P3 Toward Privacy-Preserving Photo Sharing

bes some statistically interesting aspect of the image. We evaluate the ef?cacy of attacking P3 by perform- ing SIFT feature extraction on the public part. For this, we use the implementation [32] from the designer of SIFT together with the default parameters for feature ex- traction and parison. Figure 7(c) reports the results of running feature extraction on the USC-SIPI dataset. 10This ?gure shows two lines, one of which mea- sures the total number of features detected on the public part as a function of threshold. This shows that as the 10The SIFT algorithm putationally expensive, and the INRIA data set is large, so we do not have the results for the INRIA dataset. (Recall that we need pute for a large number of threshold val- ues). We expect the results to be qualitatively similar. 10