Adaptive Privacy Based on Mutual Information for Machine Learning in Edge–Cloud Environments

Edge-cloud computing demands privacy-preserving techniques that balance data utility and computational efficiency. Traditional methods often degrade machine learning (ML) performance by applying uniform noise across all features. This work introduces Mutual Information Adaptive Differential Privacy (MIADP), an adaptive strategy that allocates privacy budgets based on feature importance and applies correlation-aware noise to preserve key statistical relationships. Experiments show that MIADP maintains strong ML performance while ensuring efficiency suitable for edge deployment.