How to: beat Chinese social media image-filtering

Originally published at:


Pretty sure all of the above will fool Google images, too.


These are all tactics currently used by youtubers to get around auto detection.


I’m surprised anyone would be using an image-recognition algorithm that could be fooled by any 2D affine transformation and/or monotonic color transformation. The types of algorithm that are currently referred to as “machine learning” can already do much better than this, especially where you’re not overly concerned about false positives.

Still, any particular image-filtering algorithm is going to have ways it can be fooled, even if it takes special tools to do it, and if it’s a famous image (like the Tiananmen Square still), even a highly distorted copy will have the same effect since people know what the original looks like.

I wonder if the Great Firewall people considered better algorithms, and concluded that it wouldn’t stop anything that’s not being stopped already. And on a more paranoid level, I wonder if they do use better algorithms, not to censor images, but to detect people circumventing the weak censorship.


So, if it’s anything like most convolutional networks, you can also add in a single bright pixel? or properly chosen noise in the least significant bits?

I wouldn’t be a bit surprised if they started with inception v3, decapitated it, and trained for a few epochs on banned images.

Wait… am I reading this right… are they basically just looking for hashes of common images like tankman + hashes of various sizes? It seems so easy to circumvent that it’s not worth doing…

link has changed, probably to correct their spelling error on the word “analysis”


This topic was automatically closed after 5 days. New replies are no longer allowed.