Abstract:In response to the challenges posed by dynamic disaster scenarios, such as inefficient model training due to device heterogeneity, limited real-time response capability caused by high communication overhead, and poor performance of the global model on resource-constrained devices due to model drift, FedDis, a federated learning framework based on adaptive distillation is proposed. FedDis employs ResNet-34 as the foundational life-monitoring model and introduces a dynamic knowledge distillation mechanism guided by a distillation feedback loop, effectively addressing model drift caused by device heterogeneity. By incorporating a multi-criteria client selection mechanism, FedDis optimizes communication overhead, and it ensures effective participation of low-computing-power devices through a fairness feedback mechanism. Experimental results demonstrate that FedDis achieves over 60% recognition accuracy within just 10 training rounds, reduces communication overhead by 30%, and outperforms existing federated learning algorithms on the CIFAR-10, CIFAR-100, and MNIST datasets. FedDis provides an efficient and reliable solution for image-based life monitoring in disaster scenarios.