The Citizen Lab form Canada have posted an extensive series on investigating and exposing WeChat on Chinese internet and specifically WeChat. They found that WeChat uses a “MD5 hashing algorithm to hash images” so if you edit an image it produces a new image that needs a new hash. They are not currently implementing image recognition. Presumably will enable this im the future.
Tencent facilitates realtime filtering by maintaining a hash index populated by MD5 hashes of images sent by users of the chat platform. If the MD5 hash of an image sent over the chat platform is not in the hash index, then the image is not filtered. Instead, it is queued for automatic analysis. If it is found to be sensitive, then its MD5 hash is added to the hash index, and it will be filtered the next time a user attempts to send an image with the same hash.
Figure 24: A photo of primatologist Jane Goodall and infant chimpanzee Flint.
The types of images they found get censored is huge. Even this national geographic image above is censored on WeChat.
Wang Liming draws anti CCP illustrations. Many of his a re documented here. And many if not all of the are blocked in we chat as documented by citizenlab. These will serve as a library of images to test with.
For more images Pro Publica have documented a library of images they found censored on Weibo.
Citizenlab also document how far image censorship spreads on different parts of the platform. The censor much more aggressively the more public the image is posted.
I tried sending images from the Pro Publica study and neither of them were blocked. So I needed to find a way to test images.
Presumably due to it being in a private chat so it wasn’t censored, so I posted some images publicly and they still weren’t blocked. I’m posting the same images Pro Publica found censored on Weibo, on Weibo. Images was going to prove much more complicated than text.
The artist Tom White has documented his process of manipulating image recognition systems. He uses Google and Amazon’s image recognition API’s to test his work with. her experiment with how they interpret images compared to what we see in images.
