Autism spectrum disease (ASD) is a neurodevelopmental disorder associated with abnormalities in brain development that impact facial appearance. Children with autism differ markedly from typically developed children in the patterns of their facial features. This work aims to provide a unique deep-learning method for diagnosing autism based on face cues utilizing a convolutional neural network with a multi-head attention (CNNMHA) mechanism, making diagnosing autism easier for families and clinicians. It mainly uses two phases such as preprocessing and classification. In this preprocessing step, the proposed system performs image resizing, image denoising by Gaussian filtering, and normalization to enhance the image quality and to make it more suitable for the specific classification task. In the classification stages, the proposed system uses CNNMHA to extract features from the preprocessed image and classifies the images into autistic children or non-autistic children. The data were obtained via Kaggle, and the preliminary computational results demonstrate that the proposed system outperformed existing methods. The technology can be used to help medical professionals validate their first screening results in order to identify youngsters with ASD illness.