Cloud computing has evolved with the merits of flexibility, scalability and increased availability by migrating the computing infrastructure to the network such that software and hardware monitoring can be achieved with reliability. The dynamic allocation and utilization of network resources which improves the efficiency of cloud infrastructure can be achieved through the incorporation of Software-Defined Networking (SDN) based cloud networking. But these SDN-based clouds are vulnerable to Distributed Denial-of-Service (DDoS) attacks, which prevents the genuine users from utilizing the required degree of subscribed services. Hence DDoS attacks in SDN based clouds need to be detected at the earlier stage with least amount of detection time for the objective of sustaining network performance. In this paper, a deep learning DDoS defense model using Self-Attention-based CNN-BiLSTM-Transformer is proposed for real time detection and mitigation of DDoS attacks at an earlier stage in SDN-based clouds. This deep learning model uses the merits of self-attention-based CNN-BiLSTM-Transformer for determining the statistical measure of entropy associated with the network traffic. It then makes use of gazelle optimization algorithm for selecting the optimal features which enhances the detection potentiality of the adopted deep learning models by facilitating accurate classification of normal and malicious traffic. The simulation experiments conducted with respect to CIDDS-001 traffic dataset confirmed minimized mean square error and mean absolute percentage error (MAPE) with maximized accuracy and precision under different forecasting entropy of source IPs compared to the benchmarked CNN-LSTM-transformers, CNN-LSTM, and LSTM.