epfml
Integrate Differential Privacy
Differential Privacy is implemented but never used. It seems that it was intentionally removed during a code refactoring (in this PR, l.179 of decentralized/base.ts)
- [ ] Integrate DP again in the new architecture code
- [ ] Look into how DP is implemented. It seems it is epsilon-DP that is implemented, which has a very strict privacy bound at the cost of lower performance. Most implementations use (epsilon, delta)-DP for a better trade-off between privacy and utility. However, the latter needs to track the privacy budget delta for each user. This is what the MomentAccount tracks in Deep learning with differential privacy.
- [ ] Create test for the differential privacy feature
@ysfalh expressed some interest in this, as they have a decentralized DP scheme which is easy to implement (only needs pairwise agreement on each edge, which we have anyway already from the secAgg implementation)
@ysfalh expressed some interest in this, as they have a decentralized DP scheme which is easy to implement (only needs pairwise agreement on each edge, which we have anyway already from the secAgg implementation)
Yes, I am happy to help with adapting our code.
