CANCELLED • Seminar • Algorithms and Complexity • On Personalized Differential Privacy

Please note: This seminar has been CANCELLED.

Juba Ziani, Assistant Professor
H. Milton Stewart School of Industrial and Systems Engineering, Georgia Tech

In this talk, I will be discussing “personalized” (or “individualized”) differential privacy, where different individuals can be offered different epsilons simultaneously within the same computation. I will be presenting two of my recent works on personalized DP in the central model:

1) In the first part of the talk, I will take a platform design view of privacy. I will study the problem faced by a data analyst or platform that wishes to collect private data from privacy-aware agents. To incentivize participation, in exchange for this data, the platform provides a service to the agents in the form of a statistic computed using all agents’ submitted data. The agents decide whether to join the platform and reveal their data by considering both the privacy costs of joining and the benefit they get from obtaining the statistic. The platform must choose what level of personalized DP to guarantee as a function of the users’ heterogeneous privacy preferences, and do so to optimize the privacy of its service or computed statistic. This is based on “Optimal Data Acquisition with Privacy-Aware Agents” with Rachel Cummings, Hadi Elzayn, Vasilis Gkatzelis, and Manolis Pountourakis that won best paper at SATML 2023.

2) The previous work considers personalized differential privacy for a simple mean estimation task. In the second part of the talk, I will focus on initial work towards extending personalized differential privacy techniques to loss minimization. In particular, I will focus on the problem of Ridge regression. I will show how output perturbation techniques for private loss minimization extend to the personalized privacy case. I will also highlight limitations, showing a need to understand how objective perturbation and gradient perturbation techniques can be adapted to the personalized DP case for loss minimization tasks. This is based on “Personalized Differential Privacy for Ridge Regression” with Krishna Acharya, Franziska Boenisch, and Rakshit Naidu.

If we have time left, I will also aim to highlight current limitations, difficulties, and impossibilities faced by personalized differential privacy.