【论文记录】Differentially Private Empirical Risk Minimization with Input Perturbation
Problem Definition and Preliminary
- privacy concerns
data privacy : when the data contributors release their own data to the database
model privacy : when the database publishes the learned model to the model user - 采用的方法
针对 model privacy : ( ε varepsilon ε, δ delta δ)-differential privacy
针对 data privacy : ( ε varepsilon ε, δ delta δ)-local differential privacy ,,, (注 : 此种定义通常用在 under the non-interactive setting) - how to assess utility
借助 empirical excess risk ,,, the empirical excess risk of M mathcal M M that randomly outputs w mathcal w w is defined as , J ( w ; D ) − J ( w ∗ ; D ) J(mathcal w ,; D)-J(mathcal w^* ; D) J(w;D)−J(w∗;D)
Ref
Fukuchi, K., Tran, Q. K., & Sakuma, J. (2017, October). Differentially private empirical risk minimization with input perturbation. In International Conference on Discovery Science (pp. 82-90). Springer, Cham.
