The use of machine learning algorithms in medical applications allows for fast and accurate diagnosis of diseases. Autism spectral disorder (ASD) is one of the common mental disorders and the importance of early diagnosing attracted researchers to use different machine learning-based methods. In this paper, we aim to classify ASD from non-ASD using the information of resting-state functional magnetic resonance imaging (rs-fMRI) multisite data. In the proposed method, at first, each region of interest (ROI) of data of each subject is decomposed using the double-density dual-tree discrete wavelet transform (D3TDWT) into time-frequency sub-bands. In the second step, generalized autoregressive conditional heteroscedasticity (GARCH) model is used for feature extraction from these sub-bands. Next, the discriminative features are selected by two-sample t-test and finally, the data are classified by support vector machine. The algorithm is tested on several datasets. The results validate the robustness of the proposed method by obtaining 71.6% classification accuracy for male subjects and 93.7% accuracy rate for female subjects. By considering the significant ROIs, Middle Temporal Gruys, Supramarginal Gyrus, and Paracingulate Gyrus, there is a reduction in anterior-posterior connections among ASDs, which can be considered in clinical approaches. The proposed method outperforms other methods.