Apple believes that privateness is a basic human proper. As AI experiences turn out to be more and more private and part of folks’s each day lives, it is essential that novel privacy-preserving methods are created in parallel to advancing AI capabilities.
Apple’s basic analysis has constantly pushed the state-of-the-art in utilizing differential privateness with machine studying, and earlier this yr, we hosted the Workshop on Privateness-Preserving Machine Studying (PPML). This two-day hybrid occasion introduced collectively Apple and members of the broader analysis group to debate the cutting-edge in PPML, specializing in 4 key areas: Non-public Studying and Statistics, Assaults and Safety, Differential Privateness Foundations, and Basis Fashions and Privateness.
The shows and discussions of those matters explored the intersection of privateness, safety, and the quickly evolving panorama of synthetic intelligence. Workshop members mentioned the theoretical underpinnings and sensible challenges of constructing AI programs that shield privateness. By addressing privateness and safety issues from each theoretical and sensible views, we intention to foster innovation whereas safeguarding person privateness.
On this put up, we share recordings of chosen talks and a recap of the publications mentioned on the workshop.
Apple Workshop on Privateness-Preserving Machine Studying 2025 Movies
Printed Work Introduced on the Workshop
AirGapAgent: Defending Privateness-Aware Conversational Brokers by Eugene Bagdasarian (Google Analysis), Peter Kairouz (Google Analysis), Ren Yi (Google Analysis), Marco Gruteser (Google Analysis), Sahra Ghalebikesabi (Google DeepMind), Sewoong Oh (Google Analysis), Borja Balle (Google DeepMind), and Daniel Ramage (Google Analysis)
A Generalized Binary Tree Mechanism for Differentially Non-public Approximation of All-Pair Distances by Michael Dinitz (Johns Hopkins College), Chenglin Fan (Seoul Nationwide College), Jingcheng Liu (Nanjing College), Jalaj Upadhyay (Rutgers College), and Zongrui Zou (Nanjing College)
Differentially Non-public Artificial Information through Basis Mannequin APIs 1: Photographs by Zinan Lin (Microsoft Analysis), Sivakanth Gopi (Microsoft Analysis), Janardhan Kulkarni (Microsoft Analysis), Harsha Nori (Microsoft Analysis), and Sergey Yekhanin (Microsoft Analysis)
Differentially Non-public Artificial Information through Basis Mannequin APIs 2: Textual content by Chulin Xie (College of Illinois Urbana-Champaign), Zinan Lin (Microsoft Analysis), Arturs Backurs (Microsoft Analysis), Sivakanth Gopi (Microsoft Analysis), Da Yu (Solar Yat-sen College), Huseyin Inan (Microsoft Analysis), Harsha Nori (Microsoft Analysis), Haotian Jiang (Microsoft Analysis), Huishuai Zhang (Microsoft Analysis), Yin Tat Lee (Microsoft Analysis), Bo Li (College of Illinois Urbana-Champaign, College of Chicago), and Sergey Yekhanin (Microsoft Analysis)
Environment friendly and Close to-Optimum Noise Technology for Streaming Differential Privateness by Krishnamurthy (Dj) Dvijotham (Google DeepMind), H. Brendan McMahan (Google Analysis), Krishna Pillutla (ITT Madras), Thomas Steinke (Google DeepMind), and Abhradeep Thakurta (Google DeepMind)
Elephants Do Not Neglect: Differential Privateness with State Continuity for Privateness Finances by Jiankai Jin (The College of Melbourne), Chitchanok Chuengsatiansup (The College of Melbourne), Toby Murray (The College of Melbourne), Benjamin I. P. Rubinstein (The College of Melbourne), Yuval Yarom (Ruhr College Bochum), and Olga Ohrimenko (The College of Melbourne)
Improved Differentially Non-public Continuous Commentary Utilizing Group Algebra by Monika Henzinger (Institute of Science and Know-how (ISTA) Austria) and Jalaj Upadhyay (Rutgers College)
