International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 06 | Jun 2025
p-ISSN: 2395-0072
www.irjet.net
An AI-Powered System for Scam Detection on Amazon’s Online Shopping Platform Yogesh J. Pawar, Adish S. Nair, Samrudhi L. Musale, Atharv R. Nalawade, Siddhi R. Naktode, Nachiket Girnar, Aditya A. Nadekar Department of Engineering, Sciences and Humanities (DESH) Vishwakarma Institute of Technology, Pune, Maharashtra, India ---------------------------------------------------------------------------***--------------------------------------------------------------------------legitimacy report, e.g., a fake-review percentage, along with Abstract — With the surge of e-commerce in recent years, online scams exploiting product reviews have become a pressing concern. Fake ratings, manipulated feedback, and deceptive sellers undermine trust in platforms like Amazon. This paper proposes an end-to-end system combining a robust Amazon review scraper with a BERT-based legitimacy classifier deployed via a Flask web application. The scraper, based on Oxylabs' implementation, collects up-to-date product reviews. The backend uses a fine-tuned BERT model trained on both real and fake review datasets to classify reviews as malicious or unworthy. The integrated platform provides users with a legitimacy score, fake-review estimates, and additional metrics to evaluate online products effectively.
Keywords — Online scam detection, Amazon reviews, fake
reviews, BERT, sentiment analysis, transformer models, product legitimacy, e-commerce security, web scraping, text classification, cybersecurity in e-commerce, consumer protection, deep learning, multilabel classification
I. INTRODUCTION THE extensive use of online reviews has exposed sites such as Amazon to fake manipulation, compromising consumer confidence and product visibility. This paper introduces an integrated system for determining the authenticity of products from user reviews. Fundamentally, the system utilizes an effective and powerful scraping engine to gather actual-time reviews from Amazon product pages. Toward this end, we integrate and improve the open-source Amazon Review Scraper developed by Oxylabs [1], which employs headless browser automation and stealth mechanisms to evade contemporary web defenses. The scraper is called by a Flask-based backend that accepts user input, initiates scraping, and processes the scraped output. The reviews are fed into a BERT-based classifier trained on data sets with both real and fake reviews [2][3]. The model returns each review with a probability score indicating whether it is malicious or not worthy. The backend then aggregates these predictions into a final
© 2025, IRJET
|
Impact Factor value: 8.315
|
confidence measures. All these elements—scraping, classification, and result determination—are orchestrated in a light-weight web application that enables real-time analysis. The modular design makes the system scalable, explainable, and able to address evolving e-commerce threat vectors.
II. LITERATURE REVIEW Several studies have explored the application of machine learning and web-based technologies to detect fraudulent activity in online shopping environments. Weng et al. [4] introduced CATS, a cross-platform machine learning system that detects counterfeit product ratings by analyzing review and rating patterns across multiple e-commerce platforms. Its ability to generalize trends makes it particularly effective at identifying widespread scam behaviors. Azzuri and Sulaiman [5] proposed a scam website detection system that classifies e-commerce sites as either legitimate or fraudulent using a Random Forest classifier, achieving a notable 93% accuracy. However, their model primarily focuses on structural website features rather than user-generated content. Cao et al. [6] addressed the problem of fake reviews by detecting collusive groups of reviewers. They utilized an unsupervised graph neural network model combined with modularity-based clustering to identify coordinated review fraud—a strategy that surpasses individual review analysis in sophistication. In another study, Ahmed et al. [7] benchmarked 11 classification algorithms on shopping behavior datasets and found that the Decision Table algorithm delivered the best performance (87.13% accuracy) using the WEKA platform, reinforcing the value of data mining for fraud detection. While they show promising outcomes, these methods tend to address isolated scopes—either web page-level metadata
ISO 9001:2008 Certified Journal
|
Page 995