Distributed Differentially-Private Algorithms for Matrix and Tensor Factorization.

Area of Science:

• Signal Processing
• Machine Learning
• Data Privacy

Background:

• Distributed datasets with private information necessitate privacy-preserving algorithms.
• Tensor and matrix factorizations are crucial in signal processing and machine learning.
• Differentially private algorithms often introduce noise, impacting performance in distributed settings.

Purpose of the Study:

• To design improved distributed and differentially private algorithms for PCA and OTD.
• To mitigate the performance degradation caused by noise in privacy-preserving algorithms.

Main Methods:

• Development of new distributed, differentially private algorithms for PCA and OTD.
• Implementation of a correlated noise design scheme to reduce the impact of added noise.
• Experimental validation on synthetic and real-world datasets.

Main Results:

• The new algorithms achieve performance comparable to centralized methods under differential privacy.
• The correlated noise scheme effectively alleviates the detrimental effects of noise.
• Outperformed previous distributed privacy-preserving factorization methods.

Conclusions:

• Meaningful utility is achievable in distributed machine learning while guaranteeing differential privacy.
• The proposed correlated noise strategy offers a significant advancement for privacy-preserving distributed tensor and matrix factorization.