Learning a robust foundation model against clean-label data poisoning attacks at downstream tasks.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Cybersecurity

Background:

• Transfer learning utilizes pre-trained models for downstream tasks, creating vulnerabilities to data poisoning.
• Attackers inject malicious data into re-training sets to manipulate model behavior.
• Existing defenses struggle against sophisticated attacks like clean-label poisoning.

Purpose of the Study:

• To develop a robust defense strategy against data poisoning attacks in transfer learning.
• To enhance the security of foundation models used in downstream applications.
• To reduce the success rate of adversarial manipulations in machine learning pipelines.

Main Methods:

• Proposing a defense strategy focused on pre-training robust foundation models.
• Implementing techniques to reduce adversarial feature distance.
• Implementing techniques to increase inter-class feature distance.

Main Results:

• The proposed strategy significantly reduces the success rate of data poisoning attacks.
• Demonstrated excellent defense performance against state-of-the-art clean-label poisoning attacks.
• Validated effectiveness within the transfer learning scenario.

Conclusions:

• The developed defense strategy offers effective protection against data poisoning in transfer learning.
• Adjusting feature distances is a viable method for building robust foundation models.
• This work contributes to securing machine learning models in real-world applications.