Lifelong DP: Consistently Bounded Differential Privacy in Lifelong Machine Learning

Phung Lai, Han Hu, Hai Phan, Ruoming Jin, My Thai, An Chen
Proceedings of The 1st Conference on Lifelong Learning Agents, PMLR 199:778-797, 2022.

Abstract

In this paper, we show that the process of continually learning new tasks and memorizing previous tasks introduces unknown privacy risks and challenges to bound the privacy loss. Based upon this, we introduce a formal definition of Lifelong DP, in which the participation of any data tuples in the training set of any tasks is protected, under a consistently bounded DP protection, given a growing stream of tasks. A consistently bounded DP means having only one fixed value of the DP privacy budget, regardless of the number of tasks. To preserve Lifelong DP, we propose a scalable and heterogeneous algorithm, called L2DP-ML with a streaming batch training, to efficiently train and continue releasing new versions of an L2M model, given the heterogeneity in terms of data sizes and the training order of tasks, without affecting DP protection of the private training set. An end-to-end theoretical analysis and thorough evaluations show that our mechanism is significantly better than baseline approaches in preserving Lifelong DP. The implementation of L2DP-ML is available at: https://github.com/haiphanNJIT/PrivateDeepLearning.

Cite this Paper

BibTeX
@InProceedings{pmlr-v199-lai22a, title = {Lifelong DP: Consistently Bounded Differential Privacy in Lifelong Machine Learning}, author = {Lai, Phung and Hu, Han and Phan, Hai and Jin, Ruoming and Thai, My and Chen, An}, booktitle = {Proceedings of The 1st Conference on Lifelong Learning Agents}, pages = {778--797}, year = {2022}, editor = {Chandar, Sarath and Pascanu, Razvan and Precup, Doina}, volume = {199}, series = {Proceedings of Machine Learning Research}, month = {22--24 Aug}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v199/lai22a/lai22a.pdf}, url = {https://proceedings.mlr.press/v199/lai22a.html}, abstract = {In this paper, we show that the process of continually learning new tasks and memorizing previous tasks introduces unknown privacy risks and challenges to bound the privacy loss. Based upon this, we introduce a formal definition of Lifelong DP, in which the participation of any data tuples in the training set of any tasks is protected, under a consistently bounded DP protection, given a growing stream of tasks. A consistently bounded DP means having only one fixed value of the DP privacy budget, regardless of the number of tasks. To preserve Lifelong DP, we propose a scalable and heterogeneous algorithm, called L2DP-ML with a streaming batch training, to efficiently train and continue releasing new versions of an L2M model, given the heterogeneity in terms of data sizes and the training order of tasks, without affecting DP protection of the private training set. An end-to-end theoretical analysis and thorough evaluations show that our mechanism is significantly better than baseline approaches in preserving Lifelong DP. The implementation of L2DP-ML is available at: https://github.com/haiphanNJIT/PrivateDeepLearning.} }
Endnote
%0 Conference Paper %T Lifelong DP: Consistently Bounded Differential Privacy in Lifelong Machine Learning %A Phung Lai %A Han Hu %A Hai Phan %A Ruoming Jin %A My Thai %A An Chen %B Proceedings of The 1st Conference on Lifelong Learning Agents %C Proceedings of Machine Learning Research %D 2022 %E Sarath Chandar %E Razvan Pascanu %E Doina Precup %F pmlr-v199-lai22a %I PMLR %P 778--797 %U https://proceedings.mlr.press/v199/lai22a.html %V 199 %X In this paper, we show that the process of continually learning new tasks and memorizing previous tasks introduces unknown privacy risks and challenges to bound the privacy loss. Based upon this, we introduce a formal definition of Lifelong DP, in which the participation of any data tuples in the training set of any tasks is protected, under a consistently bounded DP protection, given a growing stream of tasks. A consistently bounded DP means having only one fixed value of the DP privacy budget, regardless of the number of tasks. To preserve Lifelong DP, we propose a scalable and heterogeneous algorithm, called L2DP-ML with a streaming batch training, to efficiently train and continue releasing new versions of an L2M model, given the heterogeneity in terms of data sizes and the training order of tasks, without affecting DP protection of the private training set. An end-to-end theoretical analysis and thorough evaluations show that our mechanism is significantly better than baseline approaches in preserving Lifelong DP. The implementation of L2DP-ML is available at: https://github.com/haiphanNJIT/PrivateDeepLearning.
APA
Lai, P., Hu, H., Phan, H., Jin, R., Thai, M. & Chen, A.. (2022). Lifelong DP: Consistently Bounded Differential Privacy in Lifelong Machine Learning. Proceedings of The 1st Conference on Lifelong Learning Agents, in Proceedings of Machine Learning Research 199:778-797 Available from https://proceedings.mlr.press/v199/lai22a.html.

Related Material