Timezone: »

Provably Robust Metric Learning
Lu Wang · Xuanqing Liu · Jinfeng Yi · Yuan Jiang · Cho-Jui Hsieh

Wed Dec 09 09:00 AM -- 11:00 AM (PST) @ Poster Session 3 #908

Metric learning is an important family of algorithms for classification and similarity search, but the robustness of learned metrics against small adversarial perturbations is less studied. In this paper, we show that existing metric learning algorithms, which focus on boosting the clean accuracy, can result in metrics that are less robust than the Euclidean distance. To overcome this problem, we propose a novel metric learning algorithm to find a Mahalanobis distance that is robust against adversarial perturbations, and the robustness of the resulting model is certifiable. Experimental results show that the proposed metric learning algorithm improves both certified robust errors and empirical robust errors (errors under adversarial attacks). Furthermore, unlike neural network defenses which usually encounter a trade-off between clean and robust errors, our method does not sacrifice clean errors compared with previous metric learning methods.

Author Information

Lu Wang (Nanjing University & JD.com)
Xuanqing Liu (University of California, Los Angeles)
Jinfeng Yi (JD Research)
Yuan Jiang (National Key lab for Novel Software Technology)
Cho-Jui Hsieh (UCLA)

More from the Same Authors