[Paper Review] Real-world actor-based image steganalysis via classifier inconsistency detection

Daniel Lerch-Hostalot, David Megías|arXiv (Cornell University)|Jan 8, 2025

Advanced Steganography and Watermarking Techniques1 citations

TL;DR

This paper proposes a robust actor-based steganalysis method that detects guilty actors in real-world image steganography by leveraging classifier inconsistency detection (DCI) with EfficientNet for feature extraction and a Gradient Boosting classifier for final decision-making. The method maintains over 80% accuracy even under 100% cover source mismatch (CSM), outperforming baselines that degrade to random guessing, and effectively identifies and discards CSM-affected actors to prevent misclassification.

ABSTRACT

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.

Motivation & Objective

• To address the real-world challenge of cover source mismatch (CSM) in image steganalysis, where training and testing images come from different sources.
• To develop a practical, actor-based steganalysis framework that classifies actors as innocent, guilty, or unclassifiable due to excessive CSM.
• To improve reliability in real-world deployments where steganographic images may originate from diverse cameras, processing pipelines, or image modifications.
• To reduce misclassification due to CSM by detecting classifier inconsistencies rather than relying on uniform training data.

Proposed method

• Uses Detection of Classifier Inconsistencies (DCI) to detect when a classifier's predictions are inconsistent across images from the same actor, indicating CSM.
• Employs EfficientNet neural networks for deep feature extraction from images to capture source-specific statistical properties.
• Applies a Gradient Boosting Machine (GBM) classifier that uses DCI predictions as input to classify actors as innocent, guilty, or unclassifiable due to CSM.
• Introduces a threshold-based mechanism to flag actors with high CSM, discarding them from classification to avoid erroneous labeling.
• Treats each actor as a group of images sharing the same source distribution, enabling detection of inconsistencies across their image set.
• Combines DCI prediction with steganalysis to simultaneously detect steganography use and source mismatch, improving robustness in real-world settings.

Experimental results

Research questions

• RQ1 Can classifier inconsistency detection (DCI) be effectively used to detect and handle cover source mismatch (CSM) in real-world actor-based steganalysis?
• RQ2 How does the proposed method maintain high accuracy under extreme CSM conditions where baseline models fail?
• RQ3 To what extent can the method classify actors as guilty, innocent, or unclassifiable due to CSM in realistic scenarios with unknown stegosystems and payloads?
• RQ4 How does the threshold for CSM detection affect the trade-off between accuracy and the number of unclassifiable actors?
• RQ5 Can the method be applied to both spatial and transformed domain steganography (e.g., JPEG) with consistent performance?

Key findings

• The proposed method maintains an accuracy of 84% in classifying guilty actors under 100% CSM in the spatial domain, while the baseline degrades to 50% (random guessing).
• In the transformed domain (JPEG), the method achieves 81% accuracy at 100% CSM, compared to 56% for the baseline.
• At 100% CSM, 84% of actors are discarded due to excessive CSM in the spatial domain, and 67% in the transformed domain, preventing misclassification.
• The method consistently achieves over 80% accuracy across all CSM levels, demonstrating robustness in high-mismatch scenarios.
• Lowering the CSM detection threshold from 0.7 to 0.65 reduces unclassifiable actors from 84% to 75% at 100% CSM but reduces accuracy to 62%, highlighting a trade-off between precision and coverage.
• The method effectively handles both CSM and stego source mismatch (SSM), as it is robust to unknown stegosystems and varying payloads.