Extracting Information from Hypersphere-Based Machine Learning Models

This paper introduces a novel algorithm capable of executing both model inversion and parameter stealing attacks against hypersphere-based machine learning models. The algorithm consists of two main steps: reflection and infection. We evaluate the proposed algorithm on three datasets: a randomized dataset, the RT-IoT2022 dataset, and a handwritten digits dataset. The evaluation demonstrates the algorithm's effectiveness in capturing model parameters, even in higher-dimensional spaces. Additionally, it can extract meaningful information from the targeted model's training dataset.