Intrusion detection in modern networks requires models capable of capturing both the spatiotemporal structure of traffic flows and the variability that arises when deployment environments differ from training data. Traditional machine-learning approaches—although strong on tabular intrusion datasets—treat flows as independent samples and exhibit severe degradation under domain shift. Deep neural architectures improve temporal modeling but still struggle to generalize across heterogeneous traffic distributions. This work presents a spatio-temporal deep intrusion detection framework that integrates a CNN–BiLSTM backbone with a multi-head attention mechanism to learn discriminative and interpretable temporal representations. To further enhance robustness in cross-dataset scenarios, the model is extended with a Domain-Adversarial Adaptation Network (DANN), enabling the learned features to become invariant across different traffic domains. Experiments on two benchmark datasets, CIC-IDS2017 and CSE-CIC-IDS2018, show that the hybrid CNN–BiLSTM–Attention architecture delivers competitive in-domain performance while providing more stable recall and F1-score on minority and evolving attack types. When combined with DANN, the model substantially mitigates performance degradation during cross-domain transfers, outperforming both classical machine-learning baselines and deep models without adaptation. The results highlight the effectiveness of adversarial domain alignment for building transferable and reliable intrusion detection systems in heterogeneous real-world network environments.