Rethinking Label Flipping Attack: From Sample Masking to Sample Thresholding.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Deep Learning

Background:

• Machine learning (ML) and deep learning (DL) are foundational AI technologies.
• These methods are vulnerable to adversarial attacks, posing significant security risks.
• Label Flipping Attacks (LFA) corrupt models by altering training data labels.

Purpose of the Study:

• To address the scalability limitations of existing LFA methods for deep learning.
• To propose an efficient and scalable algorithm for label flipping attacks.
• To analyze the effectiveness of surrogate models in adversarial attack scenarios.

Main Methods:

• Reformulation of the Label Flipping Attack as a novel minimax problem.
• Development of the Sample Thresholding algorithm for efficient sample selection.
• Theoretical analysis of surrogate model paradigms for unpredictable victim models.
• Extension of the method to crowdsourced ranking tasks.

Main Results:

• Sample Thresholding enables efficient label flipping attacks on deep learning models.
• The proposed method is scalable and applicable to large datasets.
• Theoretical analysis demonstrates the small performance gap with surrogate models.
• Experimental validation on real-world datasets confirms the method's efficacy.

Conclusions:

• Sample Thresholding offers an effective and scalable approach to adversarial label flipping attacks.
• The method is robust and adaptable, even in the presence of surrogate models.
• This research advances the understanding and mitigation of security threats in ML/DL.