You Wont Believe It: Exploring the Advertising Ecosystem of Fake - - PowerPoint PPT Presentation

You Won’t Believe It:

Catherine Han Allen Tong CS 261N

1

Exploring the Advertising Ecosystem

• f Fake News Websites

Misinformation Flavors: Deception

2

Misinformation Flavors: Junk Science

3

Misinformation Flavors: Unreliable Clickbait

4

Misinformation Related Work

• Detection
• Fact-checking
• Prevention
• Propagation modeling

5

Online Advertisement Problem Space

6

• Big revenue stream for

“free” services

○ For news publishers too

• Impact on user experience

and privacy

○ Annoyance ○ Performance ○ Trackers ○ Fear

Online Advertisement Problem Space, con’t

• Targeting algorithms

○ Algorithmic bias

• RTB ecosystem

7

Real-Time Bidding (RTB) Simplified

8

Ad Exchange

Publisher Server

AD SLOT

P I I + r e q u e s t u s e r p r

• f

i l e request for bids Advertiser A Advertiser B Advertiser C Advertiser D User Browser $A $B $C $D winning bid ($A), creative (A) embed A A

Misinformation and Ad Revenue

9

Misinformation and Ad Revenue, con’t

10

Misinformation and Ad Revenue, con’t

11

Research Problem

• Understanding the ecosystem of online

ads on fake news sites

○ Identifying third-party mediators ○ Identifying advertisers ○ Categorizing ads (e.g., medical, health, etc.)

• Comparative analysis with popular benign

news sites

12

Construction of Site Corpus

• Fake News

○ Zimdars’ False, Misleading, Clickbait-y and/or Satirical “News” Sources List

■ Domain, “About Us”, source, style, aesthetic, social media analysis

• “Real” (Benign) News

○ Alexa Top Sites (News category)

13

foo.co m

rendered HTML for foo.com

Methodology - Collecting Ads

14

foo.com foo_1 foo_2 foo_3 foo_1_1 foo_1_2 foo_1_3 foo_2_1 foo_2_2 foo_2_3 foo_3_1 foo_3_2 foo_3_3

EasyList Malware Domains UBlock Lists

Methodology - Collecting Ads

15

foo.co m

rendered HTML for foo.com + 12 foo.com links < a h r e f = ” . . . ” > <a href=”...”> <a href= ad URLs

Methodology - Categorizing Ads

16

ad URLs URL URL URL track URL redirects final ad landing pages LDA dictionary of ad words ad topics

Methodology - Challenges & Lessons Learned

• Choosing the corpus of fake news sites
• Web crawling woes

○ Headless browsing detection ○ Dynamically loaded content (ads) ○ Asynchronicity in JS

• Categorization of websites

○ Amazon Web Information Service API

17

Limitations & Ethics

• Inability to account for fingerprinting
• Corpus size and robustness
• Gathering (consistent) corpus metadata
• Clicking on ads

18

Results - Third-Party Ad Servers

19 … (truncated)

• Third-party ads were found on 246 fake news sites and 716 benign sites

… (truncated)

Results - Site Ranking and Third-Party Ad Servers

20

Results - Advertisers

21

Misinfo Benign

Results - Site Ranking and Advertisers

22

Results - Products Advertised

23

Advertisement Category Unique Fake News Sites (N=246) Unique Benign News Sites (N=716) Technology 235 (96%) 403 (56%) Medicine & Health 133 (57%) 383 (53%) Coronavirus

(proper subset of Medicine & Health)

36 (15%) 151 (21%) Finance 84 (34%) 329 (46%) Politics 24 (10%) 40 (6%) Cannabis and CBD 23 (10%) 46 (6%)

Future Work

• Larger corpus
• Categorization of ads
• Longitudinal crawling
• Motivations of users navigating

to such sites

○ Browser extension + user study with MTurk

24