The rapid proliferation of AI-generated and digitally altered images on social media has caused significant concerns about content authenticity, impersonation, misinformation and online safety. Even though current forensic techniques can detect deepfakes and analyze image provenance separately, they often fail to provide a unified and explainable framework for verifying the authenticity of visual content. In this paper, we introduce FakeLineage, a multi-signal forensic structure for trustworthy social media image authentication and analysis of their provenance. The proposed model integrates deep feature-based provenance graph construction, verification of metadata integrity, forensic analysis at the pixel level, GAN artifact detection and perceptual similarity matching to detect manipulated images. We use a calibrated ensemble scoring strategy plus a fine ResNet-50 feature embedding model to produce directed provenance graphs that are then optimized through maximum spanning tree optimization.

The system is tested on a specially prepared dataset that contains 1,200 multi-transformation image chains and is further validated on 15,000 benchmark images gathered from FaceForensics++, Celeb-DF v2, StyleGAN2/3, COCO, and RAISE datasets. The experiments register a provenance reconstruction precision of 91.4% and a recall of 87.6%, beating the traditional perceptual hash and feature-based baseline methods. Despite composite splicing and missing intermediate images being difficult scenarios, the use of complementary forensic signals still greatly enhances robustness and interpretability. The discussed setup stands as an efficient solution to confirming visual content authenticity and paves the way for various social media security applications such as misinformation analysis, harmful content moderation, and digital forensic investigations.

Deepfake Detection,ResNet-150,Metadata Integrity,GAN Artifacts,Provenance Graph,Multi-Signal Forensics