Software defect prediction researches aim to find out potential risks in software projects by identifying files and dependencies. Deep learning models have been presented in defect prediction research, resulting in considerable achievements. Meanwhile, the recent growth in edge computing eases the high hardware requirement of machine learning methods. However, a number of irrelevant features in defect prediction data bring challenge to the performance of defect prediction models, while the lack in effective fusion of features decreases prediction accuracy. In order to improve these existing problems in deep learning defect prediction methods, we propose a defect prediction model suiting for edge computing via an attention-based gated recurrent unit neural network, in which the Bayesian algorithm is introduced for optimization. The model extracts both classic features from repository and semantic features from abstract syntax trees, fusing these features for comprehensive information. Furthermore, the attention component in our model guides the network to dynamically focus on key features for the defect prediction mission, thus overcome the issue of irrelevant features. In addition, the model is available for edge computing environment through federated learning. The experiment outcomes indicate that our method has comparatively good performance on software defect prediction dataset.