On Microtargeting Socially Divisive Ads: A Case Study of Russia-Linked Ad Campaigns on Facebook
Filipe N. Ribeiro, Koustuv Saha, Mahmoudreza Babaei, Lucas Henrique, Johnnatan Messias, Fabricio Benevenuto, Oana Goga, Krishna P. Gummadi, Elissa M. Redmiles
OOn Microtargeting Socially Divisive Ads:A Case Study of Russia-Linked Ad Campaigns on Facebook
Filipe N. Ribeiro*
UFOP/UFMG, [email protected]
Koustuv Saha*
Georgia Tech, [email protected]
Mahmoudreza Babaei
MPI-SWS, [email protected]
Lucas Henrique
UFMG, [email protected]
Johnnatan Messias
MPI-SWS, [email protected]
Fabricio Benevenuto
UFMG, [email protected]
Oana Goga
Univ. Grenoble Alpes, CNRS,Grenoble INP, LIG, [email protected]
Krishna P. Gummadi
MPI-SWS, [email protected]
Elissa M. Redmiles
University of Maryland, [email protected]
ABSTRACT
Targeted advertising is meant to improve the efficiency of match-ing advertisers to their customers. However, targeted advertisingcan also be abused by malicious advertisers to efficiently reachpeople susceptible to false stories, stoke grievances, and incite so-cial conflict. Since targeted ads are not seen by non-targeted andnon-vulnerable people, malicious ads are likely to go unreportedand their effects undetected. This work examines a specific caseof malicious advertising, exploring the extent to which politicalads from the Russian Intelligence Research Agency (IRA) runprior to 2016 U.S. elections exploited Facebook’s targeted advertis-ing infrastructure to efficiently target ads on divisive or polarizingtopics (e.g., immigration, race-based policing) at vulnerable sub-populations. In particular, we do the following: (a) We conductU.S. census-representative surveys to characterize how users withdifferent political ideologies report , approve , and perceive truth in the content of the IRA ads. Our surveys show that many ads are“divisive”: they elicit very different reactions from people belongingto different socially salient groups. (b) We characterize how thesedivisive ads are targeted to sub-populations that feel particularlyaggrieved by the status quo. Our findings support existing calls forgreater transparency of content and targeting of political ads. (c)We particularly focus on how the Facebook ad API facilitates suchtargeting. We show how the enormous amount of personal dataFacebook aggregates about users and makes available to advertisersenables such malicious targeting. KEYWORDS advertisements, targeting, social divisiveness, news media, socialmedia, perception bias We deployed a system that shows the ads and the demographics of their targetingaudiences (available at ).*
These authors contributed equally to this work.
FAT* ’19, January 29–31, 2019, Atlanta, GA, USA
ACM Reference Format:
Filipe N. Ribeiro*, Koustuv Saha*, Mahmoudreza Babaei, Lucas Henrique,Johnnatan Messias, Fabricio Benevenuto, Oana Goga, Krishna P. Gummadi,and Elissa M. Redmiles. 2018. On Microtargeting Socially Divisive Ads:[0.1em] A Case Study of Russia-Linked Ad Campaigns on Facebook. In
Proceedings of FAT* ’19: Conference on Fairness, Accountability, and Trans-parency (FAT* ’19).
ACM, New York, NY, USA, 10 pages. https://doi.org/10.1145/3287560.3287580
Online targeted advertising refers to the ability of an advertiser toselect audience for their ads. Such advertising constitutes the pri-mary source of revenue for many online sites including most socialmedia websites such as Facebook, Twitter, YouTube, and Pinterest.Consequently, these websites accumulate detailed demographic,behavioral and interest profiles of their users enabling advertisersto “microtarget”, i.e., choose small (tens or hundreds to thousands)of users with very specific attributes like people living in a zipcodethat read New York Times or Breitbart. Beyond raising numerousprivacy concerns [17, 24], targeted advertising platforms have comeunder scrutiny for enabling discriminatory advertising , where adsannouncing housing or job opportunities are targeted to excludepeople belonging to certain races or gender [4, 8, 10, 23].In this paper, we analyze the potential for a new form of abuse ontargeted advertising platforms namely, socially divisive advertising ,where malicious advertisers incite social conflict by publishing adson divisive societal issues of the day (e.g., immigration and racial-bias in policing in the lead up to 2016 US presidential elections).Specifically, we focus on how ad targeting on social media sitessuch as Facebook can be leveraged to selectively target groups ondifferent sides of a divisive issue with (potentially false) messagesthat are deliberately crafted to stoke their grievances and thereby,worsen social discord. We also investigate whether targeted adplatforms allow such malicious campaigns to be carried out instealth, by excluding people who are likely to report (i.e., alert siteadministrators or media watchdog groups about) such ads.Our study is based on an in-depth analysis of a publicly releaseddataset of Facebook ads run by a Russian agency called InternetResearch Agency (IRA) before and during the American Election on a r X i v : . [ c s . S I] N ov AT* ’19, January 29–31, 2019, Atlanta, GA, USAFilipe N. Ribeiro*, Koustuv Saha*, Mahmoudreza Babaei, Lucas Henrique, Johnnatan Messias, Fabricio Benevenuto, Oana Goga,Krishna P. Gummadi, and Elissa M. Redmiles the year of 2016 . Our analysis is centered around three high-levelresearch questions:
RQ 1: How divisive is the content of the IRA ads?
