AdFlush

Original Article ↗ Hacker News Discussion ↗

A new research project called AdFlush applies machine learning and hand-engineered features to detect online ads, claiming higher accuracy than prior academic adblockers and even a slight edge over uBlock Origin. Commenters question its practical value given heavy performance costs and the fact that community-maintained blocklists already block most ads efficiently, while also noting issues of list governance and potential abuse. The conversation broadens into worries about Google’s Manifest V3 restrictions on extension capabilities, the future arms race between adblockers and ad formats, and the appeal of hybrid or browser-level solutions that could remain effective despite tightening platform controls.

AdFlush approach and results

  • AdFlush uses a classical, feature-engineered ML model to detect ad-related resources (e.g., JS AST structure, identifier length, access patterns, graph metrics of scripts).
  • Reported F1 score is 0.98 on 10,000 real sites, outperforming prior academic systems (AdGraph, WebGraph, WTAgraph).
  • It is advertised as more CPU- and memory-efficient than those systems and more robust to adversarial manipulation.

Comparison with uBlock Origin and list-based blockers

  • Several commenters ask why comparison is not primarily against popular list-based blockers.
  • The paper does contain a uBlock Origin comparison: AdFlush F1 ≈ 0.86 vs uBO ≈ 0.84, a marginal advantage that is not claimed to be statistically significant.
  • Many view list-based blocking as a “solved” practical solution; algorithmic approaches are seen as more about research value or long-term robustness.

Performance and real-world viability

  • Significant performance penalty: median page load times reported as ~2.7s (no blocker), ~2.1–2.2s (uBO), 6.6s (AdFlush fresh), 3.4s (AdFlush with cached predictions).
  • Commenters doubt it can compete with URL-rule matching for real-time use but see potential as an offline tool to generate/augment filter lists.

Crowdsourced lists: strengths, weaknesses, and abuse

  • Lists scale well: one report can protect millions.
  • Concerns raised about abuse, “corruption,” and pay-to-whitelist/blacklist behaviors, as well as accidental overblocking and the difficulty of getting removed.
  • Others argue abuses and mistakes are observable and often quickly reversible.

Chrome Manifest V3 and ecosystem concerns

  • Discussion centers on MV3’s limits: only declarative rules, no dynamic inspection, bans on remote code, and capped rule counts (with mentions of recent increases).
  • Many see MV3 as aimed at weakening powerful adblockers, potentially breaking advanced or algorithmic blockers entirely.
  • Some predict a browser exodus (to Firefox, Brave, etc.) if adblocking becomes ineffective; others think most users won’t switch.

Hybrid and future directions

  • Multiple commenters propose hybrid systems: lists for fast blocking; ML for offline analysis or catching evasive ads.
  • Ideas extend to network-layer/MITM filtering, DNS-level tools (Pi-hole, AdGuard Home) plus browser blockers, and even AI-based page rendering that strips ads before display.
  • Native ads and server-side-integrated ads are flagged as the hardest to handle; some see them as unavoidable without changing what content users choose to consume.