Table 4.
Evaluation of collaborative and non-collaborative learning on Test and External Datasets using DP-SGD, achieving privacy parameter ε = 2.90 for δ = 0.0001.
| Source hospital | Non-collaborative training | DP-FL training | FL training | Combined training | ||||
|---|---|---|---|---|---|---|---|---|
| Test | External | Test | External | Test | External | Test | External | |
| International Genomics Consortium | 0.654 | 0.631 | 0.823 ± 0.01 | 0.707 ± 0.01 | 0.823 ± 0.01 | 0.741 ± 0.01 | 0.839 ± 0.01 | 0.768 ± 0.003 |
| Indivumed | 0.648 | 0.556 | ||||||
| Asterand | 0.709 | 0.701 | ||||||
| John Hopkins | 0.681 | 0.600 | ||||||
For FL and combined training we report the mean accuracy and standard deviation across the client’s test datasets. On the external dataset we ran the experiments using three random initializations, and report the mean accuracy and standard deviation across them.