[Paper Review] I call BS: Fraud Detection in Crowdfunding Campaigns

Beatrice Perez, Sara Machado+2 more|arXiv (Cornell University)|Jun 30, 2020

FinTech, Crowdfunding, Digital Finance4 citations

TL;DR

This paper proposes a machine learning framework that detects fraudulent crowdfunding campaigns at publication time using only textual and visual features from campaign descriptions and images. By training an ensemble classifier on a manually annotated dataset of 700 health-related campaigns, the method achieves 96.01% AUC and 91.14% precision, enabling real-time fraud detection in browsers or platform moderation systems.

ABSTRACT

Donations to charity-based crowdfunding environments have been on the rise in the last few years. Unsurprisingly, deception and fraud in such platforms have also increased, but have not been thoroughly studied to understand what characteristics can expose such behavior and allow its automatic detection and blocking. Indeed, crowdfunding platforms are the only ones typically performing oversight for the campaigns launched in each service. However, they are not properly incentivized to combat fraud among users and the campaigns they launch: on the one hand, a platform's revenue is directly proportional to the number of transactions performed (since the platform charges a fixed amount per donation); on the other hand, if a platform is transparent with respect to how much fraud it has, it may discourage potential donors from participating. In this paper, we take the first step in studying fraud in crowdfunding campaigns. We analyze data collected from different crowdfunding platforms, and annotate 700 campaigns as fraud or not. We compute various textual and image-based features and study their distributions and how they associate with campaign fraud. Using these attributes, we build machine learning classifiers, and show that it is possible to automatically classify such fraudulent behavior with up to 90.14% accuracy and 96.01% AUC, only using features available from the campaign's description at the moment of publication (i.e., with no user or money activity), making our method applicable for real-time operation on a user browser.

Motivation & Objective

• To investigate the characteristics of fraudulent donation-based crowdfunding campaigns in health-related appeals.
• To build a dataset of 700 annotated campaigns (fraud vs. non-fraud) from major platforms like GoFundMe and Indiegogo.
• To develop a machine learning classifier that detects fraud using only features available at campaign publication, without requiring user activity or transaction data.
• To evaluate whether combining textual and visual features improves fraud detection beyond human-level performance.
• To enable real-time, proactive fraud detection via deployable tools such as browser plugins or platform-side screening systems.

Proposed method

• Collected and annotated 700 crowdfunding campaigns from GoFundMe, MightyCause, Fundly, Fundrazr, and Indiegogo, focusing on medical and emergency appeals.
• Extracted textual features using NLP techniques, including sentiment, readability, and linguistic cues associated with deception.
• Extracted image-based features using computer vision and deep learning models to detect anomalies, such as manipulated or generic stock images.
• Combined textual and visual features into a unified feature vector for each campaign.
• Trained an ensemble machine learning classifier (e.g., XGBoost or similar) on the combined feature set to predict fraud status at time of publication.
• Evaluated model performance using standard metrics: AUC, precision, recall, and accuracy, with a focus on real-time applicability.

Experimental results

Research questions

• RQ1 What textual and visual features distinguish fraudulent crowdfunding campaigns from legitimate ones at the time of publication?
• RQ2 Can a machine learning model trained on pre-publication features detect fraud with higher accuracy than human judgment?
• RQ3 How effective is the combination of NLP and image processing in detecting deception in donation-based crowdfunding campaigns?
• RQ4 To what extent can such a system be deployed in real time, e.g., as a browser plugin or platform-level filter?
• RQ5 What are the main sources of bias in the annotation and model training process, and how can they be mitigated?

Key findings

• The proposed model achieves a maximum AUC of 96.01% in detecting fraudulent campaigns using only pre-publication textual and image features.
• The classifier reaches 91.14% precision and 90.77% recall, significantly outperforming human deception detection capabilities (approximately 41% improvement).
• Textual features such as emotional language, readability, and linguistic deception cues are strong predictors of fraud.
• Image-based features, including use of generic or stock photos, significantly correlate with fraudulent campaigns.
• The model’s performance demonstrates that fraud detection is feasible at the moment of campaign creation, enabling proactive intervention.
• The study reveals that fraud constitutes a small but impactful fraction of the crowdfunding ecosystem, with potentially millions in annual losses.