AI presentation and introduction - Retrieval Augmented Generation RAG 101
Socialbots www2012
1. When socialbots attack:
Modeling susceptibility of users in online social networks
Claudia Wagner, Silvia Mitter, Christian Körner, Markus Strohmaier
Lyon, 16.4.2012
2. What are socialbots?
A socialbot is a piece of software that controls a user
account in an online social network and passes itself of as
a human being
3. 3
Danger of socialbots
Social Engineering
Gaining access to secure objects by exploiting human
psychology rather than using hacking techniques
Harvest private user data such as email addresses, phone
numbers, and other personal data that have monetary
value
Spread Misinformation
Ratkiewicz et al. describe the use of Twitter bots to run
smear campaigns during the 2010 U.S. midterm elections.
J. Ratkiewicz, M. Conover, M. Meiss, B. Goncalves, S. Patil, A. Flammini, and F. Menczer. Truthy:
mapping the spread of astroturf in microblog streams. In Proceedings of the 20th international
conference companion on World wide web, WWW '11, pages
4. Danger of socialbots
Snowball effects
Boshmaf et al. show that
Facebook can be infiltrated by
social bots sending friend
requests. 102 socialbots, 6
weeks, 3.517 friend requests and
2.079 infections
Average reported acceptance
rate: 59,1% up to 80% depending
on how many mutual friends the
social bots had with the infiltrated
users
Y. Boshmaf, I. Muslukhov, K. Beznosov, and M. Ripeanu. The socialbot network. In Proceedings
of the 27th Annual Computer Security Applications Conference, page 93. ACM Press, Dec 2011.
5. How likely will she
be infected by a bot
Experimental Setup
?
Whom shall we protect to avoid large scale infiltration due to
snowball effects?
Who is a bot? Whom shall we eliminate?
Is she a bot?
src: http://adobeairstream.com/green/a-natural-predicament-sustainability-in-the-21st-century/
6. Experimental Setup
Two-stage approach
Predict Infections (binary classification task)
Who is susceptible for bot attacks – i.e. who gets
infected?
Predict Infection level (regression task)
How susceptible is a user – i.e. how often does a user
interact with bots?
Dataset: Social Bot Challenge 2011
7. Social Bot Challenge 2011
Competition organized by Tim Hwang
Aim was to develop socialbots that persuade 500 randomly Twitter
users (targets) to interact with them
Targets have a topic in common: cats
Teams got points if targets replied to, mentioned, retweeted or
followed their lead bot
14 days during which teams were allowed to develop their social
bots.
Game started on the Jan 23rd 2011 (day 1) and ended Feb 5th 2011
(day 14)
At the 30th of January (day 8) the teams were allowed to update
their codebase
9. Feature Engineering
How likely will this user become infected?
User Network
Behavior
Content
10. Network Features
3 directed networks: Follow, retweet and interaction
(retweet, reply, mention and follow) network
Hub and Authority Score (HITS)
High authority score node has many incoming edges from
nodes with a high hub score
High hub score node has many outgoing edges to nodes
with a high authority score
In-degree and Out-degree
Clustering Coefficient
number of actual links between the neighbors of a node
divided by the number of possible links between them
11. Behavioral Features
Informational Coverage
Conversational Coverage
Question Coverage
Social Diversity
Informational Diversity
Temporal Diversity
Lexical Diversity
Topical Diversity
C. Wagner and M. Strohmaier. The wisdom in tweetonomies: Acquiring latent conceptual structures
From social awareness streams. In Proc. of the Semantic Search 2010 Workshop, April 2010.
12. Linguistic Features
LIWC uses a word count strategy searching for over
2300 words
Words have previously been categorized into over 70
linguistic dimensions.
standard language categories
(e.g., articles, prepositions, pronouns including first person
singular, first person plural, etc.)
psychological processes (e.g., positive and negative emotion
categories, cognitive processes such as use of causation
words, self-discrepancies),
relativity-related words (e.g., time, verb tense, motion, space)
traditional content dimensions
(e.g., sex, death, home, occupation).
J. Pennebaker, M. Mehl, and K. Niederhoer. Psychological aspects of natural language use: Our words,
our selves. Annual review of psychology, 54(1):547-577, 2003.
13. Feature Computation
For all targets we computed the features by using all
tweets they authored during the challenge (up to the
point in time where they become infected) and a
snapshot of the follow network which was as
recorded at the 26th of January (day 4)
We only used targets which became susceptible at
day 7 or later
Features do not contain any future information (such
as tweets or social relations which were created
after a user became infected)
14. Predict Infections
Binary Classification of users into susceptible and non-
susceptible
Train 6 classifiers
97 Features
Compare classifiers via 10 cross-fold validation
Balanced dataset
17. Predict Level of Infection
Which factors are correlated with users‘
susceptibility score?
