In existing deep learning image steganography methods, the neural network structure contains ineffective or complicated standard convolutional layers. These convolutional layers do not significantly improve performance, but instead increase computational overhead. And over-parameterization will lead to over-fitting, which limits the ability to express image edges and texture features, resulting in poor steganography effect. This paper proposes an improved deep network, which optimizes the network structure by adjusting the number of layers and groups of standard convolutional layers, and adds random tensor noise, gaussian noise, and salt-and-pepper noise attacks in the coding fusion process. Different from other works on noise attacks on encoded images, our system is not only resistant to noise attacks added after encoding, but also resistant to the above three types of noise attacks during the encoding fusion process. Experimental results on Tiny ImageNet and CIFAR-10 datasets show that the model has high-quality hiding and extraction effects and is well resistant to noise attacks during the encoding process.