CSCI 1800 Cybersecurity and Internaonal Relaonscs.brown.edu/.../files/lectures/Lect18_SocialMediaPropaganda.pdfCSCI 1800 Cybersecurity and Internaonal Relaons Social Media and Propaganda

CSCI1800CybersecurityandInterna4onalRela4ons

SocialMediaandPropagandaJohnE.Savage

BrownUniversity

Outline

•  Informa4onEcosystem•  Simpleexamplesofinforma4onmanipula4on•  Informa4onapocalypsescenarios•  AdvancedAItechnologies•  AI-supportedimpersona4on•  Copingwiththethreat•  Poli4ciza4onofinforma4ontechnology•  Copingwithmisuseofsocialmedia

Lect184/4/18 ©JESavage 2

Informa4onEcosystem

•  Dosocialmediapriori4zeclicks,shares,likesandadsoverqualityinforma4on?– Doesthisresultinamplifica4onof“fakenews?”

•  Hasacri4calthresholdofaddic4veandtoxicmisinforma4onbeenreached?– Willfaithinfactsbelostonthealtaroftechnology?

•  In2016AvivOvadyapredicted*wereachedthepointofinforma4onapocalypseorInfocalypse!

Lect184/4/18 ©JESavage 3

*h^ps://www.buzzfeed.com/charliewarzel/the-terrifying-future-of-fake-news

Informa4onEcosystem

•  SocialmediatoolscoupledwithAIallowusto– Enhanceanddistortreality– Createnewreali4es– Launchinfluencecampaigns– Amplifyreac4ons,e.g.viatrollsandbots– Todoallofthisatadistance

•  Consequence:– Misinforma4on– Propaganda

Lect184/4/18 ©JESavage 4

TrollsandBots

•  NewYorkTimesvideoexplainingtrolls&bots:– h^ps://www.ny4mes.com/video/us/poli4cs/100000005414346/how-russian-bots-and-trolls-invade-our-lives-and-elec4ons.html(5:03)

Lect184/4/18 ©JESavage 5

ExamplesofAr4ficialOutputs

•  Changingaudioinavideo*– h^ps://youtu.be/9Yq67CjDqvw(8:00)– Authorslearnmouthshapesfromspeech– Superimposeshapesonstockfootage

•  Videoswithfullanimatedsuperimposedfaces†– h^ps://www.youtube.com/watch?v=ohmajJTcpNk(6:35)

Lect184/4/18 ©JESavage 6

*h^p://grail.cs.washington.edu/projects/AudioToObama/siggraph17_obama.pdf

†h^p://niessnerlab.org/projects/thies2016face.html

2016Predic4onsofAvivOvadya*

•  Manyslick,easy-to-use,andpowerfultoolstomanipulatepercep4onandfalsifyrealitycoming

•  Thesetoolswillprovidetheabilitytodistorttruthatwillandputcoreins4tu4onsatrisk!

•  Ovadyaholdsposi4onsatU.MichiganCenterforSocialMediaResponsibilityandColumbiaUniversityCenterforDigitalJournalism.

Lect184/4/18 ©JESavage 7

*h^ps://www.buzzfeed.com/charliewarzel/the-terrifying-future-of-fake-news

Ovadya’sDisrup4veScenarios*

•  Diplomacymanipula4on– CreatevideowithKimJungundeclaringwar– Thiswouldprecipitateadiploma4ccrisis

•  Politysimula4on– Createafakegrassrootscampaign– Bombardlegislatorsofficeswithrealis4cpleas

•  Laser(targeted)phishing– Spamusersw.realis4cfakemessagesfromfriends

Lect184/4/18 ©JESavage 8

*h^ps://www.buzzfeed.com/charliewarzel/the-terrifying-future-of-fake-news

AssaultonFCCPublicCommentSystem*

•  In‘17FCCsolicitedpublicinputonnetneutrality– Requiredunder1946Administra4veProcedureAct

•  23millioncommentswerereceived– Morethanallpreviousgovernmentfeedback!

•  SomecommentsgeneratedbyAI-drivenbots.– Somefake,somenot.Numberfakeisunknown

•  Openplatormscanbesubvertedbybots– Congresswantsaninves4ga4on

Lect184/4/18 ©JESavage 9

*h^ps://www.wired.com/story/bots-broke-fcc-public-comment-system/

Wired’sAnalysisofFCCComments

•  HelpedbyFiscalNote,itstudiespubliccommentsonbehalfofcorpora4ons.

•  Large-scaleanalysis*es4mates>1millionfakes•  Small-scaleanalysis†on39NicholasThompson– 6confirmedbots– 11formle^ers– 3realNicholasThompson– 19unknownsource

•  Thisanalysissuggests>halfofcommentsfake

Lect184/4/18 ©JESavage 10

*h^ps://www.wired.com/story/bots-broke-fcc-public-comment-system/†h^ps://www.wired.com/story/bots-form-le^ers-humans-fcc-net-neutrality-comments/

AudioImpersona4on

•  Realis4cimpersona4onofindividualsusingAdobe’s#VoCo– h^ps://youtu.be/I3l4XLZ59iw(7:20)

•  Whatnefarioususescouldbemadeofthistechnology?

