This page organizes papers related to artificial intelligence published around the world. This page is summarized using Google Gemini and is operated on a non-profit basis. The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.
Gaussian DP for Reporting Differential Privacy Guarantees in Machine Learning
Created by
Haebom
Author
Juan Felipe Gomez, Bogdan Kulynych, Georgios Kaissis, Flavio P. Calmon, Jamie Hayes, Borja Balle, Antti Honkela
Outline
This paper identifies problems with reporting the protection level of differential privacy (DP) machine learning algorithms and proposes a solution using non-asymptotic Gaussian Differential Privacy (GDP). Using numerical accounting tools and decision-theoretic metrics, we demonstrate that GDP can accurately represent the full privacy profile of algorithms like DP-SGD. We verify the suitability of GDP by analyzing the privacy profiles of state-of-the-art DP image classification and the TopDown algorithm for the US Census, and discuss the strengths and weaknesses of the GDP approach.
Takeaways, Limitations
•
Takeaways:
◦
GDP can be used to convey DP guarantees more accurately and completely.
◦
It can capture the full privacy profile of DP-SGD and related algorithms with almost no error.
◦
Effectively analyze and understand the privacy profiles of cutting-edge DP ML algorithms.
•
Limitations:
◦
Further research is needed on the strengths and weaknesses of the GDP approach.
◦
There is a need to explore other privacy mechanisms to which GDP can be applied.
◦
Further consideration is needed for specific implementation and practical application.
