TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

Facebook's learning defense system against attacks [pdf]

2 pointsby Revisoralmost 14 years ago

1 comment

Revisoralmost 14 years ago
As I run a forum and we assess all threats manually (a bunch of moderators watching over the forum and reacting to flagged posts), this is an interesting view of what the moderation looks like when it's automated.<p>What caught my eye:<p><i>Threats to the social graph can be tracked to three root causes. These are compromised accounts, fake accounts, and creepers.<p>Our earlier phishing classifiers made heavy use of features on IP and successive geodistance. Attackers have responded by using proxies and botnets to log in to their compro- mised inventory. Malware is a tough problem because the attacker is operating from the same machine as the legitimate user, so IP does not provide signal. To combat malware, the most effective mechanism we have discovered is to target the propagation vector using user feedback. Attackers can also try to game user feedback features. That is combatted with reporter reputation and rate limits.<p>Chain letter volume can explode when spread using the powerful viral channels of Facebook. In the past, they have been observed to reach 1-5% of total user communica- tions in minutes.<p>Chain letters exploit social engineering to trick otherwise well-behaved Facebook users into propagating the attack. As with other creeper attacks, the best long-term answer is education. In the short-term other mechanisms can be used against chain letters specifically. For example, fuzzy n-gram matching or other forms of locality-sensitive hashing on text.<p>Like users, attacks use many different channels. For the system to be ef- fective it must share feedback and feature data across channels and classifiers.</i>
评论 #2693833 未加载