PERCIVAL: Making In-Browser Perceptual Ad Blocking Practical with Deep Learning
summary
PERCIVAL is an in-browser deep learning ad blocker that hooks into Chromium/Brave at the image decode/raster chokepoint to classify every decoded image and suppress ads; prototype adds ~4.55% render overhead on Chromium and ~19.07% on Brave and ~120–140 ms on a 100-image page. Models start from SqueezeNet, retrained and later swapped to MobileNetV2 to defeat adversarial samples (retrained in ~9 minutes). Image-based blocking replicates EasyList at 96.76% and on Facebook achieves 78.4% precision / 70% recall (custom user model: 88.04% precision, 97.25% recall).
AI-written summary · about
chapters (8)
- 0:00Start
- 1:11Why Would Anyone Want To Block Ads
- 5:26Challenges of Running Deep Learning in the Browser
- 9:30Layout Phase
- 11:35In-Browser Inference
- 13:52In-Browser Evaluation of Percival on Facebook Ads
- 15:29Demo of Percival
- 17:15Adversarial Attacks
related talks
25mFiltering bad ads using LLM assisted adblocking
26mBeyond filter lists: using LLMs to rethink ad blocking
27mAgents need adblocking, too
15mBeyond the banner: The power of Web AI to personalize paid rich media ads
19mBringing Content Blocking to the Masses
01h 19mEffective Transparency and User Controls: Are We Close to a Breakthrough?
19mBlock, unblock, block! How ad blockers are being circumvented…
15mDouble-Edged Shield: On the Fingerprintability of Customized Ad Blockers