Controlled Evaluation for Numeric-Field Forgery Detection in SROIE Receipts using Fine-Tuned EfficientNetV2-S

Tousi Adak; Priyanshu Mukherjee; Sudipta Das; Jahir Uddin Paik

doi:10.36948/ijfmr.2026.v08i05.89791

Controlled Evaluation for Numeric-Field Forgery Detection in SROIE Receipts using Fine-Tuned EfficientNetV2-S

Author(s)	Ms. Tousi Adak, Mr. Priyanshu Mukherjee, Mr. Sudipta Das, Mr. Jahir Uddin Paik
Country	India
Abstract	Altered amounts, dates, and identifiers on retail receipts represent a practical source of expense and tax fraud. However, public receipt datasets such as Scanned Receipts OCR and Information Extraction (SROIE) contain authentic receipts but do not provide forged counterparts, limiting controlled evaluation of receipt-forgery detection methods. Chakraborty (2025) reported 97.1% accuracy using a convolutional neural network (CNN)-based receipt-forgery detection framework. Building on this research context, the present study investigates forgery detection at the numeric-field level, focusing on receipt components that directly carry financial information. Using SROIE, a receipt-grouped benchmark of paired authentic and synthetically forged numeric-field crops is constructed. Two controlled forgery mechanisms are considered: splicing a donor numeric field from the same receipt and erasing and re-rendering a field with a specified proportion of its digits changed. Authentic and forged crops undergo the same Joint Photographic Experts Group (JPEG) compression procedure, and receipt-level grouping is used for training, validation, and test partitioning at approximately 70/15/15. An ImageNet-pretrained EfficientNetV2-S is fine-tuned end-to-end using linear warm-up, cosine learning-rate decay, and exponential moving average (EMA) weights, with model selection based exclusively on validation accuracy. On the held-out test set of 4,602 crops, comprising 2,301 authentic and 2,301 forged samples, the reference detector achieves 98.09% accuracy, 99.42% precision, 96.74% recall, an F1-score of 0.9806, and an area under the receiver operating characteristic (ROC) curve (AUC) of 0.9971 at a fixed decision threshold of 0.5 in a single run with seed 42. Most remaining errors correspond to missed forgeries. The reported accuracy is numerically higher than the 97.1% reported by Chakraborty (2025), but direct comparison is limited by differences in datasets, sample construction, and evaluation protocols. Gradient-weighted Class Activation Mapping (Grad-CAM) analysis indicates attention to numeric text regions but also, in some cases, to crop boundaries and background regions, suggesting that synthetic generation artifacts may contribute to the observed performance. As the study relies on synthetic forgeries without multiple baseline architectures, ablation experiments, multiple random seeds, or real-world forged-receipt evaluation, the findings are limited to the defined synthetic tampering processes rather than general real-world forgery-detection performance.
Keywords	Receipt Forgery Detection, Document Tampering, Numeric-Field Forgery, EfficientNetV2-S, SROIE, Grad-CAM, Synthetic Data, Receipt Image Analysis
Field	Engineering
Published In	Volume 8, Issue 5, September-October 2026
Published On	2026-10-16
DOI	https://doi.org/10.36948/ijfmr.2026.v08i05.89791

View / Download PDF File

About IJFMR Fees & Payment Current Issue Publication Archive	Submit Research Paper Track Submission Status Publication Guidelines Publication Ethics Peer Review & Plagiarism	Join as a Reviewer Editors & Reviewers Reviewer Referral Program Get Reviewer Membership Certi.	Website/Journal Policies Usage Policy Content Policies Privacy Policy

Contact Us		+91-9687-828-838	editor@ijfmr.com

International Journal For Multidisciplinary Research

A Widely Indexed Open Access Peer Reviewed Multidisciplinary Bi-monthly Scholarly International Journal

Controlled Evaluation for Numeric-Field Forgery Detection in SROIE Receipts using Fine-Tuned EfficientNetV2-S

Share this