Neural Honeypoint: An Active Defense Framework Against Model Inversion Attacks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning Security

Background:

• Machine learning systems are vulnerable to model inversion attacks (MIAs), risking private training data exposure.
• Current MIA defenses often hinder model usability and fail to detect persistent attack attempts.
• Existing defenses focus on increasing attack overhead, not actively capturing malicious queries.

Purpose of the Study:

• To introduce an active defense framework, Neural Honeypoint, to counter model inversion attacks (MIAs).
• To address the limitations of existing defenses by enabling detection and blocking of inversion queries.
• To maintain model utility while enhancing data privacy against sophisticated attacks.

Main Methods:

• Neural Honeypoint models attacker capabilities in the frequency domain to design specialized honeypoints.
• Honeypoints are integrated into the protected model using a backdoor-like fine-tuning process.
• Attack detection is achieved by comparing input feature similarity against deployed honeypoints.

Main Results:

• Neural Honeypoint significantly reduces the attack success rate (ASR) of advanced MIAs to 0%-2%.
• The framework effectively captures inversion queries, enabling timely detection and blocking of attacks.
• Experimental results demonstrate the efficacy of active defense in safeguarding machine learning models.

Conclusions:

• Neural Honeypoint offers a robust active defense against model inversion attacks, enhancing data privacy.
• The proposed method successfully bridges the gap in detecting and mitigating persistent MIA threats.
• This framework provides a practical solution for securing machine learning models without compromising usability.