Differential Privacy has being proposed by some authors to be effective for preserving privacy in Machine learning (or deep learning). I will like to know if it has a deficiency or limitations as regards to accuracy in deep learning models .