Towards Robust Media Integrity: Semi-Supervised Image Forgery Localization with DeepLabv3 and ELA

Image forgery localization is critical in ensuring the authenticity of digital content, yet the challenge of limited labeled data persists. This study presents a semi-supervised approach leveraging Error Level Analysis and DeepLabv3 to accurately detect tampered regions. ELA highlights compression-induced artifacts, serving as input features alongside RGB channels to enhance detection. The DeepLabv3 model, known for its advanced semantic segmentation capabilities, is employed to learn and predict forgery masks effectively.Our semi-supervised framework utilizes labeled and unlabeled data, with consistency regularization to improve predictions on unlabeled images. Experiments conducted on diverse datasets evaluate the model’s performance using accuracy, precision, recall, Area Under the Curve and Intersection over Union metrics. The results demonstrate robust performance, achieving high localization accuracy across various forgery types. This approach bridges the gap between traditional preprocessing techniques and deep learning, offering a scalable solution for image forgery localization in real-world scenarios.