US Patent 11836249 System and method for counteracting adversarial attacks

Aspects of the present disclosure involve systems, methods, devices, and the like for generating an adversarially resistant model. In one embodiment, a novel architecture is presented that enables the identification of an image that has been adversarially attacked. The system and method used in the identification introduce the use of a denoising module used to reconstruct the original image from the modified image received. Then, further to the reconstruction, an adversarially trained model is used to make a prediction using at least a determination of a loss that may exist between the original image and the denoised image.