Distributed Denial-of-Service (DDoS) attacks pose a significant threat to edge–fog computing infrastructures, where the distributed and resource constrained nature of nodes can be easily overwhelmed by malicious traffic, rendering traditional defense mechanisms insufficient. Deep Learning (DL) provides a promising avenue for building intelligent and adaptive detection systems capable of operating close to the data source. This paper presents a comparative evaluation of four prominent DL architectures: a Deep Neural Network (DNN), a Binary Neural Network (BNN), a Convolutional Neural Network (CNN), and a Long Short-Term Memory (LSTM) network for DDoS detection using the contemporary CICDDoS2019 dataset. A key methodological contribution is the integration of an ensemble metaheuristic feature selection approach (combining Aquila, PSO, and GWO optimizers), which reduced the feature space by 41%. This process not only decreased model complexity but also improved F1-scores and MCC by up to 0.3% while reducing inference times by approximately 10%. Experimental findings show that while the DNN achieves the highest detection accuracy, the BNN delivers highly competitive performance with 27% faster inference, making it particularly suitable for resource-constrained, latency-sensitive edge–fog environments. Furthermore, SHAP analysis revealed that Subflow Bwd Packets, Fwd Packets Length Total and Bwd Header Length were the most influential features in detection. The study concludes that tailored DL models, especially lightweight BNNs combined with intelligent feature selection, provide an effective and computationally efficient strategy for securing next generation edge–fog infrastructures.