Filtering bad ads using LLM assisted adblocking
summary
Claims that current allowlists surface more problematic creative content, measuring a 13.6% average increase in problematic ads across a six-category taxonomy. Pipeline collects ~1,500 ads, places ~60 images per category under expert annotation, and runs a vision–language model (GP24 mini) with keyword prompts reaching ~70% agreement; synthetic ad generation (RealisticVision, Stability, OpenJourney) plus fine-tuning bumps accuracy to ~82% toward a 90% goal. Finds reseller routing (e.g., Criteo) changes when ads pass allowlists.
AI-written summary · about
chapters (11)
- 0:00Introduction to talk
- 1:48Problems with ad blocking
- 3:11Last year's research overview
- 5:19Defining problematic ads
- 7:27Research methodology
- 9:35Key findings and results
- 12:45Improving model accuracy
- 14:53Fine-tuning smaller models
- 17:43Expanding detection to scams
- 19:53Summary and takeaways
- 20:36Q&A session
related talks
21mPERCIVAL: Making In-Browser Perceptual Ad Blocking Practical with Deep Learning
12mWhy am I seeing this: Democratizing End User Auditing for Online Content Recommendations
19mAnother talk about ads, but this time what if we talked...
41mThe future of UA – Facebook ad best practices, creative marketplaces and UA bots
24mAI-Assisted Media Planning
15mDouble-Edged Shield: On the Fingerprintability of Customized Ad Blockers
01h 19mEffective Transparency and User Controls: Are We Close to a Breakthrough?
11mSession: Future of UA: Creatives, Predictions & Funding