Fig. 6
From: Efficient differentially private learning improves drug sensitivity prediction
Key trade-offs in differentially private learning. Relative improvements over baseline (10 non-private data points). a, As the dimensionality increases, more data are needed to improve performance of the private methods. b, With enough private data, adding more non-private data does not significantly increase the performance. c, More data are needed if privacy guarantees are tighter (privacy parameter ε is smaller). Data dimensionality is 10, the size of non-private data is 10, and ε=2 (except when otherwise noted)