Abstract: By using the compute unified device architecture (CUDA) and cuckoo search algorithm (CSA), this paper proposes an improved support vector machine algorithm, termed as CCS-SVM, to improve the classify speed and accuracy of classical SVM algorithm. SVM is one of the most popular classifiers for the classification process, but its training is very computationally intensive for large scale data, scales badly with the size of the data sets. Aiming at the slow training speed of SVM in large-scale data, this paper proposes a parallel SVM algorithm to accelerate training process by means of CUDA technology. Especially, the computations of SVM kernel function and the parallel SMO algorithm are implemented on the GPU. The classification accuracy of SVM is closely related to the setting of kernel function parameters, e.g., gamma (γ) for the radial basis function (RBF) kernel and penalty parameter C. In this work, CSA will be utilized to optimize these parameters. CSA is a recently developed meta-heuristic optimization algorithm and is quite suitable for solving optimization problems. The traditional CSA uses two fixed value parameters p_α and α, which may degrade its performance and increase the iteration times under the case that p_α is small and α is large. Moreover, CSA may not find the optimal solutions, although its convergence speed may be very fast. In order to overcome the above shortcoming, this paper makes two improvements. Firstly, the influence of individual fitness on step size factor α is considered during the optimization process. Secondly, the inertia weight is introduced into the random walk. Finally, the proposed CCS-SVM algorithm is verified via the intrusion detection simulation experiment on industrial network standard dataset, which shows that the proposes algorithm can improve the detection speed by almost 3 times while guaranteeing the accuracy of intrusion detection.