Purpose. This study aims to establish the characteristics of noise propagation and accumulation in convolutional neural networks. The article investigates how the accuracy of a trained convolutional network varies depending on the type and intensity of noise exposure.

Methods. White Gaussian noise sources were used as the basis for noise exposure. Two types of noise exposure were applied to artificial neurons: additive and multiplicative. Additionally, the effects of correlated and uncorrelated noise on the layers of neurons were examined.

Results. The findings indicate that additive noise (both correlated and uncorrelated) accumulates more significantly in networks with convolutional layers compared to those without. The relationship between network accuracy and the intensity of multiplicative correlated noise is similar for both types of networks. However, the impact of multiplicative uncorrelated noise is more favorable for networks with convolutional layers. The study also considered pooling layers, specifically MaxPooling and MeanPooling, which significantly enhance accuracy in the presence of additive noise within the convolutional layer. The decline in accuracy due to increasing intensity of multiplicative correlated noise is nearly identical for networks with and without pooling layers. Conversely, networks employing MaxPooling demonstrate reduced resilience to uncorrelated multiplicative noise.

Conclusion. The study demonstrates that additive noise severely degrades network performance when a convolutional layer is present, though this negative effect can be mitigated by including a pooling layer immediately following the convolutional layer. In contrast, the effects of multiplicative noise are less clear-cut. In most cases, its impact remains consistent regardless of the presence of convolution and pooling layers. However, the use of MaxPooling in the pooling layer may compromise the network’s robustness against multiplicative uncorrelated noise.