A Feature Fusion CNN Approach for Fake Educational Credential Detection and Forgery Region Localization

M, Sarala; L, Muralidhara B.; R, Suresh; Siddalingappa, Rashmi

A Feature Fusion CNN Approach for Fake Educational Credential Detection and Forgery Region Localization

M, Sarala, L, Muralidhara B., R, Suresh and Siddalingappa, Rashmi ORCID: https://orcid.org/0000-0001-9786-8436 (2026) A Feature Fusion CNN Approach for Fake Educational Credential Detection and Forgery Region Localization. International Journal of Intelligent Engineering and Systems, 19 (8). pp. 790-805.

Preview

Text
2026083141-2.pdf - Published Version
Available under License Creative Commons Attribution Share Alike.
| Preview

Official URL: https://doi.org/10.22266/ijies2026.0831.41

Abstract

Educational institutions are implementing tamper-proof techniques to prevent fraudulent degrees. In today's digital world, the creation of fake educational credentials has become prevalent, facilitated by document editing tools, which are often used to secure admissions and job placements. This study focused on identifying fake educational credentials across five categories: Handwritten, Computerized, Choice-Based Credit System (CBCS), Degree Certificate, and Non-CBCS using a machine learning (ML) approach. The entropy features from the preprocessed educational credentials dataset were extracted using Shannon entropy. We have chosen one hundred samples from the training data of each category to choose the best distance value using the Knee point method. The test dataset includes seven hundred test samples, which contain one hundred held-out normal samples, and forty fake samples as Out of Sample (OOS) in each category. It was clustered into either 'genuine' or 'fake' using Density-Based Spatial Clustering of Applications with Noise (DBSCAN) method, providing an accuracy of 99.14. Further, a customized feature fusion Convolutional Neural Network (CNN) model was designed with an attention mechanism and an Atrous Spatial Pyramid Pooling (ASPP) layer to locate fake regions in fake credentials. This proposed model shows significantly better results in locating fake regions of both image and text-based regions.