We quantify the divi-siveness of an ad by analyzing the differences in reactions of peoplewith different ideological persuasions to the ad. Specifically, usingUS census-representative surveys, we look at how conservative-and liberal-minded people differ in (a) how likely they are to re-port the ad, (b) how strongly they approve or disapprove the ad’scontent, and (c) how they perceive truthood (or falsehood) in ad’sclaims. Our analysis shows that IRA ads elicit starkly differentand polarizing responses from people with different ideologicalpursuasions.
RQ 2: How effectively done was the targeting of the socially divisiveads?
We find that the “Click Through Rate” (CTR), a traditionalmeasure of effectiveness of targeting, of the IRA ads are an order ofmagnitude (10 times) higher than that of typical Facebook ads. Thehigh CTR suggests that the ads have been targeted very efficiently. Adeeper analysis of the demographic biases in the targeted audiencereveals that the ads have been targeted at people who are morelikely to approve the content and perceive fewer false claims, andare less likely to report.
RQ 3: What features of Facebook’s ad API were leveraged in targetingthe ads?
We also analyze the construction or specification of “tar-geting formulae” for the ads, i.e., the combination of Facebook userattributes that are used when selecting the audience for the ads.We find widespread use of interest attributes such as “Black Con-sciousness movement” and “Chicano movement” that are mostlyshared by people from specific demographic groups such as African-Americans and Mexican-Americans. We show how Facebook adAPI’s suggestion feature may be exploited by the advertisers tofind interest attributes that correlate very strongly to specific socialdemographic groups.
Prior work has highlighted several forms of abuses of targetedadvertising in Facebook, such as for inappropriately exposing theprivate information of users to advertisers [24], and for allowingdiscriminatory advertising (e.g., to exclude users belonging to acertain race or gender from receiving their ads) [23]. Our efforthighlights a new and different form of potential abuse of thesetargeted advertising platforms in creating a social discord.A rich body of prior work have focused on understanding filterbubbles, echo chambers, polarization, and ideological discourse insocial media as an emergent phenomenon [7, 9, 12, 14, 15, 19, 22].We provide a complementary perspective on the topic by examininghow echo chambers and polarization can be engineered on socialmedia through targeted advertising. A recent work conducted adetailed study about Facebook Ads environment by analyzing thou-sands of ads collected through a browser plugin[2]. More closelyrelated to our work, Kim et al. gathered Facebook ads from individ-uals and analyzed who are behind divisive ad campaigns, reportingsuspicious foreign entities [16]. Differently, we focus on understand-ing the disruptive ability of microtargeting for providing divisivepolitical ad campaigns. Figure 1: Example of an Ad from the Dataset.
Finally, our effort is complementary to prior work that attemptsto understand the abuse of social media by misinformation cam-paigns, especially along political elections [18, 25]. Our work pro-vides a better comprehension about a key dissemination mechanismof fake news stories, highlighting how advertising platforms allowinjection of misinformation in social systems and choose vulnerablepeople as the target.
On May 10th, 2018 the Democrats Permanent Select Committeeon Intelligence released a dataset containing 3,517 Facebook adver-tisements from 2015, 2016, and 2017 that are linked to a Russianpropaganda group: Internet Research Agency (IRA).Each ad is composed of an image and text (Figure 1 shows anexample). Additionally, each ad contains a landing page, which isa link to the host of the ad, as well as an ad ID; an ad targetingformula, which is a combination of demographic, behavioral anduser interest aspects used to target Facebook users; the cost forrunning the ad in Russia Rubles ; the number of impressions, whichis the number of users who spent some time observing the ad; thenumber of clicks received by the ad; and, finally, the ad creationand end dates. This section provides an overview of these ads.The ads in the dataset were run between June 2015 and August2017. From the 3,517 advertisements, we found that 617 (17 . . . democrats-intelligence.house.gov/facebook-ads/social-media-advertisements.htm We converted currency of the costs to USD as of May 15th, USD = . RUB. n Microtargeting Socially Divisive Ads:A Case Study of Russia-Linked Ad Campaigns on Facebook FAT* ’19, January 29–31, 2019, Atlanta, GA, USA
Figure 2: Number of ads created, their impressions, cost, andreceived clicks over time. Shaded region shows the 2-monthperiod just before the 2016 U.S. Election.Figure 3: Top 10 Landing Pages based on the number of ads.
