Robustness of Neural Networks on Optimized Synthetic Images

Sammanfattning

Artificial neural networks have proven to be vulnerable even to small perturbations in theirinput. In order to prevent noise or maliciously handcrafted perturbations from eliciting undesir-able behavior in the network, one must understand what makes a network more or less sensitiveto these perturbations in its input. This is the study of robustness. Our main contribution isinvestigating the relationship between neural network classification confidence and robustness bysuggesting models for generating images that maximize the output of neural networks in somedifferent senses. We test if these generated images are more robust than normal images. Wealso investigate factors that contribute to making a generated image more robust, and conductadjacent analysis. We show that images generated by what we call selective output maximiza-tion are indeed more robust compared to normal images. The extent to which such generatedimages are more robust varies greatly depending on the type of image class, and less so, butnot insignificantly, on the architecture of the network that generated the images and the typeof selection function that is used to generate the images. We also investigate a concept we callrobustness transferability, a generated image’s propensity to perform robustly on networks otherthan the one which generated it. We find that the generated images are also more robust onother networks that did not generate them compared to normal images, but less robust comparedto the network that generated the images. We also propose an image generation method whichsimultaneously uses multiple neural networks to generate images. Tests hint at these so calledmulti-models resulting in slightly higher robustness transferability of their generated images