Abstract
Clickbait PDFs represent a growing but understudied threat in the cybersecurity landscape. Unlike traditional malicious PDFs that embed exploits or malware, clickbait PDFs are technically "clean" documents that use visual deception to lure victims to malicious websites. This paper presents the first systematic study of this phenomenon, analyzing a dataset of over 176,000 unique documents. The research reveals that these threats are primarily distributed via Search Engine Optimization (SEO) attacks rather than email, allowing them to persist undetected by standard antivirus software for months.
The Emergence of Clickbait PDFs
For decades, the security community treated malicious PDFs (MalPDFs) as synonymous with embedded exploits or malware. Consequently, defense mechanisms focused heavily on detecting malicious code within the file structure. However, a new breed of "Clickbait PDFs" has emerged. These files are technically "clean"—they contain only images and links—but they use sophisticated social engineering to trick victims into visiting phishing sites or downloading malware.
The authors address a significant "blind spot" in current cybersecurity: because these PDFs lack malicious payloads, they receive low risk scores from automated analysis services like VirusTotal. Furthermore, while traditional email phishing campaigns are often short-lived (lasting about a day), Clickbait PDFs represent a more persistent, web-based threat that the scientific community has largely ignored until now.
Methodology: Visual Clustering and SEO Analysis
To identify and characterize these documents at scale, the researchers developed a multi-stage pipeline:
- Visual Clustering: Since Clickbait PDFs rely on visual lures, the authors used DeepCluster—a framework combining Convolutional Neural Networks (CNNs) with k-means clustering—to group 176,208 PDF screenshots based on their appearance. This allowed them to condense thousands of files into 80 manageable categories.
- URL Analysis: They extracted "bait URLs" (links reachable by clicking the first page) and verified their maliciousness using VirusTotal's API and manual inspection.
- Distribution Vector Analysis: To determine how these files reach victims, the authors cross-referenced file hashes with corporate "spam traps" and used search engine APIs (Google and Bing) to find evidence of SEO poisoning. They looked for "keyword stuffing" and "cross-linking" patterns within the PDFs designed to game search engine rankings.
The study utilized a dataset of 176,208 unique PDF documents collected between December 2020 and June 2021, provided by industrial partners Cisco and InQuest Labs.
Key Insights: UI Mimicry and SEO Distribution
The most surprising finding is that 89% of the clickbait PDFs in the dataset belonged to just three clusters distributed via SEO attacks, not email. Attackers "poison" search results so that these PDFs appear when users search for popular terms like "Windows XP iso" or "ROBLOX."
Attackers have moved beyond simple "Click Here" buttons to mimic familiar web elements. These "visual baits" exploit user habits; people are conditioned to click a reCAPTCHA to proceed or an "Update Payment" button to resolve an account issue, making the attack feel like a standard, legitimate web interaction. Unlike email phishing, which usually spikes and disappears in 24 hours, these SEO-driven PDF clusters are remarkably stable, with some remaining active for the entire six-month duration of the study.
Key Results and Defensive Gaps
Out of 80 identified visual clusters, 44 were confirmed to lead to malicious activity, including credential phishing, malware downloads, and "adult content" scams.
The study highlights a severe failure in current defenses:
- Dominance of the "Big Three": Three massive clusters (reCAPTCHA, ROBLOX Text, and ROBLOX Picture) account for the vast majority of the volume.
- Search Engine Success: By querying search engines with titles from their dataset, the researchers successfully retrieved 10,416 documents via Bing and Google, confirming that SEO poisoning is a highly effective distribution method.
- Defensive Failure: Manual testing showed that 40% of the malicious URLs embedded in these PDFs were not blocked by popular tools like Google SafeBrowsing or common ad-blockers.
- AV Blind Spot: Standard antivirus (AV) engines struggle with these files. Because there is no "exploit" to detect, the overlap in AV scores between benign PDFs and clickbait PDFs is significant.
Conclusion and Future Work
Clickbait PDFs are a persistent threat that evades traditional content-based filtering. The authors suggest that future defense systems should move toward domain-specific detection, combining visual similarity analysis with the detection of "cross-linked" PDF structures to flag SEO-poisoning attempts before they reach the user. To foster further study, the authors have publicly released the file hashes, screenshots, and URLs from their dataset.
Summary updated