Autonomous Detection of Collaborative Link Spam Andrew G. West Wikimania `11 – August 5, 2011

Download Report

Transcript Autonomous Detection of Collaborative Link Spam Andrew G. West Wikimania `11 – August 5, 2011

Autonomous Detection
of Collaborative Link Spam
Andrew G. West
Wikimania `11 – August 5, 2011
Big Idea
Design a framework to detect link spam additions
to wikis/Wikipedia, including those employing:
(1) Subtlety; aims at link persistence (status quo)
(2) Vulnerabilities of recent literature [1]
And use this tool’s functionality to:
(1) Autonomously undo obvious spam (i.e., a bot)
(2) Feed non-obvious, but questionable instances
to human patrollers in a streamlined fashion
2
Outline
•
•
•
•
•
Motivation
Corpus construction
Features
Performance
Live implementation
• Demonstration
3
External Link Spam
• Any external link (i.e., not
wikilink) which violates the
subjective link policy [2] is
external link spam:
• Issues with the destination
URL (the landing site). For
example: Commercial
intent, shock sites, nonnotable sources.
• Issues with presentation.
For example: Putting a link
atop article.
[[http://www.google.com
Visit Google!]]
Article:
Visit Google!
4
Research Motivations
5
Motive: Social Spam
• Not entirely different from
link spam in other UGC/
collaborative applications
• Much literature [3]
• But wikis/WP are unique:
• Not append-only; less
formatting constraints
• Massive traffic in a single
installation.
• Large topic space: Spam
relevant to landing site.
• Mitigation entirely
community-driven
6
Motive: Incentive
• Much research on wiki vandalism (link spam is a subset)
• But most vandalism is “offensive” or “nonsense” [4]
• Not well incentivized; whereas link spam likely serves
poster self-interest (e.g., monetary, lobbying, etc.)
• Thus, more aggressive/optimal attack tactics expected
• In [1], examined the status quo nature of WP link spam:
• “Nuisance”: order of magnitude less frequent than
vandalism. See also [[WP:WPSPAM]].
• Less sophistication than seen in other spam domains
• Subtlety: Even spam links follow conventions. Perhaps an
attempt to deceive patrollers/watch-list: Doesn’t work
7
Motive: Vulnerabilities
• Novel attack model [1],
exploits human latency and
appears economically viable:
• High-traffic page targets
• Prominent placement
• Script-driven operation
via autonomously attained
privileged accounts
• Distributed
• Other recent concerns:
• Declining labor force [5]
• XRumer (blackhat SEO) [6]
Peak views/second
Calculated
Jan.-Aug.
2010
Highest avg. views/day
8
Corpus Construction
9
Corpus: Spam/Ham
• SPAM edits are those that:
1. Added exactly one link to an HTML document
2. The only changes made in the edit are the link addition
and its immediate supporting “context”
3. Were rolled-back by a privileged editor. Where rollback is
a tool used only for “blatantly unproductive edits”
• HAM edits are those:
1. Added by a privileged user
2. Meeting criteria (1) and (2) above
By relying on implicit actions, we save time and allow
privileged users to define spam on case-by-case basis
10
Corpus: Context (1)
Because the link was the ONLY change made.
The privileged user’s decision to roll-back that
edit speaks DIRECTLY to that link’s
inappropriateness.
11
Corpus: Context (2)
12
Corpus: Collection
238,371
188,210
2,865
human
All links
collected
Links to
HTML doc.
Was rolled
back
only link
LBL
SPAM
HAM
NUM#
1,095
4,867
PER%
18.4%
81.6%
“context”
human
Added by
priv. user
only link
50,108
“context”
1,095
SPAM
HAM
4,867
• ≈2 months of data collection in early 2011 (en.wiki)
• Done in real-time viva the STiki framework [7]
• Also retrieved the landing site for each link
• Be careful of biasing features!
13
Features
14
Features
55+ features implemented and described in [8]
For brevity, focus on features that: (1) Are strong
indicators, and (2) have intuitive presentation
Three major feature groups:
(1) Wikipedia-driven
(2) Landing-site driven
(3) Third-party based
15
Features: Wiki (1)
• Examples of Wikipedia features:
•
URL: Length, sub-domain quantity, TLD (.com, .org, etc.)
•
Article: Length, popularity, age
•
Presentation: Where in article, citation, description length
•
URL History: Addition quantity, diversity of adding users
•
Metadata: Revision comment length, time-of-day
16
Features: Wiki (2)
• Long URLs are good URLs:
• www.example.com vs. www.example.com/doc.html
• Former more likely to be promotional
• Spam is rarely used in citations
• Advanced syntax implies advanced editors
17
Features: Wiki (3)
≈ 85% of spam leaves
no “revision comment”
vs. < 20% of ham
TLDs with greater admin.
control tend to be good.
Also correlates well with
registration cost
18
Features: Wiki (4)
An article’s non-citation links “stabilize” with time
(Non-cites tend to have their own section at article bottom)
19
Features: Site
• We fetch and process the HTML source of the landing site
• Spam destinations marginally more profane/commercial (SEO?)
• Re-implement features from a study of email spam URLs [9]
• Opposite results from that work
