Source: Quartz (8/15/18)
Researchers have figured out ways to dodge censorship on WeChat
By Echo Huang
WeChat users in China have come up with creative ways to circumvent censorship, and one of the more effective methods they’ve discovered seems to be sharing images instead of text, which can be easily caught by censors. In the case of China’s #MeToo movement, which authorities tried to shut down, social-media users decided to share a university student’s censored letter by posting images of it upside down in hopes of dodging the country’s filters.
It’s an ongoing mystery how censorship works on WeChat, which appears to affect only those accounts that are linked to mainland phone numbers. But new research from Citizen Lab, a research group at the University of Toronto, is offering some clues on getting around it.
In previous studies, the team conducted tests in one-to-one and group chats, but this new research is focused on WeChat Moments, a feature similar to Facebook’s timeline. Their work helped them identify two algorithms used in WeChat Moments to filter images. The first uses optical-character recognition to identify sensitive text in images. The other is a visual-based approach, which basically compares an uploaded picture against blacklisted images. Tencent didn’t immediately respond to questions about its censorship methods.
Their tests included images of the Nobel laureate Liu Xiaobo, whose photo and likeness were heavily censored after his death last summer, and a graphic of China’s flag laid over Hong Kong’s, a symbol of Beijing’s encroachment of the city.
After trying out various methods, researchers found it was possible to evade censorship by distorting sensitive images in some way. Here are the methods they outlined (emphasis ours):
By mirroring or rotating the image, since the filter has no high level semantic understanding of uploaded images. However, many images lose meaning when mirrored or rotated, particularly images that contain text which may be rendered illegible.
By changing the aspect ratio of an image, such as by stretching the image wider or taller. However, this may make objects in images look out of proportion.
By blurring the photo, since edges appear important to the filter. However, while edges are important to WeChat’s filter, they are often perceptually important for people too.
By adding a sufficiently large border to the smallest dimensions of an image, or to both the smallest and largest dimensions, particularly if both dimensions are of equal or similar size.
By adding a large border to the largest dimensions of an image and adding a sufficiently complex pattern to it.
In the case of the last method, the researchers found that creating a complex pattern like duplicating an image multiple times in a single photo (see above) could bypass the filter, likely because the algorithm is looking at images on the whole instead of analyzing specific elements.
The researchers were also able to post 15 mirrored images that had been censored in their previous studies. Their successful maneuvering likely indicates that “no sort of high-level machine learning classification system is being used to trigger the observed filtering on WeChat,” they write in the report.
The relatively crude methods the researchers used to bypass WeChat’s filters suggest that image censorship—in its current form at least—is not all that sophisticated. Tencent, the tech titan and parent company of WeChat, doesn’t appear to be using computer vision to suss out sensitive elements from images. Instead, the researchers think it’s more likely the company maintains a specific blacklist of images, which government officials can add to. Of course, Chinese censorship has evolved as more people learn how to get around it, and it could be just a matter of time before these workarounds no longer work.