In response to the current problems in the green finance field, such as low efficiency and weak anti-attack ability of traditional encryption methods due to the large scale, complex structure and strong dynamic nature of data, this paper proposes a green financial data dynamic encryption system based on deep learning, aiming to achieve efficient and secure data protection through artificial intelligence technology. This method first uses a bidirectional long short-term memory network (Bi-LSTM) to perform multi-dimensional feature extraction and anomaly detection on green financial data, and then designs a generative adversarial network (Attention-GAN) that integrates an attention mechanism to dynamically generate differentiated encryption keys. It also combines the improved lightweight AES-256 algorithm to achieve block-parallel encryption, and optimizes the distributed computing efficiency of the model on a GPU cluster through transfer learning. Experimental results show that the system can achieve a single data encryption time of 3.1 ms on a dataset of 100,000 green bond transactions (68% improvement over traditional RSA). In the test of resisting quantum computing attacks, the key space is expanded to 2^512, and the success rate of adversarial attack defense is increased to 99.5%. At the same time, the model only occupies 9.0 MB of storage space after compression, making it suitable for edge computing environments. This research provides an intelligent solution for green financial data security that balances efficiency and reliability, supporting real-time processing of 2000 + TPS transaction data streams.