Privacy-Preserving Verification of ML Preprocessing via Model Behavior Indicators

Area of Science:

• Machine Learning
• Data Privacy
• Model Verification

Background:

• Ensuring data preprocessing integrity is crucial for machine learning model reliability, especially with sensitive data.
• Existing methods often require access to original data or labels, limiting their applicability in privacy-preserving scenarios.

Purpose of the Study:

• To introduce a novel privacy-preserving framework for verifying the correct application of data preprocessing pipelines.
• To enable model verification using only black-box access to the trained model, without original training data or ground-truth labels.

Main Methods:

• The framework combines three behavior indicators: prediction accuracy shifts, Kullback-Leibler (KL) divergence of output distributions, and explanation vectors (LIME/SHAP).
• It supports both binary correctness decisions and multi-class diagnosis of missing preprocessing steps.
• A label-free variant utilizes clustering of explanation vectors for verification.

Main Results:

• The binary detector achieved over 75% F1 score even under strong local differential privacy (ε=0.1).
• Machine-learning classifiers outperformed simple threshold rules for binary classification tasks.
• Comparable performance was observed between classifiers and threshold rules for multi-class diagnosis.

Conclusions:

• The proposed framework offers a practical and scalable solution for safeguarding preprocessing integrity in privacy-sensitive machine learning.
• The method effectively verifies preprocessing pipelines without compromising data privacy or requiring extensive data access.
• The label-free variant expands the applicability of the verification method to scenarios lacking labeled pipeline examples.