Task-Specific Adaptive Differential Privacy Method for Structured Data.

Area of Science:

• Computer Science
• Machine Learning
• Data Privacy

Background:

• Machine learning (ML) models require data for training, often including sensitive private information.
• Current anonymization techniques used in ML are vulnerable to privacy attacks, failing to fully protect sensitive data.
• Existing methods struggle to balance privacy protection with data utility for ML tasks.

Purpose of the Study:

• To propose a novel task-specific adaptive differential privacy (DP) technique for structured data in privacy-preserving machine learning (PPML).
• To address the limitations of current anonymization methods by enhancing the protection of sensitive information in ML datasets.
• To resolve the privacy-utility trade-off inherent in protecting sensitive data during ML model training.

Main Methods:

• Developed a new DP method that adaptively calibrates the amount and distribution of random noise.
• Noise calibration is based on attribute feature importance specific to ML tasks and data types.
• The technique is designed to be model-agnostic, applicable across various ML tasks and data structures.

Main Results:

• Experimental results demonstrate the effectiveness of the proposed task-specific adaptive DP method.
• The technique successfully balances privacy preservation with data utility, mitigating the privacy-utility trade-off.
• Evaluated performance across diverse datasets, ML tasks, and different DP mechanisms.

Conclusions:

• The proposed task-specific adaptive DP technique offers a robust solution for protecting sensitive data in ML.
• This method ensures privacy preservation while maintaining the utility of the data for ML models.
• The model-agnostic nature allows for broad applicability in various ML scenarios and data types.