(Mis)Information Spread in Social Media Systems

advertisement
Misinformation in Social
Media
Anupam Joshi
Oros Family Professor and Chair, CSEE
Director, UMBC Center for Cybersecurity
University of Maryland Baltimore County
joshi@umbc.edu
Power of Social Media
1.44 Billion
monthly active
users
60 million
photos shared
everyday
* 2015 Statistics
500 million
tweets posted
every day
300 hours of
video uploaded
every minute
2
Motivation
• Social media sites are rich source of information for current
events, and supplement traditional sensors
•
Events/Accidents are reported on social networks even before they
appear on news channels
• Eg: Tweets on the 29th July 2009 earthquake in Southern California appeared a
few seconds later while official news emerged 4 minutes later
•
Social Networks often provide information where more formal challens
are censored
• Eg: Arab Spring, Iran Election 2009,
• Iranians used Twitter, Flickr, Youtube and some blogs to protest and
communicate with the outside world.
• #IranElection, #Ahmadinejad, #Mousavi, and #Tehran became trendy on twitter
• Youtube set up various channels to upload videos which have been shot via cellphone
and videocam.
• Iran Protests”, “Iran Riots 2009” or “Tehran Protests“ were popular tags on Flickr
after the election.
• On Facebook, IRAN page had around 40,000 fan following.
Areas of Interest
• Building a Scalable Infrastructure to harvest social media data
• Analysis of social media text and relationship (graph) data to (Social)
Situational Awareness:
• Detect Events and their Attributes
• Temporal Evolution
• Detect Communities
• Detect Sentiment
• Detect and Prevent Misinforation
Analytics
• Can use both the Social Network and the Content
• Can discover a variety of things
•
•
•
•
Individual tastes and preferences
Groups
Influential Individuals
Identity across social media
• Does Privacy still exist ?
A Pessimistic View on Privacy
• Mankind is a social animal, we have a “need to share.”
• Internet enabled social media scales that up
• McLuhan’s view of changing media effecting social organization
• Once data is “gone” can it ever be controlled?
• Especially if the economic incentives are aligned against it
• Is “Privacy” the norm in human affairs ?
“I did not inhale” in the Internet Age
• Hat tip David Chadwick and Ravi Sandhu
• Curating data about self is much harder when “youthful indiscretions”
are on social media
• Curating data about self might not be sufficient – empowerment of
“whisper campaigns”
• Nikki Haley in South Carolina
• Privacy vs Integrity
Social Media meets Mobile Phones
• Embedded sensors in mobile phones add significant data to social
media
• Often not controlled by the users
• Are “18th century laws” sufficient to argue about these questions ?
• U.S. v. Jones, 10-1259.
• What does “expectation of privacy” mean ?
Who owns the data anyway
• Does the creator of the data own it
• Do you turn it over to the “service” and they own it ?
• I can haz your facebook password, or Can I make your access to some
service conditional upon you sharing this data
• What about data that is created as a side effect of your explicit
actions
• Location data collected by telcos or “environment” creators
Legislation, Schmeligslation
• Legislation is often thought of as a solution
• Do we want “laws to catch up to the internet”
• The debate over CISPA
• Whose legislation is it when these services cross jurisdictional
boundaries
• Explicit censorship
• Implicit “good behavior” requests
Misinformation on Social Media
11
Misinformation Tweets
$
FAKE
RUMORS
12
Background: Hurricane Sandy
• Dates: Oct 22- 31, 2012
• Damages worth $75 billion
• Coast of NE America
Faking Sandy: Characterizing and Identifying Fake Images on Twitter during Hurricane Sandy. Aditi Gupta, Hemank Lamba, Ponnurangam Kumaraguru
and Anupam Joshi. Accepted at the 2nd International Workshop on Privacy and Security in Online Social Media (PSOSM), in conjunction with the 22th
International World Wide Web Conference (WWW), Rio De Janeiro, Brazil, 2013. Best Paper Award.
14
Fake Image Tweets
15
Data Description
Total tweets
1,782,526
Total unique users
1,174,266
Tweets with URLs
622,860
Tweets with fake images
10,350
Users with fake images
10,215
Tweets with real images
5,767
Users with real images
5,678
16
Network Analysis
Node -> User Id
Edge -> Retweet
Tweet – Retweet graph for the propagation of fake images during first 2 hours
17
Role of Twitter Network
• Analyzed role of follower network in fake image
propagation
• Crawled the Twitter network for all users who tweeted
the fake image URLs
 Graph 1
- Nodes: Users, Edges: Reweets
 Graph 2
- Nodes: Users, Edges: Follow relationships
18
Network Overlap Algorithm
19
Results
Total edges in retweet network
Total edges in follower-followee network
10,508
10,799,122
Common edges
1,215
%age Overlap
11%
20
Classification
5 fold cross validation
Tweet Features [F2]
Length of Tweet
Number of Words
Contains Question Mark?
Contains Exclamation Mark?
User Features [F1]
Number of Friends
Number of Followers
Number of Question Marks
Number of Exclamation Marks
Contains Happy Emoticon
Contains Sad Emoticon
Contains First Order Pronoun
Follower-Friend Ratio
Contains Second Order Pronoun
Contains Third Order Pronoun
Number of times listed
Number of uppercase characters
User has a URL
Number of negative sentiment words
User is a verified user
Number of positive sentiment words
Number of mentions
Number of hashtags
Number of URLs
Retweet count
Age of user account
21
Classification Results
F1 (user) F2 (tweet)
F1+F2
Naïve Bayes
56.32%
91.97%
91.52%
Decision Tree
53.24%
97.65%
96.65%
• Best results were obtained from Decision Tree classifier, we got 97%
accuracy in predicting fake images from real.
