Table 3 Performance evaluation of test accuracy (%) with different network structures

From: Towards fairness-aware and privacy-preserving enhanced collaborative learning for healthcare

 

CIFAR-10

CIFAR-100

BloodCell45

 

LeNet-5

IID

Non-IID

IID

Non-IID

IID

Non-IID

Global Model Parameters (Communication Overhead)

DSFL26

69.88

62.79

51.12

42.94

69.87

60.07

1.8 M

FedET27

75.81

70.29

52.43

46.61

71.53

62.96

1.8 M

IncluFL34

74.72

68.46

51.88

47.57

73.49

65.82

1.8 M

HeteroFL28

71.39

65.26

50.67

43.71

71.94

63.44

1.8 M

DepthFL36

70.22

63.57

49.17

40.15

70.48

63.17

1.8 M

FedRolex37

74.72

69.75

52.81

47.38

72.54

65.48

1.8 M

DynamicFL

79.33

74.21

63.08

55.92

77.36

71.29

0.6 M

 

CIFAR-10

CIFAR-100

BloodCell45

 

GoogLeNet

IID

Non-IID

IID

Non-IID

IID

Non-IID

Global Model Parameters (Communication Overhead)

DSFL26

70.35

64.57

52.61

44.73

70.49

65.37

20.7 M

FedET27

72.39

66.64

54.69

48.42

73.18

66.19

20.7 M

IncluFL34

72.73

67.86

55.28

46.59

75.29

67.94

20.7 M

HeteroFL28

72.93

65.84

52.74

43.19

74.26

67.41

20.7 M

DepthFL36

71.12

63.65

50.49

42.97

73.17

66.92

20.7 M

FedRolex37

77.85

72.46

54.28

48.75

76.38

70.48

20.7 M

DynamicFL

82.84

77.16

61.52

54.28

81.49

75.13

6.9 M

 

CIFAR-10

CIFAR-100

BloodCell45

 

ResNet-18

IID

Non-IID

IID

Non-IID

IID

Non-IID

Global Model Parameters (Communication Overhead)

DSFL26

71.31

66.90

53.64

44.97

78.47

72.11

35.1 M

FedET27

75.13

69.13

54.37

46.55

81.79

74.96

35.1 M

IncluFL34

79.87

72.66

56.72

50.18

80.36

73.82

35.1 M

HeteroFL28

74.54

67.95

53.81

45.27

81.19

72.25

35.1 M

DepthFL36

75.21

66.78

52.27

44.62

78.77

68.85

35.1 M

FedRolex37

80.63

74.75

55.26

50.07

82.73

72.18

35.1 M

DynamicFL

86.82

81.14

63.27

57.20

88.64

81.76

11.7 M

  1. Compared method are run under the same wall-clock training time.
  2. The table shows the results of DynamicFL under three different network structures. Non-IID heterogeneity is introduced using the Dirichlet distribution Dir(β) with a concentration parameter β = 1.The global round number for our method is fixed at 100. And the global round of the other baselines is dependent on their respective computation time, ensuring a fair total computation time compared to our method. The results highlighted in bold indicate superior performance. Source data are provided as a Source Data file.