Seminar Presentations

Privacy-Friendliness of Feature Extractors: Empirical Insights, Metrics and Correlations

"Feature extraction is widely used to improve utility in differentially private (DP) classification, often under the assumption that pretrained or foundation models inherently provide privacy-friendly representations. In this presentation, we challenge this assumption through an empirical study of feature extraction methods across multiple image classification datasets. Our aim is to analyse how dataset metrics relate to downstream DP performance and what insights they offer into dataset characteristics. Finally, we apply these metrics to more challenging facial classification datasets. First findings suggest that strong features are not necessarily privacy-friendly features, emphasising the importance of a more systematic evaluation of representation learning in privacy-sensitive settings."