https://github.com/kartikey2807/differentially-private-sgd

Implement Differentially-private SGD. Large ε leads to loss of accuracy%.
https://github.com/kartikey2807/differentially-private-sgd

differential-privacy dpsgd opacus stochastic-gradient-descent

Last synced: about 2 months ago
JSON representation

Implement Differentially-private SGD. Large ε leads to loss of accuracy%.

• Host: GitHub
• URL: https://github.com/kartikey2807/differentially-private-sgd
• Owner: kartikey2807
• Created: 2025-05-26T09:59:36.000Z (5 months ago)
• Default Branch: main
• Last Pushed: 2025-08-16T00:17:30.000Z (2 months ago)
• Last Synced: 2025-08-16T02:35:59.303Z (2 months ago)
• Topics: differential-privacy, dpsgd, opacus, stochastic-gradient-descent
• Language: Jupyter Notebook
• Homepage:
• Size: 22.6 MB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Differentially-Private SGD

References: [Abadi, Martin et al.](https://arxiv.org/pdf/1607.00133)   

**Summary**   

* Differential privacy added to *gradient step*

* Tradeoff exists between utility and privacy

* Having privacy budget $\epsilon$ ensure weights are private

* Gradient access private --> Model is differentially private (**post-hoc**)

---

**Results**   

On the MNIST dataset

|Model|Test Accuracy|

| :- | :- |

|Vanilla|97.00%|

|Differentially Private|91.00%|