In today’s digital era, ensuring image authenticity is crucial for maintaining trust and integrity. This study presents a CNN-based approach for image forgery localization, integrating Error Level Analysis (ELA). Using the CASIA v2.0 dataset, our model achieved an F1-score of 0.5712 and an AUC score of 0.7678, showing its effectiveness. Performance analysis across different tampering percentages shows improved accuracy for higher tampering levels, reaching an F1-score of 0.8099 for 20–40% tampering, 0.8331 for 40–60%, 0.8331 for 60–80%, and 0.8217 for 80–100% tampering. Particularly, for images with more than 20% tampering, our model consistently achieves an F1-score above 0.80, making it highly reliable for significant manipulations. Unlike complex architectures such as Mantra-Net (F1-score: 0.566, AUC: 0.817) and ObjectFormer (F1-score: 0.579, AUC: 0.758) that require extensive computational resources, our approach is lightweight and computationally efficient, making it suitable for real-world applications where high-performance hardware may not be available. These results validate the robustness of our method in accurately localizing tampered regions with high reliability and efficiency.