Adversarial Feature Selection Against Evasion Attacks.

Area of Science:

• Computer Science
• Machine Learning
• Cybersecurity

Background:

• Machine learning is used in adversarial settings like spam and malware detection.
• Security against evasion attacks, where data is manipulated at test time, is not fully understood.
• Feature selection may negatively impact classifier security against evasion.

Purpose of the Study:

• Investigate the impact of feature selection on classifier security against evasion attacks.
• Propose a novel adversary-aware feature selection model to enhance security.
• Improve the robustness of machine learning models in adversarial environments.

Main Methods:

• Developed a novel adversary-aware feature selection model.
• Incorporated assumptions on the adversary's data manipulation strategy.
• Implemented an efficient, wrapper-based approach for feature selection.

Main Results:

• Classifier security can be worsened by standard feature selection.
• The proposed adversary-aware feature selection model improves classifier security.
• Experimental validation on spam and malware detection confirms the model's effectiveness.

Conclusions:

• Feature selection's impact on adversarial robustness requires careful consideration.
• Adversary-aware feature selection offers a promising direction for enhancing machine learning security.
• The proposed model effectively improves detection rates against evasion attacks.