Abstract
Purpose
This research introduces an innovative framework that integrates Federated Learning (FL), Split Learning (SL), and Differential Privacy (DP) to enable secure training of deep learning models on sensitive accounting data across multiple audit clients.
- Supports auditors in detecting anomalies in ERP journal entries using AI techniques
- Provides a privacy-preserving solution tailored for distributed, confidential financial data
- Balances technological advancement with the ethical and regulatory demands of the auditing profession
Scope
The framework addresses key challenges in applying AI to auditing by preserving data confidentiality while enhancing the effectiveness of audit procedures.
- Federated Learning Integration: Enables decentralized model training across audit clients without data pooling
- Split Learning Utilization: Distributes model architecture to further protect raw data during training
- Differential Privacy Application: Ensures individual data contributions remain untraceable
- Data Integrity Preservation: Maintains the accuracy and reliability of journal entry analysis
- Audit Enhancement: Facilitates more robust anomaly detection and risk assessment
- Ethical AI Implementation: Aligns with the values of client confidentiality, regulatory compliance, and transparency
This framework is particularly relevant to audit firms, AI researchers, and compliance professionals seeking responsible and effective ways to adopt AI in high-stakes financial environments.
Summary
Modern Challenges in Digital Auditing
The rapid digital transformation of financial reporting systems has resulted in vast volumes of complex data. Modern auditing standards call for the inspection of detailed digital accounting records, particularly journal entries. While AI and deep learning (DL) offer powerful tools to manage this information scale and complexity, concerns over data privacy and regulatory compliance have remained roadblocks to widespread adoption.
A Federated Learning-Based Solution
To address this issue, the authors present a Federated Learning (FL) framework, which allows multiple clients to train a shared AI model without exposing their data. Each client trains its local model using its private data, and a centralized model is updated using the aggregated learnings—without transferring any actual data. This preserves client confidentiality while still enabling the creation of industry-specific audit models.
Enhancing Privacy Through Differential Privacy
The study further incorporates Differential Privacy (DP), adding a layer of randomized noise to the learning process. Techniques like gradient clipping and Gaussian noise application ensure that individual data points from clients cannot be reconstructed, thus significantly mitigating the risk of data leakage. This move aligns with key regulatory guidelines, such as the GDPR and AICPA confidentiality rules.
Secure Partitioning via Split Learning
Split Learning (SL) is applied to further isolate sensitive model components. By dividing the neural network into public and private layers—where only the non-sensitive parameters are shared—SL allows more granular control over what each party exposes during training. This modular approach strengthens privacy while maintaining the model’s ability to detect critical financial anomalies.
Real-World Datasets and Experimental Design
The framework was tested on real-world civic payment datasets from cities such as Philadelphia, Chicago, and York. These datasets mirror ERP financial transactions. Through both independent and aggregated data configurations, the team evaluated the anomaly detection capabilities of the model in identifying both global anomalies (e.g., rare ledgers) and local anomalies (e.g., unusual value co-occurrences within journal entries).
Empirical Results Demonstrate Effectiveness
Results show that Federated Learning significantly improves anomaly detection performance compared to traditional centralized models, particularly when aggregating data from four or more clients. Meanwhile, the application of Differential Privacy with moderate noise levels not only preserves privacy but can even enhance model robustness, contradicting the common belief that noise always undermines accuracy.
Practical Considerations for Audit Firms
Audit entities have the flexibility to adjust how much of their model’s parameters are shared, based on their clients’ consent and regulatory conditions. This adaptability allows audit firms to balance between model performance and data protection, enabling them to tailor assurance services to specific industries or jurisdictions without violating privacy norms.
A New Era of Responsible AI Adoption
By integrating FL, DP, and SL, the framework allows auditors to harness the analytical power of deep learning while retaining compliance with data confidentiality requirements. It sets a technological benchmark for secure and effective AI applications in financial auditing, paving the way for more transparent and data-driven audits.
This research effectively demonstrates a scalable, ethical, and technically sound approach to embedding advanced AI methods in financial auditing workflows. It encourages audit firms and regulatory bodies to embrace responsible AI experimentation while safeguarding sensitive financial data, marking a pivotal step toward smarter and more secure assurance practices.
Resource
Read more in Federated and Privacy-Preserving Learning of Accounting Data in Financial Statement Audits by Marco Schreyer, Timur Sattarov, Damian Borth