We first explore the ad landing pages: the urls to which users whoclicked on the ads were redirected. There are 462 unique land-ing pages corresponding to all the ads. Figure 3 shows the top10 landing pages per number of ads posted. The most popularlanding page ( fb.com/Black-Matters-1579673598947501/ ) posted 259advertisements. Interestingly, one of the top landing pages, the musicfb.info , invites users to install a browser extension, whichwas reported to send spam to the Facebook friends of those whoinstalled it . This landing page received 24,623 impressions, 85clicks, and spent around US$112 .
38. The domain musicfb.info wasalso promoted by other pages, accounting for 3% of all ads. Wealso find that the most popular landing pages are Facebook pages,accounting for 84% of all ads, followed by blackmattersus.com (7%),and Instagram (3 . Figure 4 (a) shows the cumulative distribution functions (CDFs) ofall the ads in the dataset on their number of impressions, clicks,and amount spent to advertise. The most expensive ad cost 5,307 $ web.archive.org/web/20161019155736/musicfb.info/ wired.com/story/russia-facebook-ads-sketchy-chrome-extension/ (a) (b) Figure 4: Cumulative Distribution Function (CDF) of the adson their (a) clicks, impressions, and costs, (b) click-through-rates.
USD. The highest number of impressions generated was 1,335,000and the maximum number of clicks was 73,060.Nearly 25% of the landing pages spent more than 100 dollars,26 .
8% of the pages received more than 1,000 clicks, and around36 .
1% had more than 10 ,
000 impressions. On the other hand, morethan 25% of the ads had no impressions, clicks, and cost, suggestingthese ads were not launched or ran for a very short period of time.An average ad cost 34 . . .
18% of the total cost, 71 .
93% of the total number of impressions,69 .
47% of the total number of clicks.However, there were notable exceptions to this correlation: higherinvestment (cost) did not always lead to higher return (e.g., im-pressions, clicks). Table 1 shows the most popular landing pagesper impressions, clicks, and cost of the ads. For example, fb.com/brownunitedfront/, received the largest number of impressions(5,817,734), corresponding alone to 14 .
3% of impressions obtainedby all ads, but cost only 6 .
5% of the total cost of all ads in the dataset.Finally, we compute the click-through rate (CTR) of these ads,which is a typical metric to measure the effectiveness of an ad.It is computed as a ratio between the number of clicks and thenumber of impressions received by an ad. Figure 4 (right) shows thecumulative distribution function of the CTR of the ads, excludingthose with 0 values for clicks, impressions, and cost. The medianCTR is 10 .
8% and 75% of the ads have a CTR higher than 5 .
6. Theaverage CTR is 10 . which shows the average CTR for Facebook ads across all industriesis 0 . . . . wordstream.com/blog/ws/2017/02/28/facebook-advertising-benchmarks AT* ’19, January 29–31, 2019, Atlanta, GA, USAFilipe N. Ribeiro*, Koustuv Saha*, Mahmoudreza Babaei, Lucas Henrique, Johnnatan Messias, Fabricio Benevenuto, Oana Goga,Krishna P. Gummadi, and Elissa M. Redmiles
Impressions Clicks Cost (USD)fb.com/brownunitedfront/ 14.3% fb.com/brownunitedfront/ 18.8% fb.com/patriototus/ 6.5%fb.com/blacktivists/ 10.8% fb.com/Blacktivist-128371547505950/ 13.8% fb.com/blacktivists/ 5.4%fb.com/Blacktivist-128371547505950/ 10.5% fb.com/blacktivists/ 11.9% fb.com/blackmattersus/ 5.3%fb.com/blackmattersus.mvmnt/ 4.7% fb.com/blackmattersus.mvmnt/ 7.0% fb.com/timetosecede/ 4.7%fb.com/Woke-Blacks-294234600956431/ 3.3% fb.com/Dont-Shoot-1157233400960126/ 3.6% fb.com/Igbtun/ 4.3%fb.com/copsareheroes/ 3.3% fb.com/blackmattersus/ 2.5% fb.com/BlackJourney2Justice/ 4.1%fb.com/blackmattersus/ 3.1% fb.com/patriototus/ 2.5% fb.com/MuslimAmerica/ 3.28fb.com/South-United-1777037362551238/ 2.7% fb.com/Memopolis-450474615151098/ 2.4% fb.com/South-United-1777037362551238/ 3.2%fb.com/Dont-Shoot-1157233400960126/ 2.2% fb.com/Woke-Blacks-294234600956431/ 2.3% fb.com/blackmattersus.mvmnt/ 2.7%fb.com/patriototus/ 1.7% fb.com/South-United-1777037362551238/ 2.0% fb.com/savethe2a/ 2.5%
Table 1: Most popular landing pages per impressions, clicks, and cost.
