Timezone: »

Robust Perception through Equivariance
Chengzhi Mao · Lingyu Zhang · Abhishek Joshi · Junfeng Yang · Hao Wang · Carl Vondrick

Thu Jul 27 01:30 PM -- 03:00 PM (PDT) @ Exhibit Hall 1 #109

Deep networks for computer vision are not reliable when they encounter adversarial examples. In this paper, we introduce a framework that uses the dense intrinsic constraints in natural images to robustify inference. By introducing constraints at inference time, we can shift the burden of robustness from training to testing, thereby allowing the model to dynamically adjust to each individual image's unique and potentially novel characteristics at inference time. Our theoretical results show the importance of having dense constraints at inference time. In contrast to existing single-constraint methods, we propose to use equivariance, which naturally allows dense constraints at a fine-grained level in the feature space. Our empirical experiments show that restoring feature equivariance at inference time defends against worst-case adversarial perturbations. The method obtains improved adversarial robustness on four datasets (ImageNet, Cityscapes, PASCAL VOC, and MS-COCO) on image recognition, semantic segmentation, and instance segmentation tasks.

Author Information

Chengzhi Mao (Columbia University)
Lingyu Zhang (Columbia University)
Abhishek Joshi (Columbia University)
Junfeng Yang (Columbia University)
Hao Wang (Rutgers University)
Hao Wang

Dr. Hao Wang is currently an assistant professor in the department of computer science at Rutgers University. Previously he was a Postdoctoral Associate at the Computer Science & Artificial Intelligence Lab (CSAIL) of MIT, working with Dina Katabi and Tommi Jaakkola. He received his PhD degree from the Hong Kong University of Science and Technology, as the sole recipient of the School of Engineering PhD Research Excellence Award in 2017. He has been a visiting researcher in the Machine Learning Department of Carnegie Mellon University. His research focuses on statistical machine learning, deep learning, and data mining, with broad applications on recommender systems, healthcare, user profiling, social network analysis, text mining, etc. His work on Bayesian deep learning for recommender systems and personalized modeling has inspired hundreds of follow-up works published at top conferences such as AAAI, ICML, IJCAI, KDD, NIPS, SIGIR, and WWW. It has received over 1000 citations, becoming the most cited paper at KDD 2015. In 2015, he was awarded the Microsoft Fellowship in Asia and the Baidu Research Fellowship for his innovation on Bayesian deep learning and its applications on data mining and social network analysis.

Carl Vondrick (Columbia University)

More from the Same Authors