Anomaly detection in double-entry bookkeeping data by federated learning system with non-model sharing approach.

Area of Science:

• Financial Auditing
• Machine Learning
• Data Security

Background:

• Anomaly detection in financial auditing requires extensive data from multiple organizations.
• Confidentiality concerns prevent traditional data sharing among audit firms.
• Existing federated learning (FL) methods involve extensive communication and network exposure.

Purpose of the Study:

• To develop a non-model-sharing FL framework for anomaly detection in financial auditing.
• To enable secure data collaboration without exposing raw data or requiring constant network connectivity.
• To improve the efficiency and confidentiality of anomaly detection in audits.

Main Methods:

• Proposed a data collaboration (DC) analysis framework, a non-model-sharing FL technique.
• Utilized dimensionality reduction for secure intermediate data representations.
• Employed an autoencoder built on collaboration representations, requiring only one communication round.

Main Results:

• The DC-based approach outperformed locally trained models and traditional FL methods (FedAvg, FedProx).
• The framework demonstrated superior performance, especially under non-i.i.d. conditions.
• Achieved effective anomaly detection while preserving data confidentiality.

Conclusions:

• Organizational knowledge can be integrated for advanced auditing while maintaining data confidentiality.
• The DC framework offers a practical solution for intelligent auditing systems.
• This method advances secure and efficient anomaly detection in collaborative financial audits.