Devising Mutation-Based Attacks to Evade ML-Based Text Understanding Models

This article discusses the problem of bypassing machine learning (ML) based text understanding models in web application firewalls (WAFs). Existing WAFs rely on ML algorithms to detect and prevent SQL injection attacks, but attackers can easily bypass these models by making slight changes to the input data. The article proposes a novel approach to generate SQLi payloads that can bypass these ML-based text understanding models while preserving their maliciousness.
The proposed method involves creating semantic replacements for words in the input data using a context-free grammar. This allows attackers to modify the input data without changing its meaning, making it difficult for ML algorithms to detect any differences. The article demonstrates the effectiveness of this approach by training three machine learning models on two datasets and showing that they can bypass existing WAFs with high accuracy.
The authors also discuss the limitations of their approach and suggest future research directions to improve the effectiveness of WAFs against SQLi attacks. They emphasize that while their method can bypass ML-based text understanding models, it is not a foolproof solution and other security measures should still be used in conjunction with WAFs.
Overall, this article provides valuable insights into the challenges of detecting SQL injection attacks using ML-based text understanding models and proposes a novel approach to bypass these models while preserving their maliciousness. The authors’ findings have important implications for the security of web applications and highlight the need for more effective security measures against SQLi attacks.