Enhancing Reliability of Neural Networks at the Edge: Inverted Normalization with Stochastic Affine Transformations