[Re] Subspace Attack: Exploiting Promising Subspaces for Query-Efficient Black-box Attacks

Abstract

As part of the NeurIPS 2019 Reproducibility Challenge, we chose to attempt to reproduce the attack algorithm proposed in “Subspace Attack: Exploiting Promising Subspaces for Query-Efficient Black-box Attacks”. Our reported results are better than the original paper in terms of the median number of queries per attack, but worse in terms of failure rate. A concise assessment of our implementation is also included.

Edoardo Debenedetti

Computer Science PhD Student

My research interests include Machine Learning and Security.