Why Do Unlearnable Examples Work: A Novel Perspective of Mutual Information

Researchers propose a new method called Mutual Information Unlearnable Examples (MI-UE) to protect data privacy by preventing unauthorized AI models from learning from scraped data. The approach uses mutual information theory to create more effective data poisoning techniques that impede deep learning model generalization.

• →New MI-UE method outperforms existing data protection techniques by reducing mutual information between clean and poisoned features.
• →Researchers prove that minimizing conditional covariance of intra-class poisoned features effectively reduces mutual information between distributions.
• →The approach maximizes cosine similarity among intra-class features to impede AI model generalization.
• →Method remains effective even against defense mechanisms designed to counter data poisoning.
• →Provides theoretical foundation for unlearnable examples rather than relying on empirical heuristics.