Federated learning coordinates model training in a distributed manner within Internet of Things (IoT) systems and ensures the privacy of local client data simultaneously. Nonetheless, traditional federated learning relies primarily on a unified global model and focuses on local feature extraction, failing to accommodate the diversity and personalized needs of clients in non-independent and identically distributed (non-IID) environments. To mitigate the decline in model accuracy posed by these challenges, we propose a discrepancy-aware federated learning method based on multi-branch feature fusion (FedDM). Firstly, we design a differential-aware aggregation strategy (DA), which adjusts the contribution of each client during model aggregation using Gaussian distribution statistics, to generate personalized local models. Next, we propose a multi-branch feature fusion mechanism (MFF) that integrates diverse feature representations through multi-scale pooling and feature enhancement, enabling the incorporation of features across both spatial and channel dimensions for a more holistic representation. Experimental results demonstrate that FedDM enhances model accuracy and robustness, while exhibiting adaptability when facing challenges posed by data distribution heterogeneity.