MSG: Stealing data from pruned neural networks via malicious sparsity guidance

Area of Science:

• Machine Learning Security
• Data Privacy
• Adversarial Machine Learning

Background:

• Machine learning cloud platforms raise privacy concerns.
• Correlation Value Encoding Attacks (CVEA) embed training data into model parameters.
• Model pruning techniques can reduce the effectiveness of such attacks.

Purpose of the Study:

• To demonstrate the impact of model pruning on CVEA.
• To propose a novel, pruning-resistant data encoding attack.
• To enhance the stealthiness of data extraction attacks.

Main Methods:

• Empirical evaluation of common model pruning techniques against CVEA.
• Development of Malicious Sparsity Guidance (MSG) for data embedding.
• Integration of knowledge transfer to maintain model accuracy post-pruning.

Main Results:

• Model pruning significantly reduces CVEA effectiveness.
• MSG successfully embeds data and resists pruning.
• MSG-encoded models maintain high accuracy after pruning due to knowledge transfer.

Conclusions:

• MSG offers a robust method for data extraction attacks, even against pruned models.
• Knowledge transfer enhances the inconspicuousness of data extraction.
• This research highlights new vulnerabilities in secure model deployment.