(Preprint) Towards Understanding the Generative Capability of Adversarially Robust Classifiers
Yao Zhu ¹, Jiacheng Ma ², Jiacheng Sun ², Zewei Chen ², Rongxin Jiang 蒋荣欣 ¹, Zhenguo Li ²
¹ Zhejiang University
² Huawei Noah’s Ark Lab
arXiv, 2021-08-20

Recently, some works found an interesting phenomenon that adversarially robust classifiers can generate good images comparable to generative models. We investigate this phenomenon from an energy perspective and provide a novel explanation. We reformulate adversarial example generation, adversarial training, and image generation in terms of an energy function. We find that adversarial training contributes to obtaining an energy function that is flat and has low energy around the real data, which is the key for generative capability.

Based on our new understanding, we further propose a better adversarial training method, Joint Energy Adversarial Training (JEAT), which can generate high-quality images and achieve new state-of-the-art robustness under a wide range of attacks. The Inception Score of the images (CIFAR-10) generated by JEAT is 8.80, much better than original robust classifiers (7.50). In particular, we achieve new state-of-the-art robustness on CIFAR-10 (from 57.20% to 62.04%) and CIFAR-100 (from 30.03% to 30.18%) without extra training data.
Towards Understanding the Generative Capability of Adversarially Robust Classifiers_1
Towards Understanding the Generative Capability of Adversarially Robust Classifiers_2
Towards Understanding the Generative Capability of Adversarially Robust Classifiers_3
  • Power grid fault diagnosis based on a deep pyramid convolutional neural network
  • Xu Zhang 张旭, Huiting Zhang, Dongying Zhang, Yixian Wang, Ruiting Ding, Yuchuan Zheng, Yongxu Zhang
  • CSEE Journal of Power and Energy Systems
  • 2022-05-06
  • China's factor reallocation effect considering energy
  • Guangqing Xu, Xiaoyu Chen
  • Chinese Journal of Population, Resources and Environment
  • 2022-05-02
  • Cannabidiol prevents depressive-like behaviors through the modulation of neural stem cell differentiation
  • Ming Hou, Suji Wang, Dandan Yu, Xinyi Lu, Xiansen Zhao, Zhangpeng Chen, Chao Yan
  • Frontiers of Medicine
  • 2022-04-26
  • Cultivation of gut microorganisms of the marine ascidian Halocynthia roretzi reveals their potential roles in the environmental adaptation of their host
  • Yang Yang, Yuting Zhu, Haiming Liu, Jiankai Wei, Haiyan Yu, Bo Dong
  • Marine Life Science & Technology
  • 2022-04-26
  • Data network traffic analysis and optimization strategy of real-time power grid dynamic monitoring system for wide-frequency measurements
  • Jinsong Li, Hao Liu, Wenzhuo Li, Tianshu Bi, Mingyang Zhao
  • Global Energy Interconnection
  • 2022-04-25
  • Field distribution of the Z₂ topological edge state revealed by cathodoluminescence nanoscopy
  • Xiao He, Donglin Liu, Hongfei Wang, Liheng Zheng, Bo Xu, Biye Xie, Meiling Jiang, Zhixin Liu, Jin Zhang, Minghui Lu, Zheyu Fang
  • Opto-Electronic Advances
  • 2022-04-25
  • Advances in femtosecond laser direct writing of fiber Bragg gratings in multicore fibers: technology, sensor and laser applications
  • Alexey Wolf, Alexander Dostovalov, Kirill Bronnikov, Mikhail Skvortsov, Stefan Wabnitz, Sergey Babin
  • Opto-Electronic Advances
  • 2022-04-25
  • Graphene-empowered dynamic metasurfaces and metadevices
  • Chao Zeng, Hua Lu, Dong Mao, Yueqing Du, He Hua, Wei Zhao, Jianlin Zhao
  • Opto-Electronic Advances
  • 2022-04-25
  • Charge carrier dynamics in different crystal phases of CH₃NH₃PbI₃ perovskite
  • Efthymis Serpetzoglou, Ioannis Konidakis, George Kourmoulakis, Ioanna Demeridou, Konstantinos Chatzimanolis, Christos Zervos, George Kioseoglou, Emmanuel Kymakis, Emmanuel Stratakis
  • Opto-Electronic Science
  • 2022-04-21
  • Applications of optically and electrically driven nanoscale bowtie antennas
  • Zhongjun Jiang, Yingjian Liu, Liang Wang
  • Opto-Electronic Science
  • 2022-04-20
  • Validation of the bodily expressive action stimulus test among Chinese adults and children
  • Yunmei Yang, Wenwen Hou, Jing Li
  • PsyCh Journal
  • 2022-04-17

  • Context-Aware Candidates for Image Cropping                                Image Cropping Assisted By Modeling Inter-Patch Relations
    Copyright © PubCard