TL;DR
Researchers conducted a large-scale, multi-platform study of donation-based scams, identifying 832 fraudsters across X, Instagram, Facebook, YouTube, and Telegram. By combining LLM-based context filtering with industry validation from PayPal and Chainabuse, the study reveals how scammers leverage aged accounts, cross-platform redirection, and emotional manipulation to siphon millions in fraudulent donations.
What Problem Does This Paper Address?
Social media has democratized fundraising, but it has also created a "Pirates of Charity" problem. Fraudsters exploit the emotional nature of charitable giving to solicit funds for fake causes—ranging from disaster relief to animal rescue. While individual scams are often reported, there hasn't been a comprehensive, end-to-end study of the scammers' behavior, cross-platform operations, and financial impact.
The authors address this gap by investigating how these "charity pirates" create profiles, engage with victims, and move money across five major social media ecosystems. They highlight a systemic failure: social media platforms are currently ill-equipped to detect and block these coordinated, multi-stage abuses.
Methods
The authors developed a three-stage pipeline to identify and analyze fraudulent activity:
- Data Collection: Using 78 donation-related keywords (e.g., "givebetter," "act of kindness," "Ukraine"), they scraped over 150,000 accounts and 3 million posts.
- Scam Filtration:
- Context Check: They used GPT-4o to filter out non-donation-related content (e.g., general news about cancer vs. a solicitation for a cancer patient).
- Risk Engine Validation: Profiles were flagged if their linked URLs, emails, or phone numbers appeared in security databases like VirusTotal or IPQualityScore.
- Tracking & Scam Analysis:
- Content Clustering: They employed BERTopic (UMAP + HDBSCAN) to categorize the types of scams and CLIP embeddings to cluster profile pictures.
- Sentiment Analysis: Llama-3-8B was used to classify YouTube comments to understand user-scammer interactions.
- Network Analysis: Automated scripts (Selenium) traced external links (like Linktree) to uncover interconnected scam operations spanning 11 different platforms.
Datasets
The study utilized a massive raw dataset collected between March and May 2024:
- Total Accounts: 151,966
- Total Posts: 3,053,333
- Platforms: X (Twitter), Instagram, Facebook, YouTube, and Telegram.
- Validation Data: 1,898 email addresses shared with PayPal and 142 cryptocurrency addresses shared with Chainabuse for ground-truth verification.
Key Insights and Ideas
- The "Harvested Account" Strategy: Scammers aren't just creating new accounts; they are using "aged" or harvested profiles. On X, the median account age for scammers was 2011. Older accounts lend a veneer of legitimacy that bypasses simple bot detection.
- Visual Trust-Building: 29% of scammers used "Association Logos" (NGOs, religious symbols, or flags like Ukraine's) as profile pictures to manufacture credibility.
- Cross-Platform Redirection: Scammers rarely stay on one platform. They use social media as a "top-of-funnel" lure, redirecting victims to external sites, messaging apps (Signal/WhatsApp), or crowdfunding platforms (Patreon, Donorbox) to finalize the fraud.
- The YouTube "Gratitude" Trap: Over 53% of analyzed YouTube comments expressed gratitude. The authors suggest these are often scripted or manipulated to create a false sense of community and success, encouraging others to donate.
- Platform Preferences: Scammers on X preferred malicious URLs, while those on Facebook leaned heavily toward email-based fraud. YouTube was the primary hub for video-based lures.
Key Results
- Identified Scammers: 832 unique fraudsters were confirmed across the five platforms.
- Interconnected Operations: Network analysis uncovered an additional 1,001 external accounts linked to these scammers, extending their reach to 11 platforms total.
- Financial Impact:
- PayPal: 79.7% of shared emails were linked to active PayPal accounts; 26% of these had already been restricted for "risky operations" or compliance issues.
- Crypto: The researchers identified 96 active crypto accounts with a staggering total incoming volume of $247 million. While this includes "long-tail" transactions that might not all be scams, the scale is immense.
- Crowdfunding: 37 active fundraising links had already defrauded 3,696 contributors of over $252,620.
Limitations and Future Work
The authors acknowledge several limitations:
- Conservative Filtering: Their multi-stage filtration process was designed to minimize false positives, meaning they likely missed many sophisticated scammers who don't use flagged URLs or emails.
- Multilingual Gap: The LLM-based context check struggled with non-English posts, leading to potential under-reporting of scams in other languages.
- Ephemeral Content: Phishing URLs are often taken down quickly, making real-time detection a "cat-and-mouse" game.
Future Work: The researchers suggest that platforms should implement "fraud scores" during sign-up and monitor referral headers to see where donation traffic is originating. They also advocate for better information sharing between social media platforms and financial services to block the "charity pirates" before they can cash out.
Summary updated