search

QinbinLi / DPBoost

Privacy-Preserving Gradient Boosting Decision Trees (AAAI 2020)
MIT License
24 stars 12 forks source link

Could u explain the parameters in detail? #1

Closed hacql2004 closed 4 years ago

hacql2004 commented 4 years ago

Hi, thanks for your sharing! Follow your instruction, I've run your demo code(exp_run.py) successfully. But I still get confused of some parameters(list in the below image) in it. Could you explain them more specifically? Thanks! image

hacql2004 commented 4 years ago

Here is the result of your demo. image I notice that the rmse turns smaller along with the larger value of para:total_budget([1, 2, 4, 6, 8, 10]). Does this para represent the num of parties in FL system? If so, in my view, the rmse should turn larger since the more parties the lower accurancy. Could u help me to figure out it? Thanks.

QinbinLi commented 4 years ago

Hi @hacql2004 ,

The explanation of some parameters is available in the Parameter section here. The other parameters are used for our previous testing and may not be useful for you. We may delete them in the future.

The total_budget is the total privacy budget (\epsilon in differential privacy). If the privacy budget is bigger, the privacy level is lower and the accuracy usually is higher.

hacql2004 commented 4 years ago

Thanks for your quickly reply. I found here from your AAAI2020 paper "Practical Federated Gradient Boosting Decision Trees". It seems I need to read your "Privacy-Preserving Gradient Boosting Decision Trees" first.