Neural network based image classifier resilient to destructive perturbation influences – architecture and training method

Abstract

Modern methods of image recognition are sensitive to various types of disturbances, which actualize the development of resilient intelligent algorithms for safety-critical applications. The current article develops a model and method of training a classifier that exhibits characteristics of resilience to adversarial attacks, fault injection, and concept drift. The proposed model has a hierarchical structure of prototypes and hyperspherical boundaries of classes formed in the space of high-level features. Class boundaries are optimized during training and provide perturbation absorption and graceful degradation. The proposed learning method involves the use of a combined loss function, which allows the use of both labeled and unlabeled data, implements the compression of the feature representation to a discrete form and ensures the compactness of the distribution of classes and the maximization of the buffer zone between classes. The main component of the loss function is the value of the normalized modification of Shannon's information measure, averaged over the alphabet of the classes, expressed as a function of accuracy characteristics. Simultaneously, accuracy characteristics are calculated on the basis of smoothed versions of the distribution of statistical hypothesis testing results. It is experimentally confirmed that the proposed approach provides a certain level of disturbance absorption, graceful degradation and recovery. During testing of the proposed algorithm on the Cifar10 data set, it was established that the integral metric of resilience to tensor damage by inversion of one randomly selected bit is about 0.95 if the share of damaged tensors does not exceed 30%. Also, during testing of the proposed algorithm, it was established that an adversarial attack with a disturbance that does not exceed the L∞-norm threshold equal to 3 provides resilience that exceeds the value of 0.95 according to the integral metric. Additionally, the integral metric of resilience during adaptation to the appearance of two new classes is 0.959. The integral metric of resilience to the real drift of concepts between the two classes is 0.973. The ability to adapt to the appearance of new classes or the concept drift has been confirmed 8 times faster than learning from scratch.