Abstract:
In an era marked by unprecedented data growth and pervasive digital influence, ensuring model privacy is imperative as machine learning models gain prominence in diverse domains like healthcare, finance, and business. Despite advancements in privacy-preserving methods, current approaches struggle to shield models against imitation without compromising accuracy or escalating computational costs. To address this, we draw inspiration from the financial concept of Ring Fencing, proposing a framework that establishes a virtual barrier around machine learning models. This innovative approach enhances privacy and security, allowing for privacy-preserving model sharing across institutions. Our framework encapsulates the model, dynamically adapting to a reduced set of features while withholding feature metadata. Rigorous experiments employing decision tree classifiers in healthcare and finance domains from multiple institutions, validate the efficacy of our approach. Notably, our SurrogateML + HE architecture demonstrates optimal performance, approaching 94% accuracy without a significant increase in computational cost.