Adap CDP-ML: Concentrated Differentially Private machine learning with Adaptive Noise

Machine learning and big data, foundational pillars of artificial intelligence, are propelling the societal development, garnering significant attention. Recent studies reveal that malevolent actors can exploit vulnerabilities, leading to privacy breaches. These breaches occur through orchestrated privacy attacks on machine learning models, enabling the extraction of sensitive information from data via model parameters. Moreover, prevailing privacy-preserving machine learning approaches consume substantial privacy budgets, presenting a considerable challenge in balancing privacy guarantees and model utility. To tackle this challenge, our paper explores an adaptive differential privacy learning framework, integrating zero-concentration differential privacy techniques into machine learning to eliminate privacy threats and yield clearer quantitative results. Additionally, we propose a dynamic privacy budget allocation strategy to mitigate the high accumulation of the total privacy budget, providing enhanced privacy guarantees without compromising model utility. Theoretical analysis and extensive experiments on benchmark datasets validate that our approach effectively enhances model performance while reducing privacy loss.