What is it about?
Privacy-preserving Tensor Decomposition (PTD) method to anonymize human faces. Core tensor of Tucker decomposition generated from the original face input can effectively represent the underlying characteristics of the original face data; that is, learning only from the core tensors is sufficient for differentiating real human face images from deepfakes.
Featured Image
Photo by Dayne Topkin on Unsplash
Why is it important?
Training data may include personal information such as human faces, which requires anonymization to provide user privacy. However, after anonymization, the performance of the original machine learning (ML) model degrades due to the reduced or missing information.
Perspectives
I expect that the tensor decomposition can be useful for deep learning which mostly uses data and model parameter as tensor. We can leverage the decomposed features as the additional information of the input tensor such as in self-supervise learning.
jeonghokim Kim
Read the Original
This page is a summary of: PTD, April 2022, ACM (Association for Computing Machinery),
DOI: 10.1145/3477314.3507036.
You can read the full text:
Resources
Contributors
The following have contributed to this page
