Existing deep learning models for epileptic electroencephalogram (EEG) signal analysis frequently overlook intrinsic pathological characteristics during feature extraction and exhibit insufficient cross-dataset generalization. To address these limitations, this study proposes an innovative dual-attention epilepsy detection network (EDDANet). The model integrates a multi-band and multi-scale dual-attention module with a dynamic kernel sampling adaptive convolutional module to classify interictal and ictal EEG signals. Extensive experiments conducted on four heterogeneous public datasets demonstrate that EDDANet consistently outperforms state-of-the-art models across key evaluation metrics, including accuracy and recall. Notably, this work is the first to achieve robust generalization across varying lead configurations, sampling rates, and electrode layouts. In conclusion, this study provides a valuable methodological framework for the design and optimization of automated epilepsy detection systems in complex scenarios, providing reference for enhancing the generalizability and clinical utility of deep learning models in real-world environments.