Timezone: »
We provide an end-to-end differentially private spectral algorithm for learning LDA, based on matrix/tensor decompositions, and establish theoretical guarantees on utility/consistency of the estimated model parameters. We represent the spectral algorithm as a computational graph. Noise can be injected along the edges of this graph to obtain differential privacy. We identify subsets of edges, named ``configurations'', such that adding noise to all edges in such a subset guarantees differential privacy of the end-to-end spectral algorithm. We characterize the sensitivity of the edges with respect to the input and thus estimate the amount of noise to be added to each edge for any required privacy level. We then characterize the utility loss for each configuration as a function of injected noise. Overall, by combining the sensitivity and utility characterization, we obtain an end-to-end differentially private spectral algorithm for LDA and identify which configurations outperform others under specific regimes. We are the first to achieve utility guarantees under a required level of differential privacy for learning in LDA. We additionally show that our method systematically outperforms differentially private variational inference.
Author Information
Christopher DeCarolis (University of Maryland)
Chris is an undergraduate student at the University of Maryland with research interests in machine learning. He does research under Professor Furong Huang, and has done internships at Microsoft and Facebook in the past.
Mukul A Ram (University of Maryland)
Seyed Esmaeili (University of Maryland, College Park)
Yu-Xiang Wang (UC Santa Barbara)

Yu-Xiang Wang is the Eugene Aas Assistant Professor of Computer Science at UCSB. He runs the Statistical Machine Learning lab and co-founded the UCSB Center for Responsible Machine Learning. He is also visiting Amazon Web Services. Yu-Xiang’s research interests include statistical theory and methodology, differential privacy, reinforcement learning, online learning and deep learning.
Furong Huang (University of Maryland College Park)
More from the Same Authors
-
2021 : Privately Publishable Per-instance Privacy: An Extended Abstract »
Rachel Redberg · Yu-Xiang Wang -
2021 : Optimal Accounting of Differential Privacy via Characteristic Function »
Yuqing Zhu · Jinshuo Dong · Yu-Xiang Wang -
2021 : Optimal Uniform OPE and Model-based Offline Reinforcement Learning in Time-Homogeneous, Reward-Free and Task-Agnostic Settings »
Ming Yin · Yu-Xiang Wang -
2021 : Near-Optimal Offline Reinforcement Learning via Double Variance Reduction »
Ming Yin · Yu Bai · Yu-Xiang Wang -
2022 : Optimal Dynamic Regret in LQR Control »
Dheeraj Baby · Yu-Xiang Wang -
2023 : A Privacy-Friendly Approach to Data Valuation »
Jiachen Wang · Yuqing Zhu · Yu-Xiang Wang · Ruoxi Jia · Prateek Mittal -
2023 : Why Quantization Improves Generalization: NTK of Binary Weight Neural Network »
Kaiqi Zhang · Ming Yin · Yu-Xiang Wang -
2023 : Generative Autoencoders as Watermark Attackers: Analyses of Vulnerabilities and Threats »
Xuandong Zhao · Kexun Zhang · Yu-Xiang Wang · Lei Li -
2023 : Provable Robust Watermarking for AI-Generated Text »
Xuandong Zhao · Prabhanjan Ananth · Lei Li · Yu-Xiang Wang -
2023 Poster: Offline Reinforcement Learning with Closed-Form Policy Improvement Operators »
Jiachen Li · Edwin Zhang · Ming Yin · Jerry Bai · Yu-Xiang Wang · William Wang -
2023 Poster: Protecting Language Generation Models via Invisible Watermarking »
Xuandong Zhao · Yu-Xiang Wang · Lei Li -
2023 Poster: Differentially Private Optimization on Large Model at Small Cost »
Zhiqi Bu · Yu-Xiang Wang · Sheng Zha · George Karypis -
2023 Poster: Non-stationary Reinforcement Learning under General Function Approximation »
Songtao Feng · Ming Yin · Ruiquan Huang · Yu-Xiang Wang · Jing Yang · Yingbin LIANG -
2023 Poster: Global Optimization with Parametric Function Approximation »
Chong Liu · Yu-Xiang Wang -
2023 Expo Talk Panel: Generative AI and Science »
Srinivasan Sengamedu · Joseph Sirosh · Yu-Xiang Wang · Alexander Amini -
2022 Poster: Sample-Efficient Reinforcement Learning with loglog(T) Switching Cost »
Dan Qiao · Ming Yin · Ming Min · Yu-Xiang Wang -
2022 Spotlight: Sample-Efficient Reinforcement Learning with loglog(T) Switching Cost »
Dan Qiao · Ming Yin · Ming Min · Yu-Xiang Wang -
2019 Poster: Poission Subsampled R\'enyi Differential Privacy »
Yuqing Zhu · Yu-Xiang Wang -
2019 Oral: Poission Subsampled R\'enyi Differential Privacy »
Yuqing Zhu · Yu-Xiang Wang -
2018 Poster: Detecting and Correcting for Label Shift with Black Box Predictors »
Zachary Lipton · Yu-Xiang Wang · Alexander Smola -
2018 Oral: Detecting and Correcting for Label Shift with Black Box Predictors »
Zachary Lipton · Yu-Xiang Wang · Alexander Smola -
2018 Poster: Improving the Gaussian Mechanism for Differential Privacy: Analytical Calibration and Optimal Denoising »
Borja de Balle Pigem · Yu-Xiang Wang -
2018 Oral: Improving the Gaussian Mechanism for Differential Privacy: Analytical Calibration and Optimal Denoising »
Borja de Balle Pigem · Yu-Xiang Wang -
2018 Poster: Learning Deep ResNet Blocks Sequentially using Boosting Theory »
Furong Huang · Jordan Ash · John Langford · Robert Schapire -
2018 Poster: signSGD: Compressed Optimisation for Non-Convex Problems »
Jeremy Bernstein · Yu-Xiang Wang · Kamyar Azizzadenesheli · Anima Anandkumar -
2018 Oral: Learning Deep ResNet Blocks Sequentially using Boosting Theory »
Furong Huang · Jordan Ash · John Langford · Robert Schapire -
2018 Oral: signSGD: Compressed Optimisation for Non-Convex Problems »
Jeremy Bernstein · Yu-Xiang Wang · Kamyar Azizzadenesheli · Anima Anandkumar -
2017 Poster: Optimal and Adaptive Off-policy Evaluation in Contextual Bandits »
Yu-Xiang Wang · Alekh Agarwal · Miroslav Dudik -
2017 Talk: Optimal and Adaptive Off-policy Evaluation in Contextual Bandits »
Yu-Xiang Wang · Alekh Agarwal · Miroslav Dudik