The apprenticeship from wenyanliu

wenyanliu / apprenticeship Goto Github PK

View Code? Open in Web Editor NEW

A notebook of awesome privacy protection,federated learning, fairness and blockchain research materials.

License: MIT License

apprenticeship's People

Contributors

Stargazers

Watchers

apprenticeship's Issues

How to set the privacy budget of the derived item latent matrix?

Question

Why the privacy budget of the derived $\mathbf{V}$ is set as $\epsilon =\frac{\beta \epsilon _n}{1+\beta }$ ?

Xuying Meng, Suhang Wang, Kai Shu, Jundong Li, Bo Chen, Huan Liu, Yujun Zhang: Personalized Privacy-Preserving Social Recommendation. AAAI 2018: 3796-3803

Section Modeling Sensitive and Non-sensitive Ratings, Page 4, the right column, the second paragraph from bottom

We allocate a much smaller privacy budget $\epsilon_s$ for sensitive ratings and a larger $\epsilon_n$ for non-sensitive ones where $\epsilon_s=\beta \epsilon_n$ and the domain of $\beta$ is which is used to control the relative noise magnitude. Then sensitive ratings can receive better privacy protection with the small privacy budget $\epsilon_s$ . We set the privacy budget of the derived $\mathbf{V}$ as $\epsilon =\frac{\beta \epsilon _n}{1+\beta }$ .

Expected solution

Since sensitive and non-sensitive ratings are mutually exclusive, I am not sure if the pricacy budget of $\mathbf{V}$ is $Max(\epsilon _s, \epsilon _n)=\epsilon _n$ .

Additional doubt

If the privacy budget of $\mathbf{V}$ is $\epsilon_n$ , is it unfair to compare the level of privacy protection provided by PrivSR and DPMF in the experimental evaluation?

Recommend Projects