1: \begin{abstract}

2: 	This paper analyzes $k$ nearest neighbor classification with training

3: 	data anonymized using \emph{anatomy}. Anatomy preserves all

4: 	data values, but introduces uncertainty in the mapping between

5: 	identifying and sensitive values. 	We first study the theoretical effect of the 

6: 	anatomized training data on the $k$ nearest neighbor error rate bounds,

7: 	nearest neighbor convergence rate, and Bayesian error. We then validate 

8: 	the derived bounds empirically. We show that 1) Learning from anatomized 

9: 	data approaches the limits of learning through the unprotected data (although

10: 	requiring larger training data), and 2) nearest neighbor using anatomized data 

11: 	outperforms nearest neighbor on generalization-based anonymization.

12: \end{abstract}