Methods for Preventing Visual Attacks in Convolutional Neural Networks Based on Data Discard and Dimensionality Reduction

The article is devoted to the study of convolutional neural network inference in task image processing under influence visual attacks. Attacks four different types were considered: simple, involving addition white Gaussian noise, impulse action on one pixel an image, and attacks that change brightness values within a rectangular area. MNIST Kaggle dogs vs. cats datasets chosen. Recognition characteristics obtained for accuracy, depending number images subjected used training. was based well-known architectures pattern recognition tasks, such as VGG-16 Inception_v3. dependencies accuracy parameters obtained. Original methods proposed prevent Such are selection “incomprehensible” classes recognizer, their subsequent correction with reduced sizes. As result applying these methods, gains metric by factor 1.3 after iteration discarding incomprehensible images, reducing amount uncertainty 4–5% integration results analyses dimensions.