A Dataset Auditing Method for Collaboratively Trained Machine Learning Models.

Area of Science:

• Machine Learning
• Data Privacy
• Federated Learning

Background:

• Dataset auditing is crucial for verifying data usage in machine learning (ML) models.
• Federated Learning (FL) presents unique challenges for dataset auditing due to decentralized private data.
• Auditing facilitates regulatory compliance and user data control in FL.

Purpose of the Study:

• To establish requirements for practical dataset auditing methods in FL.
• To introduce Ensembled Membership Auditing (EMA), a novel auditing technique.
• To evaluate EMA's effectiveness compared to state-of-the-art methods.

Main Methods:

• Leveraging existing Membership Inference Attack (MIA) techniques.
• Aggregating data-wise membership scores using statistical testing.
• Experimental evaluation on benchmark, X-ray, and dermatology datasets.

Main Results:

• EMA substantially meets the proposed requirements for dataset auditing.
• The method demonstrates superior performance compared to prior state-of-the-art approaches.
• Validation across diverse medical imaging datasets confirms EMA's robustness.

Conclusions:

• EMA offers a practical and effective solution for dataset auditing in Federated Learning.
• The proposed method enhances privacy preservation and regulatory adherence in collaborative ML.
• EMA provides a significant advancement in auditing techniques for sensitive data.