Identifying Disinformation Websites Using Infrastructure Features
summary
Claim: infrastructure metadata can flag disinformation sites early. Trained a random forest on 551 disinformation, 553 news, 555 other snapshots using domain features, certificate features, and hosting features and gets strong cross‑validation ROC/PR performance; domain features alone work well, certificate features perform poorly. Built a real‑time pipeline ingesting Twitter via the streaming API across two servers for five days; live results collapse — only about 5 of 100 disinformation predictions are correct — highlighting class imbalance and dataset drift.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:31Motivation
- 1:21Key Insight
- 3:09Certificate Features
- 3:30Hosting Features
- 4:59Cross-Validation on Training Data
- 5:47Disinformation Websites Are Detectable
- 6:30Domain Features Do Well on Their Own
- 7:03Preliminary Evaluation on Real-Time Data
- 7:40Performance Degraded on Real-Time Dat
- 8:17Theories for Poor Real-Time Performance
- 9:29Conclusion
related talks
20mVulnerable to misinformation?: Verifi!
32mTricking the Trolls: A Look at the Digital Future
40mTechnology to counter misinformation/disinformation
35mDisinformation or Debate? Finding Signal Amid Political Noise
35mDisinform your Surroundings: AI and disinformation campaigns
21mMachine Learning and Cyber Security - Detecting malicious URLs in the haystack
24mThe Disinformation Industry
36mDisinform your Surroundings: AI and disinformation campaigns