SMART IMAGE FORGERY DETECTION USING COMBINED LIGHTWEIGHT AI MODELS

Abstract

Digital image forgery has emerged as a critical challenge in the era of widespread multimedia dissemination, where manipulated images can facilitate misinformation, fraud, and deceptive evidence generation. Conventional forgery detection techniques, including feature-based and deep learning approaches, often exhibit performance degradation under realworld conditions such as compression artifacts, blurring, lowtexture regions, and social media-induced distortions. To address these limitations, a robust and efficient framework is introduced for detecting copy-move forgeries, image splicing, and tampered regions. The methodology integrates multiscale feature extraction with self-correlation analysis to capture both local and global inconsistencies within images. A consistencybased learning mechanism is employed to enhance discrimination between authentic and manipulated regions, while balanced training strategies mitigate class imbalance issues. Furthermore, probability calibration is incorporated to improve confidence estimation and reduce false positives. The framework is evaluated on benchmark forensic datasets, including VISION, DEFACTO, CASIA v2, and CoMoFoD, demonstrating strong generalization capabilities. Experimental results showed strong performance, achieving AUROC scores of 0.9830 for copy-move detection and 0.9891 for splicing detection, with F1-scores of 0.8218 and 0.9460, respectively. Calibration also reduced the false positive rate from 0.0416 to 0.0284. Additionally, calibration significantly reduces false positive rates, enhancing reliability. The proposed approach offers an effective and practical solution for real-world digital image forgery detection and localization tasks.