Self-Deception: Reverse Penetrating the Semantic Firewall of Large Language Models