Privacy-Preserving k-NN Graphs with Autoencoder-Based Representations for Sensitive Features
Gustavo Lima de Oliveira, Maria da Graça Campos Pimentel, Ricardo Marcondes Marcacini · 2024
Privacy-preserving representation learning has gained significant attention for enabling secure data and model sharing by protecting sensitive information while maintaining data utility. In this paper, we present a new approach to privacy-preserving representation learning with k-NN-based graph models. This method maps the original feature space to a new space that balances feature utility, such as classification accuracy, with reducing privacy attack risks, and constructs a kNN graph from this new space. We evaluate three scenarios using real datasets to assess privacy-preserving graph representations. Experimental results show that learning a privacy-preserved representation and constructing a k-NN graph is a simple, intuitive, and competitive approach compared to other methods in the literature. Thus, this method enables graph data sharing with a lower risk of sensitive information extraction attacks.