Private Graph Extraction via Feature Explanations

Privacy and interpretability are two important ingredients for achieving trustworthy machine learning. We study the interplay of these aspects in graph learning through reconstruction attacks. The goal adversary here is to reconstruct structure training data given access model explanations. Based on different kinds auxiliary information available adversary, we propose several show that additional knowledge post-hoc feature explanations substantially increases success rate Further, investigate detail differences between attack performance with respect three classes explanation methods neural networks: gradient-based, perturbation-based, surrogate model-based methods. While gradient-based reveal most terms structure, find do not always score high utility. For other explanations, privacy leakage an increase Finally, a defense based randomized response mechanism releasing which reduces rate. Our code at https://github.com/iyempissy/graph-stealing-attacks-with-explanation.