动机
Many previous works observed that the well-generalized solutions are located in flat areas rather than sharp areas of the loss landscapes.
通常，由于每个本地模型的任务是相同的，因此每个客户端的损失情况仍然相似。直观上，与尖锐的最佳区域相比，不同客户端的平坦最佳区域部分重叠的可能性更大。换句话说，当模型收敛到重叠区域时，它可以在大多数客户端中实现高推理性能。
怎么做能达到效果？
FedMut框架
global model ：
ω g l b = { l a y e r 1 , l a y e r 2 , . . . , l a y e r L } . \omega_{glb} = \{ layer_1, layer_2,..., layer_L \}. ωglb={layer1,layer2,...,layerL}.
gradients of ω g l b \omega_{glb} ωglb :
g g l b = { l g 1 , l g 2 , . . . , l g L } . g_{glb} = \{lg_1, lg_2,..., lg_L\}. gglb={lg1,lg2,...,lgL}.
mutation weight list L m u t i L^{i}_{mut} Lmuti:
L m u t i = { v 1 i , v 2 i , . . . , v L i } . L^{i}_{mut}=\{v^i_1, v^i_2,..., v^i_L\}. Lmuti={v1i,v2i,...,vLi}.
mutated model ω m u t i \omega^{i}_{mut} ωmuti
ω m u t i = { l a y e r 1 + α v 1 i ⋅ l g 1 , . . . , l a y e r L + α v L i ⋅ l g L } \omega^{i}_{mut} = \{layer_1+\alpha v^i_1 \cdot lg_1 ,..., layer_L+\alpha v^i_L \cdot lg_L\} ωmuti={layer1+αv1i⋅lg1,...,layerL+αvLi⋅lgL}
算法
