Re-thinking Model Inversion Attacks Against Deep Neural Networks