SISA Based Machine Unlearning for Credit Card Default Prediction with Shard Optimization

Abstract

Machine unlearning addresses a critical challenge in modern AI based financial systems by enabling the removal of user data while maintaining model accuracy and efficiency without retraining the model from scratch. In applications such as credit card default prediction, models are trained on sensitive financial data, creating a need for efficient data deletion mechanisms. Traditional approaches require complete retraining of the model after removing the requested data, which is computationally expensive and impractical in real world scenarios. To tackle this issue, this paper uses a framework based on the SISA (Sharded, Isolated, Sliced, and Aggregated) architecture. The method divides the dataset into independent shards and trains models accordingly, allowing targeted retraining of only the affected shard when a deletion request occurs. Experimental evaluation on the UCI Credit Card Default dataset shows that the proposed method achieves up to ~146.6× reduction in unlearning time compared to full retraining while maintaining stable predictive performance (Accuracy $\approx 0.817, \text{AUC} \approx 0.78$, post-unlearning degradation $<0.001)$. This paper highlights the trade off between unlearning efficiency and model quality, which can be optimized through appropriate shard size selection. Verification using membership inference and cross-entropy analysis confirms that the influence of removed data is effectively eliminated, while SHAP-based evaluation shows that model interpretability remains largely unaffected. The results shows that efficient and scalable machine unlearning is feasible for real world financial systems, allowing continuous learning models to follow with regulatory requirements without compromising performance or reliability.