Security researchers have devised a technique to alter deep neural network outputs at the inference stage by changing model weights via row hammering in an attack dubbed ‘OneFlip.’ A team of ...