud) was generated tofollow documented scam patterns. A logistic regression baselineand a random forest were trained with class weighting and testedon a stratified hold-out set. The random forest reached a ROCAUC of 0.887, a fraud precision of 0.985 and a recall of 0.758 (F1= 0.857), and a five-fold cross-validated ROC-AUC of 0.907. Wealso show how moving the decision threshold trades precision forrecall. The numbers show that the pipeline works; they do notpredict accuracy on real traffic, and the limits of synthetic dataare discussed.