<p>Federated learning enables multiple clients to collaboratively train a global model using their respective local datasets. While it offers advantages such as privacy preservation and efficient data utilization, its practical deployment is still constrained by heterogeneous client computing capabilities and unstable operating environments. To address these challenges, this work proposes FedRAS, a dual-strategy federated framework that combines resource-aware partial training with attention-guided self-distillation. FedRAS dynamically allocates trainable model components according to client resource availability and aligns local and global model representations via a self-distillation mechanism, thereby improving federated learning performance along two dimensions: cross-client collaboration and local-global model alignment. Experimental results on the public dermoscopy dataset ISIC 2019 show that, for the same computation and communication overhead, the partial training strategy in FedRAS delivers significantly higher accuracy than various partial training methods, while its self-distillation strategy also surpasses multiple knowledge distillation baselines in the overall balance between accuracy and overhead.</p>