Our analysis reveals that only a few ads are responsible for mostof the cost, impressions, and clicks. Considering this, we defined aset of high impact ads as the union of the top 10% ads in terms ofcost, impressions, clicks, and CTR. We obtained 905 high impactads, corresponding to 27 .
7% of the entire dataset. These ads accounttogether to 83 .
9% of the total number of impressions, 81 .
8% of clicks,88 .
5% of the cost, and 46 .
9% of the CTR. For the purposes of ourstudy, where we require manual inspection of the ads (to identifytheir targets and to run surveys), our ensuing analyses concernthose high impact ads run before the 2016 U.S. elections: 485 ads.
This section describes and characterizes the ads in the IRA dataset.Our analysis highlights the landing pages that paid for the ads andidentifies the most successful ads in terms of impressions and clicks.We find that the ad campaigns were intensified near to the U.S.election period. Among our main findings, we show that the typicalCTR for these ads is an order of magnitude higher than typicalvalues for Facebook, meaning that these ads were very effective.
To investigate whether these ads were designed to be ideologicallydivisive – that is, designed to elicit different reactions from peoplewith different political viewpoints – we conducted three onlinesurveys on a U.S. census-representative sample (n=2,886). We usedeach survey to measure one of three axes along which ads couldpotentially be divisive: 1) reporting : whether respondents wouldreport the ads, and why, 2) approval and disapproval : whether theyapprove or disapprove the content of the ad, and 3) false claims : ifthey are able to identify any false claims in the content of the ad.Our surveys considered only those 485 high impact ads whichwere run before the elections. Each survey showed ten ads followedby demographic questions. More detail on the specific questionsused to assess each axis is provided in the corresponding axis sub-sections that follow. The survey questions were pre-tested usingcognitive interviews and all survey questions included a “I don’tknow” or “Prefer not to respond” answer choice to ensure internalmeasurement validity [6]. To obtain a demographically representa-tive sample, and ensure that we captured a wide variety of American perceptions, we deployed the surveys using the Survey Sampling In-ternational survey panel , a non-probabilistic census-representativesurvey panel. For each survey, we sampled at least 730 respondents(15 responses per ad) whose demographics were representative ofthe U.S. within 5% and who had a range of political views (40%liberal, 40% conservative, and 20% moderate or neutral); across thethree surveys we obtained a total sample of 2,886 respondents.We measured overall ideological divisiveness on the three axes(reporting, approval, and false claims) using two metrics: Within-group divisiveness.
Within-group divisiveness measuresthe extent to which respondents’ answers about a particular ad areconsistent with their political ideology. That is, do all liberals answersimilarly about a particular ad. For each ad, we first calculate thestandard deviation of all the responses, and then we calculate thestandard deviation of the responses within a particular ideologicalgroup. Next, we compute within-group divisiveness as the fractionof within-group standard deviation to the overall standard deviation.Therefore we interpret values lower than 1 as lower divisiveness(and greater agreeableness) within a group than overall, and valuesgreater than 1 as greater within-group divisiveness than overall.
Between-group divisiveness.
Between-group divisiveness mea-sures the extent to which answers from respondents of one politicalideology differ from answers of respondents who align with anotherpolitical ideology. That is, do liberals answer differently about aparticular ad than conservatives. For an ad, we calculate the differ-ence between the mean responses per ideological group, and thencompute the fraction of this difference over the maximum possibledifference given the range of values to obtain the between-groupdivisiveness measure. This limits the range of between-group divi-siveness measure between 0 and 1, where higher values indicategreater divisiveness between ideological groups.Table 2 summarizes the divisiveness of the high impact ads. Wefind that the within-group divisiveness measure is lower than 1for all our surveys. This indicates high agreeableness within theideological groups. In