Considerations on the Theory of Training Models with Differential Privacy