Adding noise is important for generalization and predictive power but it may just shift the attacking pixel from one to another, since the model itself is inher

Adding noise is important for generalization and predictive power but it may just shift the attacking pixel from one to another, since the model itself is inherently compressed, inversely correlated with its complexity.