A Survey of Deep Image Steganography with Invertible Networks

This survey presents a systematic review of image steganography and image hiding that use invertible neural networks. We first explain how a bijection with a single parameterization treats embedding as the forward mapping and recovery as the inverse mapping. Under a unified evaluation protocol, we analyze four dimensions, namely capacity, imperceptibility, security and detectability, and reversibility. We then present a concise taxonomy that covers single image methods, multiple image methods, and lossless channels at the message channel (Lossless INN). Comparisons across scenarios indicate that, for high payloads in single image settings, HiNet is generally more stable than ISN; in multiple image settings, DeepMIH offers stronger interpretable control when the number of secrets is moderate, whereas InvMIHNet shows better scaling of capacity and efficiency when the number is large. Finally, we summarize common issues, such as frequency leakage, robustness against steganalyzers, and protocol inconsistency, and we outline a practice oriented agenda including boundary focused spectral regularization with adaptive partitioning, mosaic assisted cascades, dual channel cooperation between the message path and the pixel path, and a Minimal Comparable Set benchmark with accompanying theoretical bounds.