US Patent 10915631 Deep learning on execution trace data for exploit detection

Technologies disclosed herein provide for converting a first data of a first control flow packet to a first pixel, where the first data indicates one or more branches taken during a known execution of an application, generating an array of pixels using the first pixel and one or more other pixels associated with one or more other control flow packets generated from the known execution, transforming the array of pixels into a series of images, and using a machine learning algorithm with inputs to train a behavior model to identify a malicious behavior in an unknown execution of the application. The inputs include one or more images of the series of images and respective image labels assigned to the one or more images. More specific embodiments include extracting the first control flow packet from an execution trace representing at least part of the known execution.