PDF Mirage: Content Masking Attack Against Information-Based Online Services
summary
Work demonstrates a PDF content-masking attack that alters underlying text via custom masking font mappings so humans see one document while automated systems index another. Reproducing the INFOCOM reviewer-assignment pipeline with 114 TPC members, 2,000 training papers and 100 tests, two thirds of papers require only six masked keywords (always <15) to target a reviewer. A random word replacement variant tunes Turnitin similarity to 5–15%. Bing, Yahoo and DuckDuckGo index masked content; Google blocks it (likely OCR). A lightweight font verification defense renders embedded glyphs and OCRs them, achieving near-100% detection with far lower cost than full-document OCR.
AI-written summary · about
chapters (16)
- 0:00Start
- 0:11PDF Mirage: Content Masking Attack Against Information-Based Online Services
- 0:34Motivation
- 2:06Background Information
- 4:57Content Masking Attack Against Automatic Conference Reviewer Assignment Systems
- 9:56Content Masking Attack Against Plagiarism Detection
- 11:58Content Masking Attack Against Document Indexing
- 15:18Content Masking Defense - Our proposal
- 15:56Challenges and Technical Details
- 17:09Font Training
- 17:46Font Verification Performance
- 18:04Performance - Experiment 1
- 18:57Performance - Experiment 2
- 19:40Performance - Experiment 3
- 20:01Conclusion
- 20:23Thank you!
related talks
40mFinding Secrets in PDF Files
20mUltraviolet: Turn Hidden Document Data into an AI Advantage
56mHomoglyph-Based Attacks: Circumventing LLM Detectors
15mVulnerability of Text-Matching in ML/AI Conference Reviewer Assignments...
33mPDF - Mess With The Web
26mDefining a Data Masking Framework at Scale
45mWinning by deception: tactics, techniques, and procedures of adversarial A.I.
44mComputers are Stupid: Protecting "AI" from Itself