Title Page
Abstract
국문 초록
Contents
Chapter 1. Introduction 13
Chapter 2. Related Works 16
Chapter 3. System Model 19
3.1. Federated Learning 19
3.2. Aggregation Weights 21
Chapter 4. Proposed Methods 23
4.1. Global-Local Gradients Distance 23
4.2. Local-Local Gradients Distance 25
4.3. Model-wise / Layer-wise 28
4.4. Adaptive Aggregation Algorithm 29
Chapter 5. Experiments 32
5.1. Experiments Setting 32
5.2. Experiments Results 36
5.2.1. Performance Comparison between the Proposed Algorithm and Conventional Algorithms 36
5.2.2. Performance Comparison with Different Participation Ratios 37
5.2.3. Performance Comparison with the Degree of Heterogeneity 40
5.3. Analysis 42
5.3.1. Robustness of the Proposed Adaptive Aggregation Algorithm 42
5.3.2. Analysis on the Proposed Aggregation Methods 43
Chapter 6. Conclusion 48
Reference 49
Table 5.1. Final accuracy (%) in a heterogeneous environment (α=0.1). 36
Table 5.2. Final accuracy (%) across different participation ratios (0.25, 0.5, 0.75, 1.0) in a heterogeneous environment 37
Table 5.3. Final accuracy (%) across different participation ratios (0.30, 0.35, 0.40, 0.45) in a heterogeneous environment 40
Table 5.4. Final accuracy (%) across varying degrees of heterogeneity. 40
Table 5.5. Final accuracy (%) in harsh environments. 42
Figure 1.1. Test accuracy of global model vs. FL training rounds. 14
Figure 3.1. Federated learning system model. 19
Figure 4.1. Update step applying aggregation weights based on the Global-Local gradients distance in the parameter space. 25
Figure 4.2. Update step applying aggregation weights based on the Local-Local gradients distance in the parameter space. 27
Figure 5.1. Data distribution generated using Dirichlet distribution and the sort-and-partition method 34
Figure 5.2. Final accuracy vs. participation ratio (0.25, 0.5, 0.75, 1.0) in a heterogeneous environment. 38
Figure 5.3. Final accuracy vs. participation ratio (0.25, 0.30, 0.35, 0.40, 0.45, 0.50) in a heterogeneous environment. 39
Figure 5.4. Final accuracy vs. heterogeneity. 41
Figure 5.5. Mean and standard deviation of the weighted gradients distance across rounds based on Global-Local gradients distance. 44
Figure 5.6. Mean and standard deviation of the weighted gradients distance across rounds based on Local-Local gradients distance. 45
Figure 5.7. The class distribution of data participating in the training. 46