For several decades, the detection of epileptic seizures has been an active research topic. The performance of current patient-specific algorithms is satisfactory. However, due to significant variability of EEG data between patients, cross-subject seizure characterization and detection remains a challenging task. The purpose of this study is to propose and investigate a modified convolutional neural network (CNN) architecture based on separable depth-wise convolution for effective automatic cross-subject seizure detection. The architecture is conceived with a reduced number of trainable parameters to reduce the model complexity and storage requirements to easily deploy it in connected devices for real-time seizure detection. The performance of the proposed method is evaluated on two public datasets collected in the Children’s Hospital Boston and the University of Bonn respectively. The method achieves the highest sensitivity-false positive rate/h of 91.93%–0.005, 100%–0.057 for the CHB-MIT and Ubonn datasets respectively.