This paper describes an approach to early detection of network attacks using weight agnostic neural networks. The choice of the type of neural networks is due to the specificity of their architecture that provides high processing speed and performance, which is significant in solving the problem of early attack detection. Experimental studies have demonstrated the effectiveness of the proposed approach based on a combination of multiple regression for feature selection of the training sample and weight agnostic neural networks. The accuracy of attack detection is comparable to the best results in the field with a significant time gain.