【论文记录】小应用

《Bounding User Contributions: A Bias-Variance Trade-off in Differential Privacy》

• 背景：trade-off
  解决方法：limiting the contributions of individual users in order to reduce the sensitivity.

• 举例 · 从数学原理上解释应用差分隐私时需要 trade-off 的原因 ：
  S = { x 1 , x 2 , . . . , x n } \mathcal{S} = \{ x_1,x_2,...,x_n \} S={x1​,x2​,...,xn​} \quad $g(\mathcal{S})={\sum }_i{x}_i$ \quad 由于$\Delta g$发散 , 故设置bound T , 采用 $g_T(\mathcal{S})={\sum }_{i}min(x_i,T)$   \, 则使$\Delta g_T=T$
  \quad\quad\quad\quad

• 设置 bound 的方法：根据 F-utility 导出${\tau }_0\in (0,1]$
  

《Structural-Based Graph Publishing Under Differential Privacy》

• 背景 : previous methods, such as adjacency matrix, hierarchical random graph, quadtree, etc are often applicable in small-scale graph datasets. However, real graphs are generally large and sparse, leading to high computational expense.

• 算法框架 :1.Fast-unfolding community detection algorithm     \,\,\, 2.The framework of differential privacy based on hierarchical representation

• Fast-unfolding algorithm : 不断划分   \, community使得划分后的整个网络的模块度 Q=··· 不断增大。具体的算法过程为：
  1.初始化，将每个点划分在不同的社区中；
  2.对每个节点，将每个点尝试划分到与其邻接的点所在的社区中，计算此时的模块度，判断划分前后的模块度的差值ΔQ是否为正数，若为正数，则接受本次的划分，若不为正数，则放弃本次的划分；
  3.重复以上的过程，直到不能再增大模块度为止；
  4.构造新图，新图中的每个点代表的是步骤3中划出来的每个社区，继续执行步骤2和步骤3，直到社区的结构不再改变为止。

• Hierarchical Structure : hierarchical graph中的叶子结点为原图中的结点，非叶子结点的权值由$P_r=e_r/(n_{L_r}\cdot n_{R_r})$计算。举例如 :
  
  a hierarchical graph with higher $L$ is a better representation of the network structure than those with lower likelihoods.
  

• computing the probabilities stored in internal nodes with applying Laplace noise.