Download ppt - Northwestern Networks Group

Understanding the Network and User-Targeting
Properties of Web Advertising Networks
Yong Wang1,2
Daniel Burgener1
Aleksandar Kuzmanovic1
Gabriel Maciá-Fernández3
1UESTC
(China)
2Northwestern University (USA)
3University of Granada (Spain)
Motivation
“Online advertising is a $20 billion industry that is
growing rapidly… It has become an integral and
inseparable part of the World Wide Web”
However, neither public auditing nor monitoring
mechanisms still exist in this emerging area
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Contributions
We present our initial efforts on building a network and
content-level auditing service for Web-based ad
networks.
Such an ad auditing service can effectively monitor
and regulate ad industry.
– Firstly, it helps potential new advertisers/publishers in the
decision of choosing commissioners which better meet their
requirements.
– Secondly, it allows commissioners to evaluate their own
networks, with the aim of detecting potential design flaws and
points of failure with reduced quality of service.
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Background
Commissioner
Publisher
Advertiser
…
6) Fetches ad
(and send
cookies)
Commissioner’s
Ad servers
7) Sends ad (and
cookies)
End user
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Outline
Charting Online Advertising Network Infrastructure
Network-Level Performance
Content-Level Performance
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Outline
Charting Online Advertising Network Infrastructure
–
–
–
–
–
Evaluation Platform
Candidates Selection
Finding Canonical Names
Mapping CNames to IP Addresses
Mapping IP Addresses to Locations
Network-Level Performance
Content-Level Performance
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Evaluation Platform
Open Recursive DNS
PlanetLab
Region
Countries
Servers
% of
total
N. America
25
33645
43.73
3
139
49.29
Europe
50
26294
34.18
22
103
36.53
Asia
40
14019
18.22
7
29
10.28
S. America
12
1405
1.83
3
8
2.84
Oceania
8
1111
1.44
1
3
1.06
Africa
24
456
0.60
0
0
0.0
Total
159
76930
100.00
36
282
100.00
Yong Wang
Countries Servers
Understanding the Network and User-Targeting Properties of Web Advertising Networks
% of
total
Candidates Selection
√
√
√
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Finding Canonical Names
Commissioners
Sub-companies
Representative domains
Google
Google-Google
pagead.1.google.com
Google-Doubleclick
pagead.1.doubleclick.net
AOL-Adtech
3
AOL- Adsonar
a950.g.akamai.net
AOL- Tacoda
3
AOL- Advertising
5
AOL/
Akamai
Adblade
web.adblade.com
AOL-Adtech : a627.g.akamai.net, a973.g.akamai.net, e1611.c.akamaiedge.net
AOL-Tacoda : a1131.g.akamai.net, a1406.g.akamai.net, e922.p.akamaiedge.net
AOL- Advertising:a949.g.akamai.net, a957.g.akamai, a1539.g.akamai.net,
a1626.g.akamai.net, e1066.c.akamaiedge.net
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Mapping CNames to IP Addresses
Commissioners
# of IP
Ad content servers
Ad DNS servers
Google
306
6
AOL/Akamai
11132
8361
Adblade
1
2
The difference of the discovery capacity between two
platforms 286 ÷306 = 93.5%
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Mapping IP Addresses to Locations
# of IP
Google
Region
AOL/Akamai
Adblade
Ad
DNS
Ad
DNS
Ad
DNS
N. America
154
3
6761
5426
1
2
Europe
70
2
3017
1824
0
0
Asia
24
1
994
883
0
0
S. America
14
0
144
91
0
0
Oceania
0
0
178
124
0
0
Africa
0
0
38
33
0
0
Unknowns
24
0
0
0
0
0
Total
286
0
11132
8381
1
2
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Outline
Charting Online Advertising Network Infrastructure
Network-Level Performance
– Delay Performance
– Ad vs. Publisher Networks
Content-Level Performance
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Delay Performance
Delay for ad content servers
Delay for ad DNS servers
AOL/Akamai > Google > Adblade
Ad content servers > Ad DNS servers
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Commissioner Ad vs. Publisher Networks
Ad network is worse
Ad network is better
In CDN case, Google-Google ≈ Publisher network
In No-CDN case, Google-Google > Publisher network
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Commissioner Ad vs. Publisher Networks
Ad network is worse
Ad network is better
In CDNexists
case,no
AOL-Adsonar
≈ Publisher
network
There
internal mechanism
within
a CDN to
recognize
correct
such anomalies.
In
No-CDNand
case,
AOL-Adsonar
> Publisher network
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Commissioner Ad vs. Publisher Networks
Ad networks is worse
Ad network is worse
Ad networks is better
Ad network is better
In CDN case, Adblade < Publisher network
In No-CDN case, Adblade > Publisher network
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Commissioner Ad vs. Publisher Networks
The discrepancy between publishers’ and commissioners’ ad
networks can be quite high.
There exists no internal mechanism within a CDN to recognize
and correct the huge discrepancy between commissioner and
publisher network, even if both are served by the same CDN.
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Outline
Charting Online Advertising Network Infrastructure
Network-Level Performance
Content-Level Performance
– Distribution Mechanisms
– Location-Based Advertising
– Behavioral Targeting
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Distribution Mechanisms (Similarity)
Google-Google has a large pool of ads and distributes
different ads into different servers.
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Distribution Mechanisms (Similarity)
Adblade has a smaller pool of ads and puts all of them in
the same machine (or a cluster of machines).
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Distribution Mechanisms (Similarity)
AOL uses different pools of ads depending on the location
of the servers
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Distribution Mechanisms
Regional similarities in AOL-Adsonar
AOL uses finer-grained location-based advertising, e.g., citylevel advertising, in the U.S.
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Location-Based Advertising
Percentage of vantage points observing location-based ads
Commissioner
City (%)
State (%)
No info (%)
Google
31.58
21.93
46.49
AOL-Adsonar
8.00
12.00
80.00
Adblade
37.31
0.00
62.69
Three commissioners deploy location-based advertising at
various levels of granularity
Google > Adblade > AOL
CDN-based commissioners lag behind others in achieving
finer-grained location-based advertising.
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Behavioral Targeting
Percentage increase of observed ’sport’ related ads when behavioral
targeting is enabled (’local/uniform cookie’) compared with disabled
(’no cookie’)
Commissioner
Local cookie (%)
Uniform cookie (%)
Google
25
3
AOL-Adsonar
13
5
Adblade
0
0
Uniform
Establish
Local
cookie
cookie
baseline
browsing
(enable
(enable
(disable
pattern
cookies,
cookies,
cookies,
(enable
and
and
access
cookies,
and
copyaccess
all
thewebsites,
and
cookies
allonly from
visit
one
websites,
websites
which
computer
may
fitwhich
not
in to
the
bemay
allcategory
related
PLnot
nodes,
be
to“sports”)
related
sports
and then
intoorder
sports)
retrieving
to determine
ads
again
the difference
to checkwhen
whether
behavioral
profile data
targeting
is stored
is used)
locally or
globally)
Data-center
commissioners
are capable
of collecting
Both
Googleoriented
and AOL-Adsonar
associate
a user profile
only
user an
profiles
and applying
behavioral
with
ad server
close to this
user targeting more effectively
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Conclusions
We deployed an ad auditing service that can be
universally applied to arbitrary commissioners’
networks.
Using this service, we performed an extensive
network- and content-level analysis.
Our findings bring useful auditing information to
all entities involved in the online advertising
business.
Yong Wang
Understanding the Network and User-Targeting Properties of Web Advertising Networks
Thank You