Machine Learning–Based Fraud Detection and Conventional Audit Approaches in Government Deposit Processing

Abstract

This study investigates the problem of fraud vulnerability in government deposit processing, where duplicate entries, delayed reconciliation, unauthorized adjustments, false receipts, abnormal deposit timing, and inconsistent depositor records can weaken public financial accountability and reporting reliability. The purpose of the study was to examine how machine learning based fraud detection and conventional audit approaches influence fraud detection effectiveness, audit efficiency, reporting reliability, and financial control strength in public sector deposit environments. A quantitative, cross sectional, case-based research design was adopted, using structured survey data from 200 professionals representing government finance, accounting, internal audit, external audit, compliance, treasury operations, IT audit, and enterprise financial system cases. The key variables included machine learning based fraud detection, conventional audit approaches, fraud detection effectiveness, audit efficiency, reporting reliability, financial control strength, deposit fraud risk signals, and comparative trust between machine learning and traditional audit methods. The analysis plan involved descriptive statistics, Cronbach’s alpha reliability testing, Pearson correlation analysis, regression modeling, hypothesis testing, fraud risk signal mapping, and comparative trust indexing. The findings showed strong support for machine learning based fraud detection, with a mean score of 4.21 and SD of 0.61, compared with conventional audit approaches at M = 3.89 and SD = 0.68. Fraud detection effectiveness recorded M = 4.12, audit efficiency M = 4.05, reporting reliability M = 4.08, and financial control strength M = 4.16. Reliability values were acceptable, ranging from 0.82 to 0.90. Correlation results showed that machine learning based fraud detection had a strong positive relationship with fraud detection effectiveness, r = 0.71, p < 0.01, while conventional audit approaches also showed a significant relationship, r = 0.58, p < 0.01. Regression results confirmed that machine learning and conventional audit together explained 47% of fraud detection effectiveness, with machine learning showing the stronger effect, β = 0.52, p < 0.001. The findings imply that government agencies should integrate machine learning screening with audit verification to improve accountability, detection speed, evidence reliability, and financial control.