中英对照：微信如何对图像文件进行监控？

（大中报/096.ca特稿）：微信使用两种方法对中国大陆用户的内容进行审查，即关键字和封堵网站。

WeChat uses two basic methods to implement its censorship on its Chinese Mainland users – the keyword list and a forbidden web site list for filtering and banning communications.

今天你也许关心的话题：

特鲁多的母亲和兄弟通过演讲从WE Charity拿了30万的报酬
7月10日周五中午12点疫情快报：6月份安省首次新增工作机会 全国失业率下降 省府倾向9月份让学生返校上课 加国未准备好应对第二波疫情
疫情期间在家工作 小心久坐有害健康

想要阅读与本文有关的话题？请点击本文末的链接！



监控图像内容是非常费力，费时且成本高。因此，微信通过技术俗称的散列索引的方式对每张图像进行审查。通过图像对比，微信可以提高对图像审查的速度。

It is very labour intensive, time-consuming and hence expensive for to analyze an image in real time to determine if it should be censored. WeChat would therefore subject every image to check against previously catalogued “Ban items” through an image's hash index – a form of simplified coding of the file in layman's term. Since this can be done quickly in real time, it is most efficient to find matches when compared to previous identified items.

当图像不匹配时，微信将采取“内容审查”措施。此过程是看被审查的图像是否在视觉上类似于任何已被列入黑名单的图像。同样，提取并分析图像内的文本也可以确认被审查的图像是否在被禁的名单上。如果发现图像属于被禁对象，微信就会把散列索引码添加到“被禁”名单中，从而实现对图像的即时审查。研究发现，微信从来没有实施图像的即时审查，一个第一次被上载微信的图像会被微信放行。

When an image does not have a match, it then undergo “Content surveillance”. This process will see if it is visually similar to any of the blacklisted images. As well, text inside the image is extracted and analyzed to determine if any of the text is blacklisted. If these findings are positive, then the hash code is added to the “Ban items lists” and allow for future real-time censorship. Previous it was found that content surveillance was never performed in real time and that the first time that a sensitive image file is transmitted it was not censored.



这就解释了为什么有些人曾尝试用新图像来对付微信的审查。有人将要发的内容用简单的摩尔斯码写在纸上，并将其扫描成图片后通过微信发出去。后来微信发现了这一奥秘，并通过人工智能把这条路堵死了。

This explains why some people tried to beat the system recently and was successful for a while until the system caught up to them. Apparently someone put simple Morse code messages onto a paper and then scanned it into a picture and transmitted that around as part of their messages. It worked for a while and then the system learned of this through its artificial intelligence capability and put a stop to this.

在对微信的测试中，计算机专家发现微信没有对海外的用户之间的内容进行审查。但这些海外账户仍被微信监控。测试证明，那些在海外用户之间流传的政治敏感内容会被用为对国内用户之间的审查，而这些内容从来就没有在中国用户之间流传过。

In controlled testing, computer experts did not detect censorship in communications among non-China-registered accounts. They have proof that such accounts are nevertheless subject to content surveillance. This was confirmed when politically sensitive content which was sent exclusively between non-China-registered accounts was identified as politically sensitive and subsequently censored when transmitted between China-registered accounts. These traced contents were not previously sent to, or transmitted between China-registered accounts.



测试人员设法了解微信是否使用散列索引对发往或来自中国用户的敏感文件进行监控和审查。他们将敏感文件发送到在国内的用户，发现诸如UTF8编码的纯文本（* .txt），Microsoft Word（* .docx）和可移植文档格式（* .pdf）之类的文档如包含某些敏感的关键字，例如“法轮功”和“法轮大法”就会被微信封杀。换句话说，国内用户根本就不知道有人给他们发这些文件。

The testers further explored whether sensitive documents sent to, or from, China-registered accounts were been monitored and censored using a hash index. By sending sensitive documents to a China-registered account, we could observe which files were censored. We found that documents such as UTF8-encoded plain text (*.txt), Microsoft Word (*.docx), and Portable Document Format (*.pdf) documents which contained certain sensitive keyword combinations such as “Falun Gong + Falun Dafa”were censored. In other words, these files never reached their destinations and displayed on Chinese-registered accounts.

通过发送两批文件，测试人员确认微信通过散列索引进行内容监控。微信要等到文件被列入散列索引后才会对内容进行即时审查。

By sending out the files in two different batches, testers confirmed that documents underwent file hash surveillance and that such files were not censored in real time until they had undergone non-real time content surveillance and their file hash had been added to the hash index.

通过这些检测，专家们试图了解微信进行审查的方法，以便用户可以用更好的方法来对抗微信的监控。我会在日后的文章中介绍对付微信监控的一些方法。

Through these experiments, experts seek to understand the methodology used by WeChat to conduct its censorship so users can devise a better means to combat this practice. I will further explore this in the coming weeks.