نوع مقاله : مقاله پژوهشی
نویسنده
گروه حسابداری، دانشگاه علامه طباطبائی
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسنده [English]
Anomaly and fraud detection has always been one of the primary challenges for auditors and examiners. This study aims to establish a novel framework for applying emerging deep learning techniques in financial auditing by employing Autoencoder Neural Networks (AE-NN) for anomaly detection in accounting information systems. First, the theoretical foundations of anomalies in bookkeeping and financial document registration are reviewed. Then, two real-world datasets were employed: the Rahkaran system dataset containing 36,538 accounting entries and the Sepidar system dataset containing 30,000 entries, both covering fiscal years 2021–2023. To evaluate the detection power of the proposed method, several artificial anomalies were injected into the datasets. The empirical results indicate that the AE-NN demonstrates a superior ability in identifying anomalous accounting entries compared to alternative machine learning techniques, including Decision Trees, Extra Trees Regressor, Random Forest, AdaBoost Classifier, and Quadratic Discriminant Analysis. Furthermore, increasing the network depth led to an improvement in anomaly detection accuracy. The study also reveals that reconstruction error varies across subsystems generating the documents, highlighting the necessity of considering the origin subsystem when performing anomaly detection. In addition, the most influential features in detecting anomalies were found to be counterparty account (detailed ledger), subsidiary account, cost center, and the last modification date of the document. Overall, the findings demonstrate that AE-NN provides an effective and practical approach for enhancing the reliability of accounting information systems and supporting auditors in fraud detection.
کلیدواژهها [English]
)