This research presents a novel heuristic for estimating privacy leakage in differentially private machine learning, focusing on the last iterate of stochastic gradient descent.

• Introduces a simple linear heuristic to predict privacy leakage when only the final model is released
• Experimentally validates the heuristic through privacy auditing across various training procedures
• Provides a practical way to estimate privacy risks before training begins
• Demonstrates the security advantages of releasing only the last iterate rather than all intermediate models

For security professionals, this research offers a practical tool to assess privacy vulnerabilities in machine learning deployments without expensive post-training audits, helping organizations better protect sensitive data used in AI development.

The Last Iterate Advantage: Empirical Auditing and Principled Heuristic Analysis of Differentially Private SGD