A convolutional neural network architecture SpiderNet designed for fraud detection has been proposed. The principles of pooling and convolutional layers in neural networks are very similar to the methods used by antifraud analysts in their research. In addition, the skip-connections used in neural networks make it possible to use features of various power in antifraud models. Our experiments have shown that SpiderNet provides better quality compared to Random Forest, CNN, DenseNet, and F-DenseNet (adapted for antifraud modeling problems) neural networks. Also, new approaches for anti-fraud rules (B-tests and W-tests) have been proposed. The SpiderNet code is available at https://github.com/aasmirnova24/SpiderNet .