Authors: Ayush Wankhede, Ajinkya Patil, Partth Thombre, Mohit Patil, Mahesh Korade
Abstract: Online pet adoption platforms face significant chal- lenges with fraudulent listings and attribute misrepresentation, eroding user trust. This paper presents a complete pet adoption system integrating CNN-based image verification to authenti- cate listing attributes before publication. Transfer learning with EfficientNet-B0 is applied to 110,425 images spanning 712 breed classes across dogs, cats, and birds. A two-stage training strategy first trains the classification head with frozen base layers, achiev- ing 84.7% validation accuracy, then fine-tunes the top 40 layers to reach 89.3% validation accuracy. The verification pipeline combines breed confidence, color confidence, and prediction certainty into a normalized trust score (VScore, range 0–100). Server-side scoring with an 85-point threshold prevents client manipulation while achieving 98.0% fraud detection accuracy. A one-way privacy gateway protects adopter identities, and automated digital adoption certificates with unique certification IDs formalize successful adoptions. Experimental validation on 128 verification requests demonstrates an 84.4% acceptance rate, 1.8 s average processing time, and only 1.6% false positive rate.
DOI: http://doi.org/10.5281/zenodo.20930615
