Bidirectional Texts – An Easy Hack to Escape the Hawk-Eye of Profanity Filters
Bidirectional Texts – An Easy Hack to Escape the Hawk-Eye of Profanity Filters
September 30, 2018

Profanity filters, or swear filters, as they are known have indeed done a reasonable job in keeping out profane content from the vicinity of good, meaningful content.


Whether it’s a forged image or an offensive string of text, they can identify them all at the blink of an eye! Moreover, with the advent of AI, things have become really easy for content moderators. Simply provide an input to your API, and the offensive piece will be replaced with something like ‘******’.


Pretty easy, isn’t it?


But it is NOT, because here comes the loophole!


The approach used by such filters of dumbly looking for strings of the text leads to a range of problems, including false positives (known colloquially as the Scunthorpe Problem).


In recent times, mischief-mongers amongst the online community have come up with an easy hack to bypass the profanity filters i.e. bidirectional texts.


There might be instances when source codes in a page cause the direction of an offensive text to be reversed. For example, IDIOT will be written as TOIDI.


And, there have been many cases where such texts have been successful in evading the profanity filters.


So, What Can Be Done?


A problem is not a problem if it does not have a solution!


So, there are two ways out of this situation – the first one is from a technical perspective while the latter one is a result of a scientific study done to co-relate profanity filters with deep learning.


  1. Mitigation


Characters can be filtered out from the source code that causes a reversal in the direction of such texts.


For this, your site should have a zero tolerance policy against offensive content.


And of course, if you run a site which has people writing in Arabic, Persian, Urdu, Hebrew, etc. this may not be practical.


  1. Deep Learning for Detecting Inappropriate Content in Text


A recent study has outlined the need for deep learning for preventing the penetration of inappropriate content into the online community.


This study took into account two categories of content in which profanity was likely to be found – Query completion suggestions in Search Engines (SEs) and user conversations in messengers.


For query completion suggestions, the researchers combined the strengths of both Convolutional Neural Networks (CNN) and Bi-directional LSTM (BLSTM) deep learning architectures and propose a novel architecture called “Convolutional, Bi-Directional LSTM (C-BiLSTM)” for automatic inappropriate query detection.


The models developed by them outperformed both pattern-based and other hand-crafted feature-based content moderation techniques.


A Final Word on This –


There is no dearth of individuals who are desperate to create trouble and bully the netizens. However, giving a befitting reply to such people becomes necessary if we are to survive this never-ending circus of profanity.


So, the next time you come across a bidirectional text you know exactly what to do!



Share this article