ML-Driven Audit Risk Assessment with Differential Privacy
Zikun Zhang, Ghulam Mohi-ud-din, Yang Xiong, Xu Zhang, Xiao Lin, Chen Ai, Min Qiang Hu · 2024
In contemporary auditing practices, the ability to effectively identify and assess potential fraudulent companies is crucial for maintaining market fairness and transparency. Previous research has leveraged various machine learning techniques to detect fraudulent activities in financial data. However, these models often lack robustness and fail to protect the privacy of the financial information used for training. Here we introduces a novel approach by incorporating differential privacy into the training process, ensuring data protection without compromising model performance. To enhance model performance, we employed interquartile range filtering for outlier removal and optimized feature selection using Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Principal Component Analysis (PCA). For model training, we employed Logistic Regression, Random Forest, and Decision Trees, with hyperparameters meticulously fine-tuned through grid search and Bayesian optimization techniques to achieve optimal performance. To address privacy concerns, we implemented differential privacy using the Laplace mechanism, effectively safeguarding sensitive financial data. Our empirical results demonstrate that the enhanced models maintain high accuracy and robustness, providing a reliable tool for fraud detection while ensuring financial data privacy.