[Paper Review] Federated Continual Learning to Detect Accounting Anomalies in Financial Auditing

Marco Schreyer, Hamed Hemati+2 more|arXiv (Cornell University)|Oct 26, 2022

Imbalanced Data Classification Techniques4 citations

TL;DR

This paper proposes a Federated Continual Learning (FCL) framework for detecting accounting anomalies in financial auditing, enabling decentralized, privacy-preserving, and adaptive model training across dynamic client environments. The approach combines federated learning to aggregate models across clients and continual learning to mitigate catastrophic forgetting, achieving up to 99.84% global anomaly detection precision in real-world audit scenarios with data distribution shifts.

ABSTRACT

The International Standards on Auditing require auditors to collect reasonable assurance that financial statements are free of material misstatement. At the same time, a central objective of Continuous Assurance is the real-time assessment of digital accounting journal entries. Recently, driven by the advances in artificial intelligence, Deep Learning techniques have emerged in financial auditing to examine vast quantities of accounting data. However, learning highly adaptive audit models in decentralised and dynamic settings remains challenging. It requires the study of data distribution shifts over multiple clients and time periods. In this work, we propose a Federated Continual Learning framework enabling auditors to learn audit models from decentral clients continuously. We evaluate the framework's ability to detect accounting anomalies in common scenarios of organizational activity. Our empirical results, using real-world datasets and combined federated continual learning strategies, demonstrate the learned model's ability to detect anomalies in audit settings of data distribution shifts.

Motivation & Objective

• To address the challenge of training adaptive, privacy-preserving audit models in decentralized, dynamic organizational environments with evolving data distributions.
• To mitigate catastrophic forgetting in continual learning settings where audit models must retain knowledge of past financial patterns while learning new ones.
• To reduce model interference in federated learning by handling client-specific data distribution shifts across multiple audit clients.
• To evaluate the effectiveness of combined FCL strategies in real-world financial auditing scenarios involving temporal and client-based data shifts.
• To demonstrate that FCL enables accurate, real-time detection of accounting anomalies in continuous assurance settings.

Proposed method

• The framework employs a federated learning setup where a central server orchestrates model aggregation from multiple decentralized audit clients, preserving data privacy.
• Continual learning techniques—Replay, LwF, and EWC—are applied at each client to prevent catastrophic forgetting during incremental learning over time.
• The system integrates FedProx, Scaffold, and FedYogi as federated optimization strategies to reduce model interference caused by client data distribution shifts.
• An autoencoder-based anomaly detection model is used to identify journal entries with high reconstruction error, signaling potential accounting anomalies.
• Model performance is evaluated using average precision (AP) on global and local anomaly detection tasks across multiple real-world city government datasets.
• Experiments combine FL and CL strategies using five random seeds to ensure robustness and report variance in results.

Experimental results

Research questions

• RQ1 Can a federated continual learning framework effectively detect accounting anomalies in decentralized, dynamic audit environments with data distribution shifts?
• RQ2 How do different continual learning strategies (Replay, LwF, EWC) compare in mitigating catastrophic forgetting when learning from sequential audit data?
• RQ3 To what extent do federated optimization strategies (FedProx, Scaffold, FedYogi) reduce model interference caused by client-specific data distribution shifts?
• RQ4 Which combination of FL and CL strategies yields the highest anomaly detection precision in real-world audit datasets with temporal and client-based distribution shifts?
• RQ5 How does model performance vary across different anomaly classes and AEN model sizes under FCL settings?

Key findings

• The FedProx with EWC strategy achieved the highest global anomaly detection precision of 99.84% ± 2.97 on the City of Philadelphia dataset, significantly outperforming the FedAvg-Sequential baseline.
• The FedProx with Replay strategy achieved 74.11% ± 9.34 global AP on Philadelphia, demonstrating improved stability over sequential fine-tuning and baseline FedAvg.
• Scaffold with Sequential fine-tuning achieved 93.39% ± 4.23 global AP on Philadelphia, showing effectiveness in managing model convergence across dissimilar clients.
• On the City of York dataset, FedProx with EWC achieved 87.67% ± 3.11 global AP, indicating strong performance even with smaller AEN models and sparse data.
• The combination of FedProx and LwF achieved 76.18% ± 5.81 global AP on Chicago, showing robustness in handling distribution shifts across clients.
• Across all datasets, the FCL framework reduced the gap between federated and single-client learning performance, indicating effective mitigation of both catastrophic forgetting and model interference.