• TAKEAWAY: Subtlety and link diversity impair this analysis
20
Features: 3rd Party (1)
Two third-party sources queried:
• Alexa WIS [10]: Data from webcrawling and toolbar. Including
traffic data, WHOIS, load times…
• Google Safe Browsing Project: Lists
suspected phishing /malware hosts
Google lists produce NO corpus matches:
• So worthless during learning/training
• But a match is unambiguously bad…
• …. So bring into force via statically authored rules
21
Features: 3rd Party (2)
• #1 weighted feature
• At median, ham has 850 BLs, spam
has 20 BLs (40× difference).
• Intuitive: basis for search-engine rank
• Continent of WHOIS
registration.
• Asia especially poor
• Other good, nonintuitive Alexa feats.
22
Performance
23
Perform: ADTrees
Example
ADTree:
• Features turned into ML
model via ADTree algorithm
• Human-readable
• Enumerated features
• In practice…
• Tree-depth of 15
• About 35 features
val =0.0
BACKLINKS > 200
Y: +0.8
IS_CITE == TRUE
N: -0.1
Y: +0.6
• Evaluation performed using
standard 90/10 cross
validation.
…
…
N: -0.4
COMM_LEN> 0
Y: +0.2
…
N: -0.6
…
if (final_value > 0): HAM
if (final_value < 0): SPAM
24
Perform: Piecewise
• Obviously, much better than random predictions (status quo)
• Wikipedia-driven features weighed most helpful
• But also must consider robustness issues
25
Perform: All
26
Live Implementation
27
Live: Architecture
Wikipedia
IRC
#enwiki#
Wiki-DB
Wikipedia
Article
Spam-scoring engine
Scoring/ADTree
Bot
Handler
SafeBrowse
3rd-party
<HTML>
…
</HTML>
Landing
site
STiki Client
STiki Client
if(score)
< thresh:
then:
Alexa
STiki
Services
Edit
Queue
STiki Client
Likely vandalism
Likely vandalism
Likely vandalism
-------------------
Fetch Edit
Likely innocent
else
REVERT
Queue
Mainte-nance
Display
Classify
If spam, revert and warn
• Bringing live is trivial via IRC and implemented ADTree
• But how to apply the scores:
• Autonomously (i.e., a bot); threshold; approval-pending
• Prioritized human-review via STiki [7]
• Priority-queue used in crowd-sourced fashion
28
Live: Demonstration
Software Demonstration
open-source [7], [[WP:STiki]]
29
Live: Discussion
• Practical implementation notes
• Multiple links: Score individually, assign worst
• Dead links (i.e., HTTP 404): Reported to special page
• Non-HTML destinations: Omit corresponding features
• Static rules needed to capture novel attack vectors
• Features are in place: Page placement, article popularity,
link histories, diversity quotients, Safe Browsing lists.
• Pattern matches result in arbitrarily high scores
• Offline vs. online performance
• Bot performance immediate; should be equal
• Difficult to quantify decline with prioritized humans (STiki)
30
Live: Gamesmanship
How would an attacker circumvent our system?
• Content-optimization (need for robust features)
• TOCTTOU attacks (i.e., redirect after inspection)
• Rechecking on an interval is very expensive
• But a practical concern; LTA case [[WP:UNID]]
• Crawler redirection
• Determine our system’s IP; serve better content
• A more distributed operational base
• Denial-of-service (overloading system with links) + STiki
Solutions to these kinds of problems remain future work
31
References
[1] A. G. West, J. Chang, K. Venkatasubramanian, O. Sokolsky, and I. Lee. Link
spamming Wikipedia for profit. In CEAS ‘11, September 2011.
[2] Wikipedia: External links. http://en.wikipedia.org/wiki/Wikipedia:External_links.
[3] P. Heymann, G. Koutrika, and H. Garcia-Molina. Fighting spam on social web sites:
A survey of approaches and future challenges. IEEE Internet Comp., 11(6), 2007.
[4] R. Priedhorsky, J. Chen, S. K. Lam, K. Panciera, L. Terveen, and J. Riedl.
Creating, destroying, and restoring value in Wikipedia. In GROUP ’07, 2007.
[5] E. Goldman. Wikipedia’s labor squeeze and its consequences. Journal of
Telecommunications and High Technology Law, 8, 2009.
[6] Y. Shin, M. Gupta, and S. Myers. The nuts and bolts of a forum spam automator.
In LEET’11: 4th Wkshp. on Large-Scale Exploits and Emergent Threats, 2011.
[7] A.G. West. STiki: A vandalism detection tool for Wikipedia.
http://en.wikipedia.org/wiki/Wikipedia:STiki. Software, 2010.
[8] A.G. West, A. Agarwal, P. Baker, B. Exline, and I. Lee. Autonomous Link Spam
Detection in Purely Collaborative Environments. In WikiSym '11.
[9] A. Ntoulas, M. Najork, M. Manasse, and D. Fetterly. Detecting spam web pages
through content analysis. In WWW’06: World Wide Web Conference, 2006.
[10] Alexa web information service. http://aws.amazon.com/awis/.
32
Backup slides (1)
33
Backup slides (2)
L
I
N
K
S
Prevention
Detection
Blacklists
Patrollers Watchlisters Readers
Latency:
Immediate
Seconds
Mins./Days
∞
LEFT: Pipeline of
typical Wikipedia link
spam detection,
including both actors
and latency
RIGHT: log-log plot showing
average daily article views
versus article popularity rank.
Observe the power-law
distribution. A spammer could
reach large percentages of
viewership via few articles.
34
Backup slides (3)
ABOVE: Evaluating time-of-day (TOD) and
day-of-week (DOW)
AROUND: Evaluating feature strength. All
features here are “Wikipedia-driven”.
35