Secure ERP systems through LLM-based fraud detection

• Applicant: Professor Dr. Bernd Scheuermann
• Subject Area: Data Management, Data-Intensive Systems, Computer Science Methods in Business Informatics ; Methods in Artificial Intelligence and Machine Learning ; Security and Dependability, Operating-, Communication- and Distributed Systems
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 554438972

The widespread use of IT systems for managing business processes provides multiple opportunities for fraudulent behavior, negatively impacting the company's results. Employees capable of committing fraud due to their position in the company and their personal skills cause significant financial damage. The Association of Certified Fraud Examiners estimates this loss at around five percent of a company's annual turnover. To counteract this, measures are needed to prevent and detect risks and fraud. In a constantly changing digital environment, the use of machine learning (ML) for fraud detection seems to be appropriate. ML offers the opportunity to identify risks and fraud cases that cannot be detected using common risk assessment rules. ML is already being used successfully for fraud detection in areas such as credit card and insurance fraud. ERP systems, particularly those from SAP, capture extensive business data that can be used for fraud detection. However, a key challenge is that companies are reluctant to provide real data. At the same time, synthetic data is often not realistic or does not cover all relevant business areas. The SeLLMa knowledge transfer project addresses this problem with three goals: (1) Generation of realistic SAP fraud data: With the help of LLM-based software agents, both legitimate and fraudulent business processes are to be simulated in an SAP system. This enables the creation of a realistic benchmark library that is publicly accessible and serves to develop and objectively evaluate fraud detection methods. (2) Fraud detection through LLM-based data analysis: SeLLMa explores the potential of generative AI and data analysis to detect fraud. LLMs can recognize patterns, analyze user similarities, and evaluate business processes to identify potential fraud cases. In addition, the combination of traditional ML techniques and LLM-based approaches to fraud detection will be investigated. (3) Chatbot-based fraud verification: A chatbot is designed to facilitate the verification of suspected fraud cases by identifying relevant SAP data and presenting it in the business process context. By integrating knowledge of SAP processes and authorization concepts, the LLM can support the user in verifying a suspected case. The collaboration between Karlsruhe University of Applied Sciences and Pointsharp GmbH in SeLLMa enables a well-founded scientific investigation of the suitability of LLMs for fraud detection and verification and a transfer of the findings into practice.

• DFG Programme: Research Grants (Transfer Project)
• Application Partner: Pointsharp GmbH