Privacy-Preserving Verification of ML Preprocessing via Model Behavior Indicators

We present a privacy-preserving framework to verify whether a declared data preprocessing pipeline was correctly applied before training a machine learning model on sensitive data. The verifier has only black-box query access to the model and combines three behavior indicators: shift in prediction accuracy, Kullback-Leibler (KL) divergence between output distributions, and explanation vectors from Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP). The method requires neither the original training records nor ground-truth labels. It supports two tasks: (i) a binary decision on correctness and (ii) a multi-class diagnosis identifying which step is missing. Experiments on three tabular datasets (Diabetes, Adult-Income, Student-Record) show that the binary detector maintains over 75% F1 even under strong local differential privacy ( $\epsilon =0.1$ ). Machine-learning classifiers consistently outperform simple threshold rules in the binary setting, while the two approaches perform comparably for multi-class diagnosis. A label-free variant that clusters explanation vectors achieves competitive accuracy, enabling verification when no labeled pipelines are available. These results demonstrate a practical and scalable approach for safeguarding preprocessing integrity in privacy-sensitive machine learning workflows.