•  Howcanweprotectourselves?

Lect184/4/18 ©JESavage 11

NeuralNetworks

•  Ar4ficialneuralnetworks(ANNs)trainedtorecognizevarioustypesofdata,e.g.cats,words

•  Inputsrepresentedbybits,suchaspixelsforimagesorfrequenciesandintensi4esforaudio.

•  Outputsassociatedwithcategories

Lect184/4/18 ©JESavage 12

•  Nodesvaluesareintegers,edgeshaveweights•  Valuesmul4pliedbyweights,passedthroughnon-linearac4va4onfunc4on,givingintegers

•  Weightsadjustedtomakeimproverecogni4on– Adjustmentsmadeviabackpropaga4onoferrors

NeuralNetworks

Lect184/4/18 ©JESavage 13

Mul4plelayerspossible

Genera4veAdversarialNetworks(GANs)

•  GANsarepairsofcompe4ngneuralnets– Onenetgeneratesexamples– Secondnetevaluatestheexamples

•  Compe44ondrivesnetstoimprove– E.g.counterfeitersvspolice

•  GANswereinventedbyIanGoodfellowin2014tomakemachine-learningsystemssmarter– Themethodisunsupervised

Lect184/4/18 ©JESavage 14

CommentsontheThreat

•  Anyonecouldmakeit“appearasifanythinghappened,regardlessofwhetheritdidornot.”– AvivOvadaya*

•  GANshaveboth“imagina4onandintrospec4on”Theycouldsetnewsconsump4onback100years–  IanGoodfellow

•  Computa4onalpropagandaisnowareality!

Lect184/4/18 ©JESavage 15

*h^ps://www.buzzfeed.com/charliewarzel/the-terrifying-future-of-fake-news

WhatDoTechnologiesPutatRisk?

•  Erosionofauthen4cityandintegrityofofficialstatements

•  Electoraloutcomes•  Poten4albreakdownofsociety

Lect184/4/18 ©JESavage 16

Poli4ciza4onofInforma4onTechnologies

•  Russianswerequicktoseethedangers*–  “InvitesallMemberStatestoinform…oftheirviews...on...•  Advisabilityofdevelopinginterna4onalprinciplesthatwouldenhancethesecurityofglobalinforma4onandtelecommunica4onssystemsandhelptocombatinforma4onterrorismandcriminality;

•  AlsoinvolvedinmeddlinginUS2016elec4ons†–  Launched“complexblendofhacking,publicdisclosuresofprivateemails,anduseofbots,trolls,targetedadver4sing”

Lect184/4/18 ©JESavage 17

*h^p://undocs.org/A/RES/53/70†h^ps://www.cfr.org/report/countering-russian-informa4on-opera4ons-age-social-media

RussianMo4va4onforMeddling*

•  Goalistoweakenadversaries,i.e.neighbors,NATOandUSbyanymeans

•  Planted&spreadfalsestoriesin20thcentury•  Socialmediahasmadeiteasier&moreeffec4ve•  FrenchthwartedRussiana^emptinMay2017– Law:can’treportoncampaignsinlast48hours– Ci4zenslargelygetnewsviatradi4onaloutlets

•  RussianstriedtoexploitNFLprotestsin2017Lect184/4/18 ©JESavage 18

*h^ps://www.cfr.org/report/countering-russian-informa4on-opera4ons-age-social-media

Characteris4csofSocialMedia

•  Ci4zensrevealtheirsocial/poli4calpreferences– Profileseasilyassembledandanalyzed– Example:CambridgeAnaly4ca

•  Easytotargetindividualsandgroups– Messagespersonalizedorreceivedfromfriends– Differentgroupscangetcontradictorymessages– Falsereportscanbeinsertedingenuineoutlets

•  Botsspreadtrue&falsenewsatsamerate*

Lect184/4/18 ©JESavage 19

*h^p://science.sciencemag.org/content/359/6380/1146.full

WhatCanBeDoneAboutIt?*•  Improvethetoolsforauthen4ca4ngdocuments–  Images&audiocancryptographicallyverified

•  Iden4fyauthorita4vesourcesandnetworks– Majornewsoutletshavehighstandards– Userecognizedtrustedinterna4onalsources

•  Transparencyanddiscussionareeffec4ve– Quan4fyimpactofinfluencecampaigns–  Iden4fyandpublishsourcesofmisinforma4on

•  Regula4onofexploitablesocialnetworks–  Studyeffectsofsocialnetworksonhumanpsychology

Lect184/4/18 ©JESavage 20

*h^p://www.helsinki4mes.fi/finland/finland-news/domes4c/13884-report-haglund-was-quick-to-pick-up-on-russia-s-informa4on-campaigns.html*h^ps://www.cfr.org/report/countering-russian-informa4on-opera4ons-age-social-media

Review

•  Informa4onEcosystem•  Simpleexamplesofinforma4onmanipula4on•  Informa4onapocalypsescenarios•  AdvancedAItechnologies•  AI-supportedimpersona4on•  Copingwiththethreat•  Poli4ciza4onofinforma4ontechnology•  Copingwithmisuseofsocialmedia

Lect184/4/18 ©JESavage 21