README.md

Benchmarking Federated Learning Algorithms

Dependencies

• PyTorch >= 1.0.0
• torchvision >= 0.2.1
• scikit-learn >= 0.23.1

Centralized training

The maximum classification performance achived by the ResNet-50 when trained and validated on CIFAR-10 dataset. This acts as the upper bound for all the below compared federated learning methods.

Parameters	Network	Dataset	Learning Rate	Batch-size	Epochs	Optimizer	Schedular
Values	ResNet50	CIFAR-10	1e-4	256	150	Adam	OneCycleLR

Model	Normalization Layer	Number of Parameters	Accuracy
ResNet-50	BN	23528522	86.91
ResNet-50	GN	23528522	87.48

Federated Average

Parameters	# of clients	Learning Rate	Comm Rounds	Optimizer	Client-BatchSize	Client Epochs	Beta (NonIID)	Client Fraction
Values	100	1e-4	150	Adam	64	20	0.5	0.1

IID Distribution

Model	Normalization Layer	Accuracy
ResNet-50	BN	81.14
ResNet-50	GN	78.60

Non-IID Distribution

Model	Normalization Layer	Accuracy
ResNet-50	BN	60.00
ResNet-50	GN	60.51

Federated Group Knowledge Transfer

Parameters	# of clients	Learning Rate	Comm Rounds	Optimizer	Client-BatchSize	Client Epochs	Beta	Client Fraction
Values	100	1e-4	50	Adam/SGD	256	5	0.5	0.1

Parameters	Alpha	Temperature	Server Epochs
Values	1	3	50

IID Distribution

Model	Normalization Layer	Accuracy
ResNet-49	BN	55.28
ResNet-49	GN	53.27

Non-IID Distribution

Model	Normalization Layer	Accuracy
ResNet-49	BN	27.01
ResNet-49	GN	19.79

Additional Experimentation

Model-Contrastive Federated Learning

Parameters	# of clients	Learning Rate	Comm Rounds	Optimizer	Client-BatchSize	Client Epochs	Beta (NonIID)	Client Fraction
Values	100	0.01	50	SGD	64	5	0.5/1/5	0.1

Method	Accuracy @Beta=0.5	Accuracy @Beta=1	Accuracy @Beta=5
FedAvg	49.87	54.17	57.34
FedProx	59.10	62.88	67.61
MOON	63.62	68.23	75.93

Method	# of Rounds @beta=0.5	SpeedUp @beta=0.5	# of Rounds @beta=1	SpeedUp @beta=1	# of Rounds @beta=5	SpeedUp @beta=5
FedAvg	50	1x	50	1x	50	1x
FedProx	35	1.42x	34	1.47x	27	1.85x
MOON	25	2x	23	2.17x	16	3.12x

MOON-Prox

MOON-Prox is an intuitive extention of MOON. In this, we add a proximal term that acts as regulariser enforcing the local model to be close to the global model. The loss for local device can be written as:

$$ l = l_{cross-entropy} + l_{contrastive} + \frac{\mu}{2\alpha}\Vert \omega_{local} - \omega_{global} \Vert^2 $$

The main advantage of this method is that it has the same number of parameters as moon.

Method	Acc@Epochs = 50@Beta=0.5	Accuracy @Epochs = 100Beta=0.5
FedAvg	49.87	56.96
FedProx	59.10	65.21
MOON	63.62	77.03
MOON-Prox($\alpha$= 100)	66.50	77.25
MOON-Prox($\alpha$= 500)	63.16	77.59

Gradient Attack in Federated Learning

Acknowledgement

We thank @QinbinLi and @FedML for their open source code that helped in our project and @debcaldarola being a wonderful mentor throughout the project.