Differential Privacy for Classifier Evaluation.

Area of Science:

• Computer Science
• Machine Learning
• Privacy-Preserving Technologies

Background:

• Differential privacy (DP) offers strong data protection guarantees.
• Existing DP methods primarily focus on model training, not performance evaluation.
• Reporting ML model performance metrics can inadvertently disclose private information.

Purpose of the Study:

• To develop differentially private mechanisms for computing ML evaluation metrics.
• To address the privacy risks associated with reporting model performance.
• To enable secure and private evaluation of machine learning models.

Main Methods:

• Investigated differentially private algorithms for calculating key performance metrics.
• Focused on metrics such as Area Under the Receiver Operating Characteristic (ROC) Curve.
• Developed and analyzed novel privacy-preserving computation techniques.

Main Results:

• Successfully identified effective differentially private mechanisms for ROC AUC calculation.
• Demonstrated the feasibility of computing average precision with differential privacy.
• Quantified the privacy-utility trade-offs for the proposed methods.

Conclusions:

• Differentially private computation of ML evaluation metrics is achievable and necessary.
• The developed methods enable private reporting of model performance.
• This work advances the field of privacy-preserving machine learning evaluation.