Abstract: Steganography is to conceal the presence of secret communication. Traditional steganographic schemes rely on complex artificial rules that are difficult to construct. Steganalysers based on the rich models and deep learning achieve state-of-the-art performance. The security performance of existing steganographic methods is being challenged. In this paper, a search method based on image steganography model against attack is proposed to find a suitable steganography policy. The steganographic model constructs the parametric policy. The adversary model distinguishes the distribution of stego from cover to find the potential hiding artefacts. To obtain the corresponding evaluations, the adversarial attack is performed on adversary model. The security game between steganographic part and adversary is established via corresponding information, thus finding the target steganographic policy. The steganographic model and adversary model are implemented as deep neural networks. On the data set Bossbase, the payload is 0.2 and 0.4 bpp, the steganalysers are SRM and maxSRMd2. Four configurations with three steganographic schemes are compared. The experimental results show that the scheme proposed in this paper can obtain effective policy for image steganography, and the security performance is competitive compared with these schemes.