Privacy-Preserving Image Template Sharing Using Contrastive Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Machine Learning as a Service (MLaaS) presents significant privacy risks due to potential data access by adversaries.
• Common attacks include reconstruction attacks and attribute inference attacks, aiming to reveal sensitive user information.
• Existing methods often struggle to balance utility and privacy effectively.

Purpose of the Study:

• To develop and evaluate two distinct training frameworks for image classification that preserve user data privacy.
• To mitigate risks associated with reconstruction and attribute inference attacks in MLaaS environments.
• To achieve a superior trade-off between data utility and privacy preservation.

Main Methods:

• Framework 1 (Reconstruction Attack): Employs supervised contrastive loss for feature discrimination and an obfuscator module for redundant information removal, jointly trained with a classifier.
• Framework 2 (Attribute Inference Attack): Utilizes supervised and private contrastive loss for attribute independence, with an adversarially trained obfuscator module.
• Both frameworks train an encoder with contrastive loss to improve the utility-privacy balance.

Main Results:

• The proposed frameworks demonstrate effectiveness in preserving user data privacy against both reconstruction and attribute inference attacks.
• The contrastive loss and obfuscator module contribute to minimizing information leakage while retaining classification performance.
• Experimental validation on the CelebA dataset confirms the efficacy of the developed privacy-preserving methods.

Conclusions:

• The study successfully presents two novel frameworks for privacy-preserving image classification in MLaaS.
• The proposed methods offer a robust solution for enhancing user data security against sophisticated adversarial attacks.
• The findings highlight the potential of contrastive learning and adversarial training for achieving strong privacy guarantees without compromising model utility.