search

kdmalc / personalization-privacy-risk

Privacy analysis for ML and classical filtering personalization parameters
0 stars 0 forks source link

Global/Client PCA Transform Model #19

Closed kdmalc closed 9 months ago

kdmalc commented 1 year ago

Maybe have PCA transform as a hyperparameter for FL? Deal with this later

Normally, you would train a PCA model (ie the transform) on all data then use that transform to fit the testing set (for us, each user's update). What I am currently doing is creating a new PCA model (transform) every time I advance to the next update and then passing my input data through that. This may be okay since for our dataset we don't know that the PCs are the same since the orientation is, AFAIK, random placement on the forearm, thus the PCs could be representing different things.

PCA may be bad since if the 2 decs we are adding have very different orientations then aggregating them may have no real meaning

kdmalc commented 9 months ago

Actually have opted to turn PCA off for this reason, Momona believes that having each client run and "train" (instantiate) their own PCA object will be a form of personalization. The ML model should learn to do any dimensionality reduction that is necessary.