Implementation code is still under development.
Data Example
You can find the processed data in this link
MNIST (IID 10 clients) :
| user_id | data_num | classes_num | entropy | 0 | ... | 9 |
|---|---|---|---|---|---|---|
| 0 | 5000 | 10 | 3.318743 | 518 | ... | 501 |
| 1 | 5000 | 10 | 3.319638 | 503 | ... | 468 |
| ... | ... | ... | ... | ... | ... | ... |
| 9 | 5000 | 10 | 3.319255 | 503 | ... | 490 |
IID Partition:
Each client has the same classes and the same number of classes. Its purpose is to test whether a contribution prediction method yields similar contribution results for approximately similar clients.
Same Distribution & Different Size:
Data with the same distribution but different sizes; clients 1 and 2 each have 5% of the data; clients 3 and 4 each have 7.5%; clients 5 and 6 each have 10%; clients 7 and 8 each have 12.5%; clients 9 and 10 each have 15%.
Label Noise:
Flip the labels of a certain proportion of samples within each client. Clients 1 and 2 have no noise; clients 3 and 4 have a noise rate of 5%; clients 5 and 6 have a noise rate of 10%; clients 7 and 8 have a noise rate of 15%; clients 9 and 10 have a noise rate of 20%. Clients with lower noise should have a higher contribution.
Feature Noise:
Add Gaussian noise to the features of samples within each client. Participants 1 and 2 have no noise; participants 3 and 4 have 5% noise; participants 5 and 6 have 10% noise; participants 7 and 8 have 15% noise; participants 9 and 10 have 20% noise. The lower the noise percentage, the higher the contribution.
Different Distribution & Same Size:
Different distributions with the same data size, using Dirichlet(1.0) to partition.