Mastering Machine Learning for Penetration Testing

Evasion attacks

To perform machine learning evasion attacks, cyber criminals try to bypass the learning outcomes by observing how the model works, especially the outcome, by trying many different samples simply by feeding the model with different inputs and trying to find the learning patterns. This technique is very popular. For example, if an attacker wants to bypass a machine learning spam filter, he needs to feed the system with different emails and search for a pattern that makes a spam email goes through (Not detected as a spam email) and bypass detection by doing only a few modifications to previously detected emails.

The following workflow illustrates how an evasion attack works: