<p>Cloud computing is a promising technology for both individual users and organizations, as it provides accessible, internet-based computing resources as a flexible service. However, due to its distributed and open nature, cloud environments are highly vulnerable to security threats. Therefore, the Intrusion Detection System (IDS) is employed as an effective solution for preventing attacks by classifying network behavior as normal or suspicious. Traditional IDS approaches suffer from various issues including neuron death, vanishing gradient, and ineffective feature generalization. To address these issues, a Self-Gated Rectified Linear Unit (SGR) based Convolutional Neural Network (CNN) is proposed for developing an effective IDS. The SGR activation function is used to eliminate the problems of output offset, neuron death, and vanishing gradient. It also introduces sparsity during the learning process, which improves generalization performance. Additionally, the most relevant feature subset is identified through chi-square feature selection, based on the computation of statistical dependency among the features. As a result, the developed SGRCNN-IDS effectively provides security against Advanced Persistent Threats (APT) in cloud computing architecture. Two datasets namely, CIC-IDS2017 and CSE-CIC-IDS2018 are used to evaluate the proposed SGRCNN-IDS. The proposed SGRCNN-IDS is evaluated based on its accuracy, precision, recall and F1-score. The existing approaches, namely BMRF-RF, Two-phase IDS and RC-NN are used for comparison with SGRCNN-IDS. The accuracy of SGRCNN-IDS on the CIC-IDS2017 dataset is 99.78% which is higher than that of BMRF-RF and Two-phase IDS.</p>