LFighter: Defending against the label-flipping attack in federated learning

Area of Science:

• Machine Learning
• Cybersecurity
• Distributed Systems

Background:

• Federated learning (FL) enables collaborative model training while preserving data privacy.
• Malicious participants can compromise FL models through poisoning attacks, such as label-flipping (LF).
• Existing LF defenses face limitations regarding data distribution assumptions and high-dimensional models.

Purpose of the Study:

• To investigate the behavior of label-flipping attacks in federated learning.
• To propose a novel defense mechanism, LFighter, against label-flipping attacks.
• To evaluate the effectiveness and superiority of LFighter compared to existing defenses.

Main Methods:

• Investigated the impact of LF attacks on parameter gradients of source and target classes.
• Developed LFighter to dynamically extract, cluster, and analyze gradients from local updates.
• Filtered out malicious updates before global model aggregation.

Main Results:

• LFighter effectively detects label-flipping attacks by leveraging gradient features.
• The defense demonstrates robustness across different data distributions and model dimensionalities.
• LFighter significantly outperforms state-of-the-art defenses in accuracy, error rates, and stability.

Conclusions:

• Parameter gradients offer discriminative features for detecting label-flipping attacks in FL.
• LFighter provides an effective and robust defense against label-flipping attacks.
• The proposed method enhances the security and reliability of federated learning systems.