A System-Driven Taxonomy of Attacks and Defenses in Adversarial Machine Learning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning Security

Background:

• Machine Learning (ML) and Computational Intelligence (CI) power critical applications like autonomous vehicles and biometrics.
• Adversarial attacks on ML systems can cause misclassifications, leading to unreliable operations and erroneous decisions.
• Developing robust ML systems resistant to these attacks is a priority in adversarial machine learning.

Purpose of the Study:

• To propose a fine-grained, system-driven taxonomy for specifying ML applications and adversarial system models.
• To enable unambiguous experimental replication by independent researchers.
• To facilitate the advancement of robust ML through a structured understanding of the attack-defense landscape.

Main Methods:

• Developed a taxonomy covering datasets, ML architectures, adversary attributes (knowledge, capability, goal), adversary strategies, and defense responses.
• Analyzed the interrelationships between these taxonomic elements.
• Proposed an adversarial machine learning cycle to model the interaction between ML systems and adversaries.

Main Results:

• A comprehensive system-driven taxonomy integrating all proposed elements was created.
• The taxonomy provides a standardized framework for representing the adversarial ML landscape.
• It facilitates the evaluation and comparison of research contributions and identifies research gaps.

Conclusions:

• The proposed taxonomy enhances the reproducibility of adversarial ML research.
• It serves as a valuable tool for researchers to understand, categorize, and advance the field of robust ML.
• This structured approach is crucial for escalating the arms race towards more evolved and reliable ML applications.