A Targeted Machine Unlearning Method for Sensitive Data in Military Helicopter Models

This paper addresses the need for machine unlearning technology in deep learning models, particularly for applications in military domains. While deep learning models exhibit outstanding performance across various fields, they also pose privacy risks, as training data is embedded in the model’s parameters. This makes them vulnerable to model extraction attacks that can potentially leak sensitive information. To counter such risks, machine unlearning technology enables models to selectively “forget” specific data post-training, but existing studies are largely limited to toy datasets. In response, this study introduces a novel unlearning method tailored to military helicopter models, enabling the removal of specific data points without affecting overall model performance. By applying this technique to an actual military dataset, our approach is capable of deleting data from particular operational scenarios without requiring entire class removal or model retraining. The paper’s contributions include presenting a structured machine unlearning technique for military data, evaluating its performance via various metrics, and validating its effectiveness through experiments with a ResNet18 model and a simulated military dataset.