PreCurious: How Innocent Pre-Trained Language Models Turn into Privacy Traps.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Data Privacy

Background:

• The standard pre-training and fine-tuning paradigm is widely used for language models.
• Community platforms allow easy access to pre-trained models, but lack strict validation.
• Pre-trained models can pose privacy risks to fine-tuning datasets.

Purpose of the Study:

• Introduce the PreCurious framework to expose a new attack surface.
• Demonstrate how attackers can leverage pre-trained models for privacy attacks.
• Escalate privacy risks including membership inference and data extraction.

Main Methods:

• Propose the PreCurious framework for attacking fine-tuned models.
• Manipulate the pre-training memorization stage.
• Guide fine-tuning with deceptive configurations.

Main Results:

• PreCurious bypasses defenses of parameter-efficient and differentially private fine-tuning.
• The framework enables stealthy privacy attacks.
• Data extraction is possible even under strict differential privacy budgets.

Conclusions:

• Users must be cautious about downloading pre-trained models from untrusted sources.
• Common-sense defenses and tutorials may not be sufficient.
• Even sanitized datasets can be vulnerable to extraction attacks.