Optimus Prime: Linear Transformations to Secure Machine Learning Systems