Algorithm 1 BreastSwimFedNeXt: Federated Swin Transformer Ensemble for Breast Cancer Classification.

Require:  Distributed datasets $\mathcal{X}={\left\{(x_i,y_i)\right\}}_{i=1}^N$ from multiple clients, Swin Transformer variants (Tiny, Small, Base, Large), FL parameters $\theta$    1:Initialize FL parameters $\theta$ and set Swin Transformer variants and RF meta-learner    2:for each client i do    3:    Perform local data preprocessing: apply augmentation (rotation, flipping, brightness, Gaussian noise, blur)    4:     Train local Swin Transformer model to extract features: $$f_{\mathrm{swin}}=\mathrm{SwinTransformer}\left(X\right)$$    5:     Apply MHSA: $$\mathrm{Attention}(Q,K,V)=\mathrm{softmax}\left({\displaystyle \frac{Q{K}^T}{\sqrt{d_k}}}\right)V$$    6:     Optimize with local cross-entropy loss: $$L_{\mathrm{local}}=-\sum ylog\left({\widehat{y}}_{\mathrm{local}}\right)$$    7:end for    8: Server aggregates model weights using FedAvg: $$W_{\mathrm{global}}={\displaystyle \frac{1}{N}}\sum W_{{\mathrm{local}}_i}$$    9: Server constructs stacking ensemble: base predictions from all Swin variants → RF meta-learner $$\widehat{y}={\mathrm{RF}}^{\ast }(f_{\mathrm{SwinTiny}},f_{\mathrm{SwinSmall}},f_{\mathrm{SwinBase}},f_{\mathrm{SwinLarge}})$$    10:return Final prediction $\widehat{y}$