Internet of Things (IoT) technology has evolved rapidly to become an integral part of our daily life. The numerous number of IoT devices and the limited security measures that could be applied on such devices attracted attackers to exploit these devices to act as botnets, generating massive Distributed Denial of Service (DDoS) attacks. IoT botnet-based DDoS attacks are growing dramatically both in frequency and sophistication and thus call for urgent development of powerful detection mechanism and deploy those mechanisms in an efficient way.In this paper, we propose attention-based deep learning approach for IoT botnet-based DDoS attacks. Specifically, we apply deep learning attention mechanism on two popular deep learning models LSTM and CNN for detection of IoT botnet DDoS attacks. We also examine the effect of using the deep Autoencoder for feature reduction and examine its effect on training time and data size needed to be exchanged through the network for training and updating the model, especially for solutions deployed at the Edge computing network. Performance evaluation shows that combining attention mechanism with LSTM model achieves superior performance with an accuracy reaching 100% which is due to combining the memorizing ability by the LSTM and the focusing ability of the attention in one model. In addition, results show that using Autoencoder for feature reduction resulted in reducing the training time up to 66.25%, 64.39%, 72.30% and 78.64% for ALSTM, LSTM, ACNN, CNN models respectively. Also achieved reduction in size of the dataset up to 67%.