Generative-adversarial-network-based steganographic methods typically create images through adversarial confrontations between a generator and discriminator. However, these steganographic methods are characterized by the specificity of their structures, which makes them prone to generating large numbers of variants that are not easy to detect. Unfortunately, limited research has focused on detecting variants of generative-adversarial-network-based steganographic images. This chapter describes a framework for detecting generative-adversarial-network-based steganographic image variants. The framework is developed by controlling and guiding the features of images in the middle layer of a neural network and by constructing image samples that capture the common characteristics of generative-adversarial-network-based steganographic images. A detector trained on these images has improved generalization ability when faced with generative-adversarial-network-based steganographic image variants. Experimental results demonstrate the effectiveness of the approach compared with other generative-adversarial-network-based models.