Dataset Security for Machine Learning: Data Poisoning, Backdoor Attacks, and Defenses.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning Security

Background:

• Machine learning (ML) systems demand vast datasets for optimal performance.
• Automated and outsourced data curation are common but introduce security risks.
• Lack of human oversight in data collection can lead to model manipulation.

Purpose of the Study:

• To systematically categorize dataset vulnerabilities and exploits in machine learning.
• To discuss existing and novel approaches for defending against data-driven attacks.
• To identify and highlight open research problems in dataset security.

Main Methods:

• Literature review and synthesis of existing research on data poisoning and adversarial attacks.
• Categorization framework development for dataset vulnerabilities.
• Analysis of defense mechanisms against data manipulation.

Main Results:

• Identified a taxonomy of dataset vulnerabilities, including data poisoning, backdoor attacks, and data integrity issues.
• Evaluated the effectiveness of various defense strategies, such as data sanitization and robust training methods.
• Highlighted the limitations of current defenses and the need for more resilient systems.

Conclusions:

• Dataset security is a critical, yet often overlooked, aspect of machine learning.
• Proactive defense mechanisms and robust data governance are essential for trustworthy AI.
• Further research is needed to develop comprehensive solutions for securing ML training data.