FRAUDLENS DETEKSI FRAUD LAPORAN KEUANGAN BERBASIS BENFORD DAN ISOLATION FOREST

Abstract

Financial fraud detection is often hindered by the fundamental limitations of single method approaches, where statistical methods are prone to overlooking micro scale manipulation while

Artificial Intelligence (AI) algorithms risk bias when facing massive data distortion. To address these security gaps, this study aims to design and build a fraud detection system named FraudLens to produce an effective and affordable computer assisted audit instrument for micro entities. Using a Research and Development (R&D) approach with a prototyping model, this research utilizes simulation data (synthetic data) divided into three dataset scenarios (normal, minor manipulation, and massive) to overcome real data privacy constraints. The system development uses Python 3.10 and the Pandas library with a Streamlit web based interface, applying a hybrid analysis tool that integrates Benford's Law statistical method and the Unsupervised Machine Learning Isolation Forest algorithm. Research findings reveal significant contradictory results: in the minor manipulation scenario, the statistical method failed to detect anomalies, yet the Isolation Forest algorithm achieved a 100% detection rate, conversely, in the massive manipulation scenario, the AI algorithm experienced a fatal detection failure with only a 5.3% success rate due to the Masking Effect phenomenon, while Benford's Law successfully identified manipulation indications with a MAD score of 0.0817. The novelty of this research lies in the empirical proof that without cross validation, AI is susceptible to Normalization of Deviance, making the integration of both methods absolutely necessary to create a double layered defense mechanism that covers the blind spots of each approach in maintaining audit integrity.