Abstract: DNA steganography is an important technique for ensuring information security and communication privacy. However, with the advancement of technology, traditional DNA steganography methods gradually reveal limitations in steganographic performance and undetectability. These methods generally overlook the statistical characteristics of DNA sequences and lack robust evaluation mechanisms, thereby restricting the quality and security of steganographic data. To address these issues, a deep learning-driven DNA adaptive dynamic grouping steganography method (DAGS) is proposed. This method integrates deep learning, adaptive dynamic grouping strategies, and multidimensional dynamic evaluation mechanisms. The statistical characteristics of natural DNA sequences are captured using convolutional neural networks and bidirectional LSTM networks. Information hiding is performed through an adaptive dynamic grouping strategy, and the generated sequences are evaluated in real time during training for statistical similarity to natural DNA sequences. Key statistical metrics such as GC content distribution, melting temperature deviation, and KL divergence are employed to optimize the quality and security of steganographic data. Experimental results demonstrate that the generated sequences using this method exhibit closer statistical properties to natural DNA sequences, significantly improving the undetectability of hidden information. Furthermore, the proposed method outperforms existing approaches in both steganographic performance and security, confirming the application potential of DAGS in DNA steganography.