As AI systems increasingly rely on public data sources, FLORAL emerges as a powerful defense mechanism against adversaries who manipulate training labels to compromise model integrity.

• Introduces a support vector-based adversarial training strategy specifically designed to protect against label poisoning
• Addresses a critical security vulnerability in large language models that use human-annotated labels
• Offers a practical solution for maintaining model performance even when training data has been tampered with
• Particularly valuable for high-stakes AI deployments where model reliability is essential

This research provides security teams with a concrete method to fortify AI systems against an emerging threat vector, ensuring models remain trustworthy even when training on potentially compromised datasets.

Adversarial Training for Defense Against Label Poisoning Attacks