Susceptibility score
counts number of interactions between a target and
any lead bot
Method: Regression Trees
can handle strongly nonlinear relationships with high order
interactions and different variable types
Fit the model to our 75% of the susceptible users
18. Users who
• use more negation words (e.g. not, never, no),
• tweet more regularly 1
(i.e. have a high temporal balance)
Predicting Levels of Susceptibility
• use more words related with the topic death
negemo
(e.g. bury, con, kill) < 0.40068 >= 0.40068
tend to interact more often with bots
2
temp_bal
< 0.37025 >= 0.37025
3
death
< −0.16389 >= −0.16389
Node 4 (n = 25) Node 5 (n = 7) Node 6 (n = 9) Node 7 (n = 15)
8 8 8 8
6 6 6 6
4 4 4 4
2 2 2 2
19. Predicting Levels of Susceptibility
Rank correlation of hold-out users given their real
susceptibility level and their predicted susceptibility level
(Kendall τ up to 0.45)
Goodness of fit (R2 up to 0.3)
Potential Reasons:
Dataset is too small (we only had 81 susceptible users
and 61% of them had level 1, 17% had level 2, 10% had
level 3, very few users had more than 3 interactions)
20. Summary & Conclusions
Approach to identify susceptible users
Features of all three types contributed to the
identification
Users are more likely to be susceptible if
they are emotional Meformers
they use Twitter mainly for communicating
their communications are not focused to a small circle of
friends
they are social and active (i.e., interact with many others)
21. Summary & Conclusions
Active Twitter users are more susceptible
They are more likely to see the messages/requests of
social bots
But we expected that they develop some skills to
distinguish social bots from human by using Twitter
frequently
Predicting users’ susceptibility score is difficult
More data and further experiments are required
22. Future Work
Repeating experiments on larger datasets
Taxonomy of social bot strategies
Massive numbers of con-messages (brute force)
Manipulation of messages through false retweets (changing pro-
to con messages)
Diverting attention by adding con-hashtags to pro-hashtags
Susceptibility of users for different strategies
23. Emotional Meformers which are active, communicative and social
Experimental Setup
are more likely to be infected
THANK YOU
claudia.wagner@joanneum.at
http://claudiawagner.info
src: http://adobeairstream.com/green/a-natural-predicament-sustainability-in-the-21st-century/
Editor's Notes
What makes a socialbot different from self-declared bots is that hide the fact that they're robots and usually try to pursue a variety of latent goals, such as to spread information or influence users. Tim Hang defined a socialbot as a machine with social impact.
And finally, recent research has shown that socialbots are extremely dangerous due to snowball effects. The more users a bot has infected in a network, the easier he can infect new users in that network. Boshmaf et al conducted in a very controversial experiment where they setup a network of 102 fb-bots which sent friend requests to others within a time period of 6 weeks. Their results show how a network of bots can infect fb user. Interestingly the average acceptance rate of friend requests was 59:1%, which, depends on howmany mutual friends the socialbots had with the inflltrated users, and can increase up to 80%.
So whatcanwe do toprevent large scaleinfilitrations due tosocial bot attacks? The traditional thingistotrytoidentifybotsandeliminatethem. In ourworkwesuggest a complementaryappraochwhichaimstoidentifyuserswhoaremostsuscepibleforsocial bot attacks. Wewantedtoknowiftheseusersshowspecialcharacteristicsand
Toanswerthisquestionweuse a 2-stage approach. First weaimtoidentifyuserswhoaresusceptibleto bot attacks in general– i.e., userswhobecameaffected–Wewereinterested in iftheseuserswhoanyspecificcharacteristicsoriftheseusersaraverageuserslikeyouandme.
In ourexperimentweuseddatafromthesocial bot challenge 2011 –whichis a competionthatwasorganizedby...
The dataset which we got contained all tweets which were published by the targets and bots during the challange and snapshots of the follow network between these users at different points in time. The figure shows how many users became susceptible at which day. One can see that most targets became susceptible at day 1. One possible explanation is the auto-follow feature which some of the targets might have used.
Sincewewereinterested in thefactorsthatimpactwhether a usergetsinfectedor not, wefirsthadto design featuresthatdescribe potential factors. In ourworkweused 3 different typesoffeature: featuresthatarebased on usernetworks, featuresthatarebased on users‘ tweetingbehaviorandfeaturesthatarebased on thelinguisticsofusers‘ tweetcontent.
