Many new products nowadays were made possible by advances in machine learning, in particular image recognition and automatic speech recognition. Examples are spam filters, virtual personal assistants, traffic prediction in GPS devices, or face recognition. Real-world machine learning applications see more widespread use and enjoy ever increasing accuracy on common benchmark datasets. Unfortunately, it has been observed that virtually all models return incorrect results if data is fed to the model that is not from common benchmark datasets, data that was purposefully but imperceptibly altered, and plain garbage data. In this blog post I will talk about the purposeful, imperceptible input modifications, so-called adversarial examples. I will present possible kinds of modifications in computer vision and requirements for successful attacks before discussing attempts to improve models. Finally, I will make two predictions about the future of machine learning models. The appendix contains remarks about the misinterpretation of the largest eigenvalue of a weight matrix; the modulus of this value is almost meaningless without considering the smallest eigenvalues, too.
Continue reading The Fundamental Importance of Adversarial Examples to Machine Learning