Abstract: In order to solve the problem of less redundant space in database watermarking algorithm, and the databases' availability and authenticity affected by the modification of data, a large-capacity database watermarking algorithm based on generative networks and convex packages is proposed. The algorithm used LSTM and evaluation functions to generate pseudo-tuples, and the convex-pack algorithm to distinguish between real and pseudo-tuples. By inserting flag bits into the watermark information, the error in the watermark could be positioned between two flag bits, which reduced the impact of the attack on the watermark extraction. Experimental results show that the algorithm has good invisibility, and the ability to resist database watermark attacks such as data modification, deletion and addition, which is better than that of reversible watermarking algorithms.