Occasion-Optimum Non-public Density Estimation within the Wasserstein Distance by Vitaly Feldman, Audra McMillan, Satchit Sivakumar (Boston College), and Kunal Talwar
Leveraging Mannequin Steerage to Extract Coaching Information from Customized Diffusion Fashions by Xiaoyu Wu (Carnegie Mellon College), Jiaru Zhang (Purdue College), and Steven Wu (Carnegie Mellon College)
Native Pan-privacy for Federated Analytics by Vitaly Feldman, Audra McMillan, Man N. Rothblum, and Kunal Talwar
Almost Tight Black-Field Auditing of Differentially Non-public Machine Studying by Meenatchi Sundaram Muthu Selva Annamalai (College School London) and Emiliano De Cristofaro (College of California, Riverside)
On the Value of Differential Privateness for Hierarchical Clustering by Chengyuan Deng (Rutgers College), Jie Gao (Rutgers College), Jalaj Upadhyay (Rutgers College), Chen Wang (Texas A&M College), and Samson Zhou (Texas A&M College)
Operationalizing Contextual Integrity in Privateness-Aware Assistants by Sahra Ghalebikesabi (Google DeepMind), Eugene Bagdasaryan (Google Analysis), Ren Yi (Google Analysis), Itay Yona (Google DeepMind), Ilia Shumailov (Google DeepMind), Aneesh Pappu (Google DeepMind), Chongyang Shi (Google DeepMind), Laura Weidinger (Google DeepMind), Robert Stanforth (Google DeepMind), Leonard Berrada (Google DeepMind), Pushmeet Kohli (Google DeepMind), Po-Sen Huang (Google DeepMind), and Borja Balle (Google DeepMind)
PREAMBLE: Non-public and Environment friendly Aggregation through Block Sparse Vectors by Hilal Asi, Vitaly Feldman, Hannah Keller (Aarhus Univiersity; work completed whereas at Apple), Man N. Rothblum, Kunal Talwar
Privateness amplification by random allocation by Vitaly Feldman (Apple) and Moshe Shenfeld (The Hebrew College of Jerusalem)
Privateness of Noisy Stochastic Gradient Descent: Extra Iterations with out Extra Privateness Loss by Jason Altschuler (MIT) and Kunal Talwar
Privately Estimating a Single Parameter by John Duchi (Stanford College), Hilal Ali, and Kunal Talwar
Scalable Non-public Search with Wally by Hilal Asi, Fabian Boemer, Nicholas Genise, Muhammad Haris Mughees, Tabitha Ogilvie, Rehan Rishi, Man N. Rothblum, Kunal Talwar, Karl Tarbe, Ruiyu Zhu, and Marco Zuliani
Shifted Composition I: Harnack and Reverse Transport Inequalities by Jason Altschuler (College of Pennsylvania) and Sinho Chewi (IAS)
Shifted Interpolation for Differential Privateness by Jinho Bok (College of Pennsylvania), Weijie Su (College of Pennsylvania), and Jason Altschuler (College of Pennsylvania)
Tractable Settlement Protocols by Natalie Collina (College of Pennsylvania), Surbhi Goel (College of Pennsylvania), Varun Gupta (College of Pennsylvania), and Aaron Roth (College of Pennsylvania)
Tukey Depth Mechanisms for Sensible Non-public Imply Estimation by Gavin Brown (College of Washington) and Lydia Zakynthinou (College of California, Berkeley)
Person Inference Assaults on Giant Language Fashions by Nikhil Kandpal (College of Toronto & Vector Institute), Krishna Pillutla (Google), Alina Oprea (Google, Northeastern College), Peter Kairouz (Google), Christopher A. Choquette-Choo (Google), and Zheng Xu (Google)
Universally Occasion-Optimum Mechanisms for Non-public Statistical Estimation by Hilal Asi, John C. Duchi (Stanford College), Saminul Haque (Stanford College), Zewei Li (Northwestern College), and Feng Ruan (Northwestern College)
“What would you like from idea alone?” Experimenting with Tight Auditing of Differentially Non-public Artificial Information Technology by Meenatchi Sundaram Muthu Selva Annamalai (College School London), Georgi Ganev (College School London, Hazy), and Emiliano De Cristofaro (College of California, Riverside)
Acknowledgments
Many individuals contributed to this workshop together with Hilal Asi, Anthony Chivetta, Vitaly Feldman, Haris Mughees, Martin Pelikan, Rehan Rishi, Man Rothblum, and Kunal Talwar.
