Learning to forget specific concepts in text-to-image diffusion models is crucial for privacy protection. While existing methods achieve some performance success, they often suffer from partial forgetting, retaining knowledge of the specific concept and enabling unintended generation. In this work, we propose DCF (Deep Concept Forgetting), which leverages textual inversion to exploit and forget the relationship between text and image. To achieve this, we reverse the function of textual inversion (generation \(\rightarrow \) forgetting) and design a reversed inversion loss to fine-tune cross-attention neurons. Additionally, a regularization loss is introduced to prevent excessive forgetting. Theoretically, we show that DCF approximates retraining without the target concept. Extensive experiments show that our method achieves: 1) Deep forgetting with great maintaining ability, at all levels outperforming previous methods in forgetting target concepts while preserving others and maintaining high generation quality; 2) Efficiency, with reduced training time and memory costs; and 3) Illegal-oriented protection, generating blank-like images without any concept-related information to prevent illegal or malicious uses.