Automatic Transformation Search Against Deep Leakage From Gradients

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Data Privacy

Background:

• Collaborative learning enables joint Deep Learning model training without data sharing, offering privacy benefits.
• However, reconstruction attacks can extract sensitive training data from shared gradients, threatening collaborative learning security.
• Effective mitigation strategies are crucial to address these privacy vulnerabilities.

Purpose of the Study:

• To systematically analyze existing reconstruction attacks in collaborative learning.
• To propose and validate a novel defense mechanism using data augmentation to mitigate these attacks.
• To develop metrics and a search method for optimal data augmentation policies.

Main Methods:

• Leveraging data augmentation by applying carefully selected image transformations to sensitive data.
• Designing two novel metrics to quantify the trade-offs between data privacy and model usability.
• Developing an automated search method to discover effective data augmentation policies from a library.
• Integrating the defense method into existing collaborative training systems without protocol modification.

Main Results:

• The proposed data augmentation policies effectively prevent adversaries from reconstructing sensitive training samples from gradients.
• The discovered policies demonstrate high efficiency in defeating state-of-the-art reconstruction attacks.
• The defense method shows negligible impact on the overall model performance.
• Comprehensive experiments validate the effectiveness across various system settings.

Conclusions:

• Data augmentation serves as a robust defense against gradient-based reconstruction attacks in collaborative learning.
• The developed metrics and search method enable the automated discovery of privacy-preserving augmentation policies.
• This approach enhances the security of collaborative learning without compromising model utility or requiring system modifications.