This study proposes an Adaptive Image Steganography Algorithm based on Deep Reinforcement Learning(AISA-DRL)to realize a balance among lightweight steganography,optimal embedding positioning,and high hidden outputs.A lightweight secure steganography network is designed to enhance the ability of the model to extract stego features of an image based on the premise of reducing the steganography cost and optimizing the security and stability of the stego image.First,an EPSANet with efficient feature fusion characteristics is introduced into EfficientnetV2-s to obtain an improved EPSA-EfficientnetV2-s.This integration enhances the representation ability of the pixel-level embedding process to obtain the optimal tensor for the pixel modification bit.The stego image is then calculated using the weighted sum of the secret information and the optimal tensor of the pixel modification.Finally,by learning the steganographic analysis network,the optimal pixel-level reward is assigned to the stego image,and the network parameters are updated by gradient backpropagation according to the designed minimum distortion function.This process ensures that the best embedding position is obtained,and the optimal embedding of the secret information is realized.The experimental results show that the parameters of the AISA-DRL are reduced by 94.22％,and FLOPs are reduced by 24.88％.Compared with steganographic methods based on reinforcement learning,the error detection rate of the AISA-DRL increases by 2.48％-6.55％.Additionally,the PSNR values of the stego images generated under different loads are all greater than 30 dB.Thus,the proposed AISA-DRL improves the locational accuracy of the model for the pixel modification bit.The steganographic network of the AISA-DRL demonstrates a stronger characterization capability compared to other algorithms.
维普
万方数据