Forthenetworkfeatureswecreated 3 different typesofusernetworksfromourdatasetandcomputedthefollowingmeasures on these 3 networks.
Coveragebasedmeasuresdescribe e.g. howmanymessagesof a usercontain links orareconversationalorcontainquestionmarks.Diversitybasedmeasuresdescribe e.g. withhowmany different users‘ a usercommunicatesandhowevenlydistributed a users‘ communicationeffortsareacrosstheseusers. A userwhocommunicateswithmanyusersequallymuchwouldhave a high socialdiversitywhile a userwhotendstocommunicatewith a smallcirceloffriendshas a lowsocialdiversity.
Linguistic Inquiry and Word By mapping words in tweets to these 2300 words one gets linguistic annotations of tweets which we used as features.
Wecomputedourfeaturesforeachtargetuserbased on all tweetsthetargetuserhasauthoredduringthechallangeuptothepointwhen he becameinfected. Thatmeanswedid not takeanyinformationintoaccountwhichhappened after a user was alreadyinfectedwhichisimportantsincewewanttopredictinfections. Thereforeweneedtoensurethatwe do not takeanyfutureinfromationintoaccountwhichcouldfalsifyourresults. Forthefollownetworkbasedfeaturesweused a snapshotfromday 4 –allsour sample usersbecamesusceptibelatday 7 orlater.
Soourfirstaim was toidentifyuserswhoarelikelytobecomeinfected. Thatmeanswehad a binaryclassificationproblemandouraim was todiffersusceptiblefrom non-susceptibleusers. Webalancedourdataset, compared 6 classifiersandconducted a 10 corss-foldvalidation. Ourresultsshowthat a generalizedboostedregressionclassifierperformed best. Thereforeweusedthisclassifiertofurtherinspectwhich variables were kost usefulfordifferentiatingbetween...
weusedthebestperformingclassificationmodeltofurtherinspectwhichfeaturesweremostusefulfordifferentiatingbetween...Onecanseefromthisslidethatthe most important features is the out-degree of a user node in the interaction network.It is interesting to note that the top 3 features contain one network feature, one linguistic feature and one behavioral feature which shows that all 3 types of features seem to contribute to our task.ROC curve plots the true positive rate vs. false positive rate. Idea would be if the Area under the ROC curve would be 1.
Wefurtherinspectedthefeaturedistributionsofthe top 20 featuresforeach user-class (i.e. suscepand non-suscept) togainfurtherinsightsintohowfeaturesofsusceptibleusersaredistributedandhow different theirdistributionsarefromthedistributionof non-.susceptibleusers.Best networkfeature: outdegreeofinteractionnetwork– i.e. userswhoactivlycreateinteractionswithothersaremorelikelytobecomeinfected. Best linguisticfeature: verbsandpresenttenseBest behavioralfeature: conversationalvarietyandcoverage
After havingidentifiesuserswho will becomeinfectedduring an attackwe also wanttopredicttheirlevelofinfection: i.e. doestheuserinteract just oncewiththe bot or do theydevelop a closedfrienshiprelation. Thatmeanstheaimofoursecondtaskistopredicthowoftentheyinteractedwith a bot. Toadressthisquestionweusedregressiontreessincetheycan handle...
By fitting the model to our dataset we learned the following tree structure which shows which features and thresholds are used internally by the model. The leaves show the distribution of the suscept score of users who were used as samples for this branch. From this tree structure we can see that…
Toassessthequalityofthismodelwemeasuredthe rank correlationof hold-out usersgiventheir real suscept score andgiventheirpredictedsusceptscores. The correlationcoefficient was prettylowand also the R-squaredvalueofthemodel was prettylow. One potential reasonforthatisthesizeofourdatasetandthatwedid not havemanysamplesofuserswhohadlengthydiscussionswithbots.
So letmestartconcludingmytalk. What I haveyoupresentedtodayis an approachtoidentifysuscepibleuser. Wehaveintroduced a varietyoffeatureswhichcancapturecharacteristicsofuserswhoareindeedmoresuscepibleto bot attacksthanothers.
The factthatactiveTwitterusersaremoresusceptibleis on thehand not reallysurprisingsince...But on theotherhanditissurprisingsinceonewouldexpectthatactiveusersdevelopsomesortofskillytodifferbetween...
Wehopethatourresearch will not onlyinform modern socialmediasecuritysystems but also supportthedevelopmentofgoodsocialbotswhichare e.g. usedtoincreasethefitnesslevelof a community.