Adaptive Semi-supervised Federated Learning with Selective Knowledge Distillation

This work builds on the work of FedMD

Hypothesis 01 : Introducing poisonous nodes in the federated network can have negative impact on collaborative learning

Experiemnt: Introduce varying number of poisoned node in all four environments of FedMD an observe the impact on collaborative learning.

Results:

FEMNIST Balanced:

FEMNIST Imbalanced:

CIFAR Balanced:

CIFAR Imbalanced:

Hypothesis 02 : Selective Knowledge Distillation (SKD) can minimize the impact of poisoned nodes in collaborative learning

Framework:

Experiemnt: We test different variations of SKD algorithm on all four environements with 40% nodes poisoned

Results:

FEMNIST Balanced:

FEMNIST Imbalanced:

CIFAR Balanced:

CIFAR Imbalanced:

Hypothesis 03 : Selective Knowledge Distillation (SKD) will also work on semi-supervised setting

Semi-supervised learning methodology:


Experiemnt: We test SKD algorithm on all four environements with 40% nodes poisoned

Results:

FEMNIST Balanced:

FEMNIST Imbalanced:

CIFAR Balanced:

CIFAR Imbalanced:

Reproduce Resutls

In order to reproduce the results:

• Create an anaconda environment from the environment.yml file
• Run the specific .py files for specific experiments