Choosing parameters for Local Differential Privacy

Privacy & Security
1

Hi everyone,

I implemented a Federated Learning simulation where I train a Logistic Regression model and implemented local differential privacy that adds noise to the parameters with the LocaldpMod, since the server will be regarded as a untrusted party. I have a question on what to choose as parameters.

For the delta I chose a small number, 0.001 and the epsilon I will choose different values between 1-10 to compare the results.

I am having trouble to find what are ‘good’ values for the clipping_norm and sensitivity. I have found multiple resources on what they do, but I cannot find anything on what values to choose.

Does anyone have advice for me on this? Does anyone have references that explain how to decide these values for your simulation given your dataset and model?

Thanks!

1 Like
2

Hello @chielflower, these are great questions. I think @mohammad can give you some pointers on what to make good choice for these parameters.

3

Hi @chielflower, Can you please let me know what is the type of the dataset?

4

Hi @mohammad ,

It is a large but simple dataset of 1 feature, a float/percentage between 1-100, and a binary outcome label [0-1]. I am running a logistic regression model on it. The dataset is quite imbalance with only 14% of the labels being 1, and the dataset size is very different between clients.