• Tweet based features are very effective in distinguishing fake images tweets
from real, while the performance of user based features was very poor.
22
Building and Evaluating a Real-time System
• Learning to Rank model for assessing credibility of Tweets
• Model based on ground truth data for 25 real world events
and 45 features
• System evaluation using year long real world experiment
• 1800+ users requested for credibility score of more than
14.2 million tweets.
TweetCred: Real-Time Credibility Assessment of Content on Twitter. Aditi Gupta, Ponnurangam Kumaraguru, Carlos Castillo and Patrick Meier. Proceedings of the 6th
International Conference on Social Informatics (SocInfo), Barcelona, Spain, 2014. Honorable Mention for Best Paper.
24
TweetCred Score
Score
User Feedback
25
Features for Real-time Analysis
Feature set
Features (45)
Tweet meta-data
Number of seconds since the tweet; Source of tweet (mobile /
web/ etc); Tweet contains geo-coordinates
Tweet content (simple)
Number of characters; Number of words; Number of URLs;
Number of hashtags; Number of unique characters; Presence of
stock symbol; Presence of happy smiley; Presence of sad smiley;
Tweet contains `via'; Presence of colon symbol
Presence of swear words; Presence of negative emotion words;
Presence of positive emotion words; Presence of pronouns;
Tweet content (linguistic) Mention of self words in tweet (I; my; mine)
Number of followers; friends; time since the user if on Twitter;
Tweet author
etc.
Tweet network
Tweet links
Number of retweets; Number of mentions; Tweet is a reply;
Tweet is a retweet
WOT score for the URL; Ratio of likes / dislikes for a YouTube
video
26
Annotation
• 500 Tweets per event (six different events)
• Used CrowdFlower
• Step 1
• R1. Contains information about the event
• R2. Is related to the event, but contains no information
• R3. Not related to the event
• R4. Skip tweet
*45% (class R1), 40% (class R2), and 15% (class R3)
• Step 2
• C1. Definitely credible
• C2. Seems credible
• C3. Definitely incredible
• C4. Skip tweet.
*52% (class C1), 35% (class C2), and 13% (class C3)
27
Ranking Model Evaluation
NDCG@25
NDCG@50
NDCG@75
NDCG@100
Time (training)
Time (testing)
AdaRank
0.6773
0.6861
0.6949
0.6669
Coord.
Ascent
0.5358
0.5194
0.7521
0.7607
SVMrank
0.3951
0.4919
0.6188
0.7219
RankBoost
0.6736
0.6825
0.689
0.6826
35-40 secs
<1 sec
1 min
<1 sec
35-40 secs 9-10 secs
<1 sec
<1 sec
28
Identity Problem
@theUSpresident
@BarakObama
Which one is real??
@BarackObama
Why?
• Security Applications
• Detect malicious user accounts!
• Detect compromised user accounts!
• Automatic Social Aggregation
• Smartly aggregating information, managing
privacy risks via other measures.
• Characterizing User behavior
across OSN
• Users activities across OSN?
• Targeted Phishing / Spam attacks
Our Approach
GROUND TRUTH ??
Our Approach(Contd..)
Content
Words in
tweets
Meta Data
Gender
Hash tags
Location
Followers
Links
Following
Replied
Re- tweets
Mentions
Age
Initial Results
Tweets about ‘Romney’ and ‘Massachusetts’
frequently with Tf-Idf scores of 0.16 and 0.13
Fake profile tweets about ‘dudes’ and ‘excuses ‘
with Tf-Idf scores of 0.168 and 0.15
4 out of 6 articles mentioning the
US president talk about him
mentioning ‘Romney’ and
‘Massachusetts’ with an average
TF-Idf of 0.093 and o.o85
Initial Results
Talks about ‘CSIR’ most frequently
Tf-Idf score of 0.1
A fake profile talks about a ‘polar’
‘satellite’ ‘launch’ with TF-IDF s of
0.206
TOI articles (2/2) mention ‘CSIR’ with
respect to the PMO with an average
Tf-Idf score of 0.09
Current Work
Research Tasks
• Task 1: Identify factors to compute the magnitude / severity of a
misinformation using predictive models
• Task 2: Building mathematical graph based models for the
misinformation cascades
• Task 3: Designing and formulating strategies to mitigate
misinformation propagation on social networks
• Task 4: Evaluating and Prototyping
Proposed Methodology
Motivation: Saffir-Simpson scale
• Scale to measure Hurricane category
• Based on wind speed
• Adapting to online social media
• Based on speed of propagation
• Identify other factors
Image: https://en.wikipedia.org/wiki/Saffir%E2%80%93Simpson_hurricane_wind_scale
Compute the magnitude of a misinformation
• Identify factors that effect misinformation propagation
• Based on users who are propagating
• Topic of information
• Location of an event
• Develop predictive models
Graph-based Models for Misinformation
Diffusion
• Various models exsist for information diffusion
• SIR, SIS and SEIS
• Threshold and independent cascades models
• Literature shows features and properties of rumor and true content
are distinct
• Need to build new mathematical models for misinformation
propagation
Literature Review
• Friggeri et al. tracked propagation of numerous rumor cascades on Facebook,
their results showed that rumor cascades run deeper than the normal re-share
cascades on Facebook.
• Mendoza et al. compared rumor and true news tweets and found that tweets
related to rumors contained more questions than news tweets containing true
news.
• From our preliminary work for events such as Hurricane Sandy, we concluded
that temporal, network and user based properties of rumor tweets are distinct
from true news tweets.
Download