Brief
Hyderabad researchers unveil AI model to spot SAP fraud
The hybrid model combines graph neural networks, Transformers and variational autoencoders, achieving an ROC‑AUC of 0.956 on synthetic data.
By Felo News Desk · Published
Researchers from Osmania University and Jawaharlal Nehru Technological University, Hyderabad have created a hybrid deep‑learning framework to detect fraudulent activity in SAP S/4HANA systems, the New Indian Express reported on 11 October 2026.
The system merges three techniques: a graph neural network to map process relationships, a Transformer‑based model to analyse transaction sequences, and a variational autoencoder that learns patterns while handling uncertainty. An interpretability layer highlights the attributes and process stages that trigger alerts, aiding auditors.
Testing on synthetic SAP data yielded a receiver operating characteristic area under the curve (ROC‑AUC) of 0.956, a precision‑recall AUC of 0.691 and a false‑positive rate of 7.1%. Benchmarks using the Numenta Anomaly Benchmark produced an ROC‑AUC of 0.901. The authors note that results are based on synthetic and benchmark datasets, and further real‑world validation is required.
Key facts
- The model combines a graph neural network, a Transformer and a variational autoencoder. (newindianexpress.com)
- Synthetic tests gave an ROC‑AUC of 0.956 and a false‑positive rate of 7.1%. (newindianexpress.com)
- Researchers are from Osmania University, JNTU‑Hyderabad, Symbiosis Institute of Technology, Graphic Era Hill University and others. (newindianexpress.com)
Sources
- [1] newindianexpress.com — originally reported as “Hyderabad universities find new AI model to help detect SAP fraud”
