Presentation - SNOW Workshop

Report
Challenges of
Computational Verification in Social Media
Christina Boididou1, Symeon Papadopoulos1, Yiannis Kompatsiaris1,
Steve Schifferes2, Nic Newman2
1Centre
2City
for Research and Technology Hellas (CERTH) – Information Technologies Institute (ITI)
University London – Journalism Department
WWW’14, April 8, Seoul, Korea
How trustworthy is Web multimedia?
Real photo
captured April 2011 by WSJ
but
heavily tweeted during Hurricane Sandy
(29 Oct 2012)
Tweeted by multiple sources &
retweeted multiple times
Original online at:
http://blogs.wsj.com/metropolis/2011/04/28/weatherjournal-clouds-gathered-but-no-tornado-damage/
#2
Disseminating (real?) content on Twitter
• Twitter is the platform for sharing newsworthy
content in real-time.
• Pressure for airing stories very quickly leaves very
little room for verification.
• Very often, even well-reputed news providers fall for
fake news content.
• Here, we examine the feasibility and challenges of
conducting verification of shared media content with
the help of a machine learning framework.
#3
Related: Web & OSN Spam
• Web spam is a relatively old problem, wherein the spammer
tries to “trick” search engines into thinking that a webpage is
high-quality, while it’s not (Gyongyi & Garcia-Molina, 2005).
• Spam revived in the age of social media. For instance,
spammers try to promote irrelevant links using popular
hashtags (Benevenuto et al., 2010; Stringhini et al., 2010).
Mainly focused on characterizing/detecting sources of spam
(websites, twitter accounts) rather than spam content.
Z. Gyongyi and H. Garcia-Molina. Web spam taxonomy. In First international workshop on
adversarial information retrieval on the web (AIRWeb), 2005
F. Benevenuto, G. Magno, T. Rodrigues, and V. Almeida. Detecting spammers on twitter. In
Collaboration, Electronic messaging, Anti-abuse and Spam conference (CEAS), volume 6, 2010
G. Stringhini, C. Kruegel, and G. Vigna. Detecting spammers on social networks. In Proceedings of
the 26th Annual Computer Security Applications Conference, pages 1–9. ACM, 2010.
#4
Related: Diffusion of Spam
• In many cases, the propagation patterns between
real and fake content are different, e.g. in the case of
the large Chile earthquakes (Mendoza et al., 2010)
• Using a few nodes of the network as “monitors”, one
could try to identify sources of fake rumours (Seo
and Mohapatra, 2012).
Still, such methods are very hard to use in real-time
settings or very soon after an event starts.
M. Mendoza, B. Poblete, and C. Castillo. Twitter under crisis: Can we trust what we rt? In
Proceedings of the first Workshop on Social Media Analytics, pages 71–79. ACM, 2010
E. Seo, P. Mohapatra, and T. Abdelzaher. Identifying rumors and their sources in social networks.
In SPIE Defense, Security, and Sensing, 2012
#5
Related: Assessing Content Credibility
• Four types of features are considered: message,
user, topic and propagation (Castillo et al., 2011).
• Classify tweets with images as fake or not using a
machine learning approach (Gupta et al., 2013) 
Reports an accuracy of ~97%, which is a gross overestimation of expected real-world accuracy.
C. Castillo, M. Mendoza, and B. Poblete. Information credibility on twitter. In Proceedings of the
20th international conference on World Wide Web, pages 675–684. ACM, 2011.
A. Gupta, H. Lamba, P. Kumaraguru, and A. Joshi. Faking sandy: characterizing and identifying
fake images on twitter during hurricane sandy. In Proceedings of the 22nd international
conference on World Wide Web companion, pages 729–736, 2013
#6
Goals/Contributions
• Distinguish between fake and real content shared on
Twitter using a supervised approach
• Provide closer to reality estimates of automatic
verification performance
• Explore methodological issues with respect to
evaluating classifier performance
• Create reusable resources
– Fake (and real) tweets (incl. images) corpus
– Open-source implementation
#7
Methodology
• Corpus Creation
– Topsy API
– Near-duplicate image detection
• Feature Extraction
– Content-based features
– User-based features
• Classifier Building & Evaluation
– Cross-validation
– Independent photo sets
– Cross-dataset training
#8
Corpus Creation
• Define a set of keywords K around an event of interest.
• Use Topsy API (keyword-based search) and keep only
tweets containing images T.
• Using independent online sources, define a set of fake
images IF and a set of real ones IR.
• Select TC ⊂ T of tweets that contain any of the images in
IF or IR.
• Use near-duplicate visual search (VLAD+SURF) to extend
TC with tweets that contain near-duplicate images.
