Combined Semi-supervised learning with Privacy Perserving Data Mining algorithm, which outperformed all previous privacy preserving methods.
LLP is framework to enable semi-supervised learning on classification algorithm.
I implemented LLP on logistic regression, so the model learned from bag proprotion rather than
individual label.
Adding random noise sample from Laplace distribution which ensures differential privacy.
Tested on Adult dataset on income classification, and the model converged after adding enough laplace noise to both label proprotion and data matrix. Tested on Instagram hostile comment dataset, and the model converged after adding enough laplace noise to both label proprotion and data matrix.
Adding random noise sample from Laplace distribution which ensures differential privacy.
Thanks supporting from Professor Aron and his PHD students.