Digital image watermarking method based on adversarial attack

The invention discloses an image watermarking method based on adversarial attack. The method comprises the following specific steps: (1) establishing a neural network; (2) generating a training set; (3) training a neural network; (4) partitioning the carrier image into blocks; (5) dividing and converting the watermark sequence; (6) generating an adversarial image block by using the adversarial attack to obtain a watermark-containing image; (7) partitioning the image containing the watermark; (8) obtaining a watermark sequence fragment by using a neural network; and (9) splicing the watermark sequence fragments to obtain the watermark sequence. The method has the advantages that on the premise that indexes such as imperceptibility and safety of a watermark algorithm are met, the performance such as embedding capacity can be enhanced along with improvement of the classification capacity of a network model, and meanwhile the algorithm has good robustness to attacks such as filtering attacks, clipping and compression.