Abstract
SMS spam has evolved from simple commercial advertisements into sophisticated fraudulent schemes, yet the security community lacks large-scale, public datasets to study this "dark" communication channel. This paper introduces SpamHunter, an automated pipeline that leverages Twitter as a real-time sensor for SMS spam by extracting and analyzing screenshots shared by users. By processing over 21,000 unique messages across 75 languages, the study reveals a significant shift toward fraud and the use of advanced infrastructure to evade detection.
1. Introduction
SMS spam is a multi-billion dollar problem that has transitioned from nuisance advertising to dangerous "smishing" (SMS phishing) attacks. These attacks often exploit current events, such as COVID-19 or delivery notifications, to steal sensitive information. Despite its prevalence, SMS remains difficult for researchers to monitor because it is a private, point-to-point communication medium. Existing datasets are either outdated or proprietary, hindering the development of effective defenses.
2. The SpamHunter Pipeline
To overcome the data scarcity problem, the authors developed SpamHunter, a multi-stage pipeline designed to identify and extract spam content from Twitter reports.
2.1 Tweet Collection and Classification
The pipeline begins by using the Twitter Academic API to collect tweets containing keywords like "scam sms" or "phishing" that include image attachments. Because not all such tweets are actual spam reports, a Spam-Reporting Tweet Classifier (SRTC)—a 3-layer neural network—is used to distinguish between genuine complaints and irrelevant content.
2.2 Image Detection and Text Extraction
Once a relevant tweet is identified, the SMS Image Detector (SID) uses the YOLOv3 object detection algorithm to locate specific UI elements of an SMS interface, such as message bubbles and sender icons.
To ensure high-fidelity text extraction, the SMS Text Recognizer (STR) utilizes the Google Vision API. Unlike standard OCR, which often captures irrelevant metadata (e.g., battery levels or carrier names), the STR focuses specifically on the coordinates identified by the SID. By applying a custom "OverlapRatio" metric, the system achieves a 99% word accuracy rate, significantly outperforming general-purpose OCR.
3. Dataset and Insights
The resulting dataset comprises 21,918 unique SMS spam messages collected between 2018 and 2021. This represents the largest public SMS spam dataset currently available.
3.1 The Shift to Fraud
Analysis of the dataset shows that modern spam is overwhelmingly fraudulent. While historical datasets from 2012 were dominated by commercial ads, 62% of the messages in the SpamHunter dataset involve fraud, including fake bank alerts, tax refund scams, and account recovery lures.
3.2 Infrastructure and Evasion
Spammers are increasingly employing sophisticated technical infrastructure to bypass traditional IP-based blacklists:
- Port Forwarding Services (PFS): Tools like ngrok are used to hide backend servers.
- Dynamic DNS: Services like duckdns allow spammers to change domain mappings rapidly.
- Anycast IPs: These are used to distribute traffic and complicate geographic blocking efforts.
Furthermore, the study identified "multi-lingual" campaigns where identical malicious URLs were distributed in different languages across various regions, indicating highly organized, global operations.
4. Evaluation of Defenses
The research highlights a significant gap in current SMS security:
- Early Warning: Malicious URLs often appear in user-reported tweets at least one week before they are flagged by industry-standard platforms like VirusTotal.
- Filter Ineffectiveness: Bulk SMS providers (e.g., Twilio, TextMagic) were found to block only 6% to 17% of the spam samples identified in this study.
- False Positives: While some anti-spam services have high detection rates, they frequently misclassify legitimate messages, such as One-Time Passwords (OTPs), as spam.
5. Conclusion and Future Work
SpamHunter demonstrates that social media can serve as a powerful, real-time intelligence source for "dark" communication channels. The findings suggest that carriers and law enforcement should integrate social media monitoring into their threat intelligence workflows to update blacklists more rapidly. Future work will focus on reducing the 5% noise rate in the dataset and mitigating the risk of spammers "poisoning" the collection process with fake reports.
Summary updated