نوع مقاله : مقاله پژوهشی
نویسنده
گروه حسابداری، دانشگاه علامه طباطبائی
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسنده [English]
The detection of anomalies and fraudulent activities in accounting records has become increasingly critical in modern auditing practices, particularly in the era of big data where traditional sampling methods are insufficient. This study proposes a novel approach utilizing deep autoencoder neural networks for anomaly detection at the transaction level within accounting information systems. Two large-scale datasets were used: 36,538 journal entries from the Rahkaran system and 30,000 from the Sepidar system. Artificial anomalies were injected to evaluate performance. The autoencoder was trained in an unsupervised manner using PyTorch, with reconstruction error as the anomaly indicator. The empirical results indicate that the proposed model significantly outperforms conventional detection techniques, demonstrating a strong ability to identify both global anomalies (e.g., unusual amounts or transaction timings) and contextual anomalies (e.g., rare attribute combinations). Key features included subsidiary account, general ledger code, cost center, and last modification date. The findings provide strong evidence that deep learning-based anomaly detection can substantially improve fraud risk assessment and enhance the reliability of financial reporting, thereby offering a powerful tool for auditors, regulators, and financial system designers.
کلیدواژهها [English]
)