Securing Self-supervised Data Curation for Foundation Models Robustness
Researchers have developed a Poisoned Data Detector (PDD) to ensure the integrity of datasets curated using self-supervised learning for foundation models. This defense mechanism combines the ImageBind model with traditional classifiers like SVM to identify and mitigate data poisoning risks. Evaluations showed SVM-PDD performed effectively across various datasets and adversarial attacks, demonstrating scalability and ensemble integration capabilities. AI
IMPACT Enhances the security and reliability of training data for large AI models, potentially improving their robustness against adversarial attacks.