01.10.2025
5 min
457
The article discusses the basic approaches, the operation of the generator and discriminator, as well as the challenges associated with discovering such methods. It is useful material for cybersecurity researchers, IT professionals, and anyone interested in the future of artificial intelligence in the field of information hiding.
Steganography is defined as the science of covert data transmission, where the main goal is to disguise the very fact of information exchange. The term itself comes from the Greek words стеганоς — “hidden” and γράφω — “I write”, which literally translates as “secret writing”. In the modern development of this discipline, machine learning technologies, in particular generative adversarial networks (GANs), play an increasingly important role, which open up new approaches to hiding messages.
Classical steganography has long served as a means of secret communication — examples include invisible ink or hidden messages in poetry. In modern conditions, this role is performed by electronic communications and digital technologies. Any modern stegosystem consists of two key elements: embedding and extraction algorithms. The embedding algorithm receives as input a secret message, a key and a container into which the data will be integrated. The result is a stegomessage.
The hidden message can be any type of information – text, graphics, sound, etc. Each format has its own methods of concealment. For example, in text steganography, rules are applied regarding the position and number of characters, which are invisible to the average reader. One method is formatting: widening the gaps between words, where a single space indicates a bit “0”, and a double space indicates a bit “1”. In network steganography, hidden data can be transmitted over computer networks using the features of protocols. Typical approaches involve modifying the properties of individual network packets.
Within the framework of this review, attention is focused on image steganography. In this area, three fundamental architectures are distinguished that determine the methods of obtaining stegoimages.
There are basically two types of approaches to steganography by modification. The first type is associated with the support of a statistical model, and the second implements embedding as a result of minimizing a specific distortion function.
Image selection involves choosing a suitable container for hiding information and is based on two main approaches. The first is to select a candidate image for further modification. In this case, a container is first determined from the database, after which the message is embedded in it. Despite the fact that the methods of this class belong to the category of “image selection”, in fact they are a modification with an additional stage, and therefore are not considered as an independent direction.
The second approach is to use the container as a ready-made stegoimage without any changes. In this case, the key task becomes the definition of clear rules of correspondence between the transmitted message and the selected stegoimage.
The third approach in steganography is related to image synthesis. Since creating realistic graphical objects is a difficult task, classical methods implemented this strategy by generating “unnatural” images. Examples of such methods include artificially created textures or images that imitate fingerprints.
Generative adversarial network (GAN) is an unsupervised machine learning algorithm based on a combination of two neural networks. It was first introduced by Google in 2014. In a GAN, one of the networks generates samples, while the other tries to distinguish “real” samples from false ones.
The generative network, using a set of latent space variables, always tries to create a new sample by mixing several original samples. The discriminative network D learns to distinguish between real and fake samples. Its results are fed to the input of the generative network G so that it can choose the best set of latent parameters and the discriminative network would no longer be able to distinguish real samples from fake ones. As we can see, the goal of the G network is to increase the error rate of the D network, and the goal of the D network is, on the contrary, to improve the recognition accuracy.
GANs have found great applications in computer vision and natural language processing (in image and text generation). But they have also been used in steganography.
The capabilities of GANs can be viewed from different angles: a game, a generator, or a mapping function. They are consistent with the classification of the main strategies in steganography, i.e. modification, synthesis, and selection.
GAN approaches are based on a “game” between a steganographer and a stegananalyzer. Three key strategies are distinguished:
Stego-image creation — the generator creates an image that contains a hidden message and is difficult to detect. The scheme uses three networks: generator G, discriminator D, and classifier S.
Probabilistic modification matrix — an example is ASDL-GAN, which minimizes the distortion function and determines the probability of changing each pixel.
Competitive game — the ADV-EMB method, where the image is changed in such a way as to confuse the analyzer. Some of the pixels are used to hide the data, and the rest are used to create interference.
Here, the GAN establishes a correspondence between the message and a ready-made image. The container itself is converted into a stegoimage without direct modification, so classical statistical analysis almost does not work.
