Article’s

Abstract — The rapid expansion of e-commerce has precipitated a significant rise in delivery and return fraud, resulting in substantial financial losses and erosion of consumer trust. Conventional verification approaches—such as barcode scanning and manual inspection—are insufficient for detecting sophisticated fraudulent activities including product substitution, missing accessories, and fabricated return claims. This paper proposes a Visual Audit Framework (VAF) that leverages computer vision and deep learning to perform automated product verification across the three critical transaction stages: packing, delivery, and return. The framework integrates YOLOv9 for real-time object detection and a Siamese Neural Network for cross-stage visual similarity assessment. Transactions are classified as genuine, suspicious, or fraudulent using a multi-modal decision engine. Experimental results demonstrate a fraud detection accuracy of 96.4%, with a mean Intersection over Union (mIoU) of 0.91 for product localization and a similarity threshold of 0.82 for authenticity confirmation. The proposed system enhances supply-chain transparency, reduces manual inspection overhead by approximately 73%, and provides auditable digital evidence for dispute resolution in e-commerce ecosystems.