Table 2 Performance of the LR10 classifier on aggregated crowd features across privacy-preserving mechanisms when using the mode, median, and mean aggregation methods, respectively.
Privacy mechanism | Accuracy (%) | Precision (%) | Sensitivity [Recall] (%) | Specificity (%) | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
Mode | Median | Mean | Mode | Median | Mean | Mode | Median | Mean | Mode | Median | Mean | |
Unaltered | 96.0 | 92.0 | 90.0 | 100.0 | 88 | 85.7 | 92.0 | 96.0 | 96.0 | 100.0 | 88.0 | 84.0 |
Face box | 94.0 | 88.0 | 82.0 | 95.8 | 85.2 | 73.5 | 92.0 | 92.0 | 100.0 | 96.0 | 84.0 | 96.0 |
Pitch shift | 82.0 | 82.0 | 88.0 | 83.3 | 73.6 | 71.4 | 80.0 | 100.0 | 100.0 | 84.0 | 64.0 | 60.0 |
Face box and pitch shift | 86.0 | 78.0 | 80.0 | 84.6 | 70.1 | 71.4 | 88.0 | 96.0 | 100.0 | 84.0 | 60.0 | 60.0 |