A pre-trained generator is used to create new images. One approach is supervised learning, where neural networks play the role of Alice, Bob, and Eve, reproducing the process of message exchange and their hiding.
In this approach, the message is not embedded directly into an existing image, but is converted into a noise vector that is fed to the generator. The generator is pre-trained on a large sample, after which it is able to create realistic images. At the next stage, an extractor is used, which is specially trained to restore messages from already synthesized stegoimages. Thus, an additional level of protection appears: instead of making changes to the already prepared images, a completely new image is created in which the message is “sewn” into. This makes detecting the hidden channel much more difficult.
The method is based on water-gradient GANs (Wasserstein GAN with Gradient Penalty), which are famous for their stability of training and the quality of synthesized data. In this case, the generator and the message extractor are trained simultaneously. They work in the format of a competitive game: the generator strives to create the most realistic stegoimages, and the extractor learns to extract hidden messages, despite the complexity of the conditions. Additionally, a discriminator is connected that controls the plausibility of the received images. As a result, a balance is achieved between the image quality and the accuracy of data transmission, which allows to increase the hidden bandwidth of the channel.
ACGAN (Auxiliary Classifier GAN) modifies the classic GAN by adding an auxiliary classifier. The idea is not only to create realistic images, but also to teach the system to establish a connection between class labels and secret information. The generator receives not only noise, but also the corresponding class label as input, which allows you to embed the message directly into the logic of building the image. The discriminator, in addition to determining the authenticity of the image, performs an additional task – restores parts of the hidden message. This approach increases the accuracy of data reproduction and makes the system more manageable.
Steganography with limited sampling
This method focuses on minimizing the changes between the original image and the created stegoimage. In essence, hiding a message becomes a task of finding the most suitable container. The system generates a large number of candidates, and then chooses the one in which the embedded message changes the original image the least. This reduces the probability of detecting a hidden channel, since the differences are statistically close to the noise that is natural for any digital file.
This approach uses the well-known CycleGAN model, developed for image-to-image transformations without paired examples. It learns to transform images from one class (X) to images of another class (Y), while maintaining the consistency of the reverse transformation (cyclicity). In the context of steganography, this allows us to consider the transformation as a process of encoding a message. For example, a regular image can be transformed into a stylized or modified one, but at the same time additional data is hidden in it. This approach is characterized by high flexibility and complexity for steganalysis, since the image looks plausible and corresponds to the natural patterns of the target class.
GAN-based methods are currently considered the most promising direction in steganography development. They allow creating stegimages that are almost indistinguishable from real ones, but they have a number of limitations and open questions. Further development of this technology focuses on three main areas:
Current implementations like GAN-CSY suffer from pixel instability, which results in poor message extraction accuracy. In many cases, the data is not stored in its classical form, but is transformed into category attributes or noise vectors. This limits the amount of information that can actually be hidden.
In the future, an important direction will be:
stabilization of image generation;
development of mechanisms to increase channel throughput;
finding a balance between image quality and the amount of data transmitted.
The problem remains the lack of objective metrics. Currently, IS (Inception Score) and FID (Frechet Inception Distance) are actively used, which measure plausibility and quality. But these indicators are not enough, because they do not take into account the hidden load and the specificity of stegoimages.
In the future, we should expect:
creation of new metrics that will combine the assessment of image quality with the reliability of message transmission;
application of combined approaches – automated and with elements of human expertise;
development of standardized assessment tools for scientific and applied research.
Today, GANs are capable of producing images that are indistinguishable to the human eye. Classical steganography uses numerous analysis methods to detect hidden changes, but as generative models become more powerful, these methods become less effective.
Further research is aimed at:
development of tools capable of distinguishing between real and synthetic images;
improvement of systems for recognizing hidden messages in generated containers;
combination of classical steganalysis with deep neural networks to increase accuracy.
The future of GAN-based steganography is related to solving three big problems: increasing capacity, developing new evaluation metrics, and strengthening steganalysis. The balance between concealment capabilities and the ability to detect hidden data will determine the development of the entire industry in the coming years.
