This paper presents an improved method for estimating the accuracy of a model based on images intended for prediction, enhancing the standard Detection of Classifier Inconsistencies (DCI) method. The conventional DCI method typically requires a large enough set of images from the same source to provide accurate estimations, which limits its practicality. Our enhanced approach overcomes this limitation by generating a set of images from a single original image, thereby enabling the application of the standard DCI method without requiring more than one target image. This method ensures that the generated images maintain the statistical properties of the original, preserving any embedded steganographic messages, through the use of non-destructive image manipulations such as flips, rotations, and shifts. Experimental results demonstrate that our method produces results comparable to those of the traditional DCI method, effectively estimating model accuracy with as few as 32 generated images. The robustness of our approach is also confirmed in challenging scenarios involving cover source mismatch (CSM), making it a viable solution for real-world applications.

In this paper, we propose a robust method for detecting guilty actors in image steganography while effectively addressing the Cover Source Mismatch (CSM) problem, which arises when classifying images from one source using a classifier trained on images from another source. Designed for an actor-based scenario, our method combines the use of Detection of Classifier Inconsistencies (DCI) prediction with EfficientNet neural networks for feature extraction, and a Gradient Boosting Machine for the final classification. The proposed approach successfully determines whether an actor is innocent or guilty, or if they should be discarded due to excessive CSM. We show that the method remains reliable even in scenarios with high CSM, consistently achieving accuracy above 80% and outperforming the baseline method. This novel approach contributes to the field of steganalysis by offering a practical and efficient solution for handling CSM and detecting guilty actors in real-world applications.

Steganalysis is a collection of techniques used to detect whether secret information is embedded in a carrier using steganography. Most of the existing steganalytic methods are based on machine learning, which requires training a classifier with “laboratory” data. However, applying machine-learning classification to a new data source is challenging, since there is typically a mismatch between the training and the testing sets. In addition, other sources of uncertainty affect the steganlytic process, including the mismatch between the targeted and the actual steganographic algorithms, unknown parameters –such as the message length– and having a mixture of several algorithms and parameters, which would constitute a realistic scenario. This article presents subsequent embedding as a valuable strategy that can be incorporated into modern steganalysis. Although this solution has been applied in previous works, a theoretical basis for this strategy was missing. Here, we cover this research gap by introducing the “directionality” property of features concerning data embedding. Once a consistent theoretical framework sustains this strategy, new practical applications are also described and tested against standard steganography, moving steganalysis closer to real-world conditions.

In this paper, a methodology to detect inconsistencies in classification-based image steganalysis is presented. The proposed approach uses two classifiers: the usual one, trained with a set formed by cover and stego images, and a second classifier trained with the set obtained after embedding additional random messages into theoriginal training set. When the decisions of these two classifiers are not consistent, we know that the prediction is not reliable. The number of inconsistencies in the predictions of a testing set may indicate that the classifier is not performing correctly in the testing scenario. This occurs, for example, in case of cover source mismatch,or when we are trying to detect a steganographic method that theclassifier is no capable of modelling accurately. We also show how the number of inconsistencies can be used to predict the reliability of the classifier (classification errors).