Fed-DNN-Debugger: Automatically Debugging Deep Neural Network Models in Federated Learning

Table 3

Effects of the proportion of selected high-quality samples in the retraining data and the number of retraining samples.

Dataset	Ratio	500 (%)	1000 (%)	1500 (%)	2000 (%)	2500 (%)	3000 (%)	3500 (%)	4000 (%)	4500 (%)
MNIST	0.1	94.76	95.56	92.5	95.44	95.37	96	96.15	96.75	95.48
	0.15	93.85	90.89	94.35	92.27	96.32	95.32	95.25	96.42	94.94
	0.2	92.82	95.21	94.21	95.6	93.39	95.99	96.29	96.1	96.36
	0.25	94.65	94.03	94.58	95.35	96.27	95.9	94.95	96.09	95.56
	0.3	94.81	94.84	95.28	94.76	94.69	96.09	96.53	96.96	96.69
	0.35	93.36	92.13	95.03	94.35	96.08	96.35	95.84	96.39	94.81
	0.4	93.88	83.24	94.89	94.14	94.23	94.5	94.98	95.5	96.31
	0.45	93.89	93.67	93.46	94.44	93.61	95.13	93.96	96.41	96.64
CIFAR-10	0.1	69.63	71.79	72.69	72.28	74.05	71.60	74.44	72.99	74.52
	0.15	71.53	69.98	73.43	73.27	73.68	72.69	67.20	73.04	74.64
	0.2	72.36	70.46	73.25	71.53	68.30	73.52	72.63	74.02	74.34
	0.25	71.88	71.70	72.52	72.83	72.01	74.09	72.04	74.07	73.95
	0.3	73.40	70.14	72.39	70.72	68.84	69.68	72.79	74.86	74.01
	0.35	70.91	69.24	70.91	70.82	62.61	74.08	70.47	74.03	74.12
	0.4	70.27	70.69	71.45	72.88	74.30	73.55	73.66	74.38	73.17
	0.45	71.31	66.23	70.91	65.67	72.83	73.32	74.08	73.74	72.70

The two bold values indicate the combination of parameters for the best model performance on these two datasets.