• Manually check that the returned near-duplicates indeed
correspond to the images of IF or IR.
#9
Features
#
Content Feature
# User Feature
1
Length of the tweet
1
Username
2
Number of words
2
Number of friends
3
Number of exclamation marks
3
Number of followers
4
Number of quotation marks
4
Number of followers/number of friends
5
Contains emoticon (happy/sad)
5
Number of times the user was listed
6
Number of uppercase characters
6
If the user’s status contains URL
7
Number of hashtags
7
If the user is verified or not
8
Number of mentions
9
Number of pronouns
10
Number of URLs
11
Number of sentiment words
12
Number of retweets
#10
Training and Testing the Classifier
• Care should be taken to make sure that no
knowledge from the training set enters the
test set.
• This is NOT the case when using standard
cross-validation.
#11
The Problem with Cross-Validation
Training/Test tweets are randomly selected.
One of the reference fake images
Multiple tweets per reference image.
Training set
#12
Testing set
Independence of Training-Test Set
Training/Test tweets are constraint to correspond to
different reference images.
IR1
TR11
TR12
IR2
TR13
TR21
TR22
IR5
TR23
TR51
Training set
#13
TR52
TR53
Testing set
Cross-dataset Training-Testing
• In the most unfavourable case, the dataset used for
training should refer to a different event than the
one used for testing.
• Simulates real-world scenario of a breaking story,
where no prior information is available to news
professionals.
• Variants:
– Different event, same domain
– Different event, different domain (very challenging!)
#14
Evaluation
• Datasets
– Hurricane Sandy
– Boston Marathon bombings
• Evaluation of two sets of features
(content/user)
• Evaluation of different classifier settings
#15
Dataset – Hurricane Sandy
Natural disaster held around the USA from October 22nd to 31st, 2012. Fake
images and content, such as sharks inside New York and flooded Statue of
Liberty, went viral.
Hashtags
Hurricane Sandy
#hurricaneSandy
Hurricane
#hurricane
Sandy
#Sandy
#16
Dataset – Boston Marathon Bombings
The bombings occurred on 15 April, 2013 during the Boston Marathon
when two pressure cooker bombs exploded at 2:49 pm EDT, killing three
people and injuring an estimated 264 others.
Hashtags
Boston Marathon
#bostonMarathon
Boston bombings
#bostonbombings
Boston suspect
#bostonSuspect
manhunt
#manhunt
watertown
#watertown
Tsarnaev
#Tsarnaev
4chan
#4chan
Sunil Tripathi
#prayForBoston
#17
Dataset Statistics
Hurricane Sandy
Boston Marathon
Tweets with other image URLs
343939
Tweets with other image URLs
112449
Tweets with fake images
10758
Tweets with fake images
281
Tweets with real images
3540
Tweets with real images
460
#18
Prediction accuracy (1)
• 10-fold cross validation results using different classifiers
~80%
#19
Prediction accuracy (2)
• Results using different training and testing set from the
Hurricane Sandy dataset
~75%
• Results using Hurricane Sandy for training and Boston
Marathon for testing
~58%
#20
Sample Results
• Real tweet
My friend's sister's Trampolene in Long Island.
#HurricaneSandy
Classified as real
• Real tweet
23rd street repost from @wendybarton
#hurricanesandy #nyc
Classified as fake
• Fake tweet
Sharks in people's front yard #hurricane #sandy #bringing
#sharks #newyork #crazy http://t.co/PVewUIE1
Classified as fake
• Fake tweet
Statue of Liberty + crushing waves. http://t.co/7F93HuHV
#hurricaneparty #sandy
Classified as real
#21
Conclusion
• Challenges
– Data Collection: (a) Fake content is often removed (either
by user or by OSN admin), (b) API limitations make very
difficult the collection after an event takes place
– Classifier accuracy: Purely content-based classification can
only be of limited use, especially when used in a context of
a different event. However, one could imagine that
separate classifiers might be built for certain types of
incidents, cf. AIDR use for the recent Chile Earthquake
• Future Work
– Extend features: (a) geographic location of user (wrt.
location of incident), (b) time the tweet was posted
– Extend dataset: More events, more fake examples
#22
Thank you!
• Resources:
Slides: http://www.slideshare.net/sympapadopoulos/computationalverification-challenges-in-social-media
Code: https://github.com/socialsensor/computational-verification
Dataset: https://github.com/MKLab-ITI/image-verification-corpus
Help us make it bigger!
• Get in touch:
@sympapadopoulos / [email protected]
@CMpoi / [email protected]
#23
Sample fake and real images in Sandy
• Fake pictures shared on social media
• Real pictures shared on social media
#24

similar documents