Gnothi Seauton: Empowering Faithful Self-Interpretability in Black-Box Models