Table 2 Performance evaluation of DynamicFL and other methods on ResNet-18 and ViT-Base architectures

Model	Methods	Ratio (7:2:1)	Ratio (5:2:3)	Ratio (4:1:5)	Ratio (4:3:3)	Ratio (3:6:1)
ResNet-18	DSFL²⁶	70.12	73.34	68.45	74.23	71.56
	FedET²⁷	71.78	74.56	69.78	75.45	73.12
	IncluFL³⁴	72.45	75.12	70.45	76.12	73.78
	FedMD³⁵	70.89	73.01	68.12	74.01	71.67
	FedTGP⁴⁰	74.56	76.23	74.45	75.12	73.34
	FedDF²⁵	73.34	76.01	71.78	76.89	74.23
	pFedHR³⁹	74.45	77.34	73.89	77.45	75.67
	FCCL³⁸	73.78	76.12	74.78	77.01	73.89
	HeteroFL²⁸	72.34	75.45	70.56	76.34	73.12
	DepthFL³⁶	68.45	71.12	66.12	72.34	69.78
	FedRolex³⁷	73.89	76.23	71.45	77.01	74.45
	DynamicFL	82.12	83.34	80.12	83.45	81.23
ViT-Base	DSFL²⁶	69.34	72.12	67.45	73.01	70.45
	FedET²⁷	70.89	73.34	68.78	74.12	71.67
	IncluFL³⁴	71.56	74.12	69.45	74.89	72.34
	FedMD³⁵	70.12	72.89	67.78	73.34	70.89
	FedDF²⁵	72.45	75.12	70.34	75.89	73.12
	DynamicFL	80.45	81.34	78.89	82.45	79.67

The table presents the performance of DynamicFL under different client distribution ratios: (7:2:1), (5:2:3), (4:1:5), (4:3:3), and (3:6:1). These ratios represent the proportion of data distributed among clients with varying computational and data resources, reflecting heterogeneous federated learning scenarios. The results highlight that DynamicFL consistently outperforms baseline methods across all distribution ratios, demonstrating its ability to handle heterogeneous client distributions effectively and achieve superior performance in both ResNet-18 and ViT-Base architectures. The results in bold indicate the best performance. Source data are provided as a Source Data file.

Quick links

Search