Multi-Domain Information-Theoretic Features and Kolmogorov Complexity for Lightweight Image Splicing Detection

Multi-Domain Information-Theoretic Features and Kolmogorov Complexity for Lightweight Image Splicing Detection

Stať ve sborníku mimo WoS a Scopus

Image splicing is a prevalent form of digital forgery that challenges the reliability of visual content across forensic, legal, and media platforms. In this study, a novel and lightweight image splicing detection framework has been proposed grounded in multi-domain information-theoretic principles. Unlike deep learning-based methods that require large datasets and high-end GPU resources, the proposed approach leverages handcrafted entropy, mutual information, and complexity-based features that are computationally efficient and interpretable. The proposed framework extracts 20 features from spatial, cross-channel, and multi-scale domains—highlighting entropy variations, statistical dependencies, and information complexity. Notably, Kolmogorov complexity approximation, edge entropy, and multi-scale mutual information are incorporated as discriminative indicators of tampering. The model is evaluated on the Columbia Image Splicing Detection Dataset using ten classical machine learning algorithms. A maximum AUC-ROC of 0.934, cross-validation performance of 84.29%(±4.19%) and test accuracy of 85.32% with a strong F1-score of 0.8571 were achieved with classical machine learning classifiers, demonstrating competitive performance without deep models. Feature importance analysis further improves interpretability by ranking the most significant contributors. The results validate proposed framework as a reliable, resource-efficient, and explainable alternative to complex end-to-end deep learning pipelines for splicing detection.

Image splicing is a prevalent form of digital forgery that challenges the reliability of visual content across forensic, legal, and media platforms. In this study, a novel and lightweight image splicing detection framework has been proposed grounded in multi-domain information-theoretic principles. Unlike deep learning-based methods that require large datasets and high-end GPU resources, the proposed approach leverages handcrafted entropy, mutual information, and complexity-based features that are computationally efficient and interpretable. The proposed framework extracts 20 features from spatial, cross-channel, and multi-scale domains—highlighting entropy variations, statistical dependencies, and information complexity. Notably, Kolmogorov complexity approximation, edge entropy, and multi-scale mutual information are incorporated as discriminative indicators of tampering. The model is evaluated on the Columbia Image Splicing Detection Dataset using ten classical machine learning algorithms. A maximum AUC-ROC of 0.934, cross-validation performance of 84.29%(±4.19%) and test accuracy of 85.32% with a strong F1-score of 0.8571 were achieved with classical machine learning classifiers, demonstrating competitive performance without deep models. Feature importance analysis further improves interpretability by ranking the most significant contributors. The results validate proposed framework as a reliable, resource-efficient, and explainable alternative to complex end-to-end deep learning pipelines for splicing detection.

Image Splicing Detection, Information Theory, Entropy, Mutual Information, Kolmogorov Complexity, Lightweight Model, Forgery Detection

Image Splicing Detection, Information Theory, Entropy, Mutual Information, Kolmogorov Complexity, Lightweight Model, Forgery Detection

TYAGI, N.; JOSHI, R.; DAS, S.; KUNAL, .; SCHILLER, V.; JEŽEK, Š.; DUTTA, M.

2026

03.11.2025

IEEE

2025 17th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT)

227

233

7