An analysis of automatic image filtering on WeChat Moments

We report results from a series of experiments that uncover mechanisms used to filter images on WeChat, the most popular social media platform in China. Our results inform strategies for evading image filtering on the application. By performing tests on a collection of politically sensitive images filtered by WeChat, we found that WeChat uses two different algorithms to filter, an Optical Character Recognition (OCR)-based algorithm that filters images containing sensitive text, and a visual-based algorithm that filters images that are visually similar to those on an image blacklist. The OCR-based algorithm has implementation similarities to many common OCR algorithms that allow us to create text images that evade filtering. We found that the visual-based algorithm does not use any machine learning approach that uses high level classification of an image to determine whether it is sensitive; however, we discovered multiple implementation details of the visual-based algorithm that inform the creation of images that are visually similar to those blacklisted but that evade filtering. This study is the first in-depth technical analysis of image filtering on WeChat, and we hope that our methods will serve as a road map for studying image censorship on other platforms.

[1]  Jeffrey Knockel,et al.  Chat program censorship and surveillance in China: Tracking TOM-Skype and Sina UC , 2013, First Monday.

[2]  Jeffrey Knockel,et al.  Three Researchers, Five Conjectures: An Empirical Analysis of TOM-Skype Censorship and Surveillance , 2011, FOCI.

[3]  Grant Potter,et al.  China's "Networked Authoritarianism" , 2018 .

[4]  Michael Chau,et al.  Assessing Censorship on Microblogs in China: Discriminatory Keyword Analysis and the Real-Name Registration Policy , 2013, IEEE Internet Computing.

[5]  Jeffrey Knockel,et al.  Measuring Decentralization of Chinese Keyword Censorship via Mobile Games , 2017, FOCI @ USENIX Security Symposium.

[6]  Joan Bruna,et al.  Intriguing properties of neural networks , 2013, ICLR.

[7]  Jeffrey Knockel,et al.  Every Rose Has Its Thorn: Censorship and Surveillance on Social Video Platforms in China , 2015 .

[8]  Logan Engstrom,et al.  Black-box Adversarial Attacks with Limited Queries and Information , 2018, ICML.

[9]  Jeffrey Knockel,et al.  One App, Two Systems: How WeChat uses one censorship policy in China and another internationally , 2016 .

[10]  N. Otsu.  A threshold selection method from gray level histograms , 1979 .

[11]  C. Cairns,et al.  Why autocrats sometimes relax online censorship of sensitive issues: A case study of microblog discussion of air pollution in China , 2016 .

[12]  Brendan T. O'Connor,et al.  Censorship and deletion practices in Chinese social media , 2012, First Monday.

[13]  Jonathon Shlens,et al.  Explaining and Harnessing Adversarial Examples , 2014, ICLR.

[14]  Ramarathnam Venkatesan,et al.  Robust perceptual image hashing via matrix invariants , 2004, 2004 International Conference on Image Processing, 2004. ICIP '04..

[15]  Dan S. Wallach,et al.  The Velocity of Censorship: High-Fidelity Detection of Microblog Post Deletions , 2013, USENIX Security Symposium.

[16]  Roberto Brunelli,et al.  Template Matching Techniques in Computer Vision: Theory and Practice , 2009 .

[17]  Jeffrey Knockel,et al.  Remembering Liu Xiaobo: analyzing censorship of the death of Liu Xiaobo on WeChat and Weibo , 2017 .

[18]  Jinfeng Yi,et al.  ZOO: Zeroth Order Optimization Based Black-box Attacks to Deep Neural Networks without Training Substitute Models , 2017, AISec@CCS.

[19]  Margaret E. Roberts,et al.  How Censorship in China Allows Government Criticism but Silences Collective Expression , 2013, American Political Science Review.

[20]  Nina Narodytska,et al.  Simple Black-Box Adversarial Perturbations for Deep Networks , 2016, ArXiv.