HYDERABAD: Researchers have developed a hybrid deep-learning model to detect suspicious transactions, unauthorised access and unusual system changes in enterprise software such as SAP S/4HANA.
The study by researchers associated with Osmania University (OU) and Jawaharlal Nehru Technological University, Hyderabad (JNTU-H), proposes a framework combining three deep-learning techniques to identify unusual financial postings, configuration changes, authorisation failures and violations of established workflows. Unlike conventional monitoring systems that rely on predefined rules and thresholds, the model analyses relationships between business processes and transaction patterns over time.
It combines a graph neural network to map process relationships, a Transformer-based model to analyse transaction sequences and a variational autoencoder to learn patterns while accounting for uncertainty. An interpretability component identifies the transaction attributes and process stages behind alerts, helping auditors investigate potential irregularities.
Tests using synthetic SAP-related data produced a receiver operating characteristic area under the curve (ROC-AUC) score of 0.956, a precision-recall area under the curve (PR-AUC) score of 0.691 and a false-positive rate of 7.1%. The researchers said the model outperformed comparison models. Tests using adapted data from the Numenta Anomaly Benchmark yielded an ROC-AUC score of 0.901.
The framework could support continuous auditing, risk management and regulatory compliance. However, its main quantitative results were based on synthetic and benchmark datasets. Although the researchers examined actual SAP S/4HANA audit logs in a qualitative case study, further testing on large-scale, real-world deployments is needed to establish the model’s reliability and generalisability.
The paper was authored by Ram Reddy Jonnalagadda of OU; Kiran Kumar Reddy and Pell Reddy Rajender Reddy of JNTU-H; Shshank Chaube of Symbiosis Institute of Technology’s Hyderabad campus; BP Joshi of Graphic Era Hill University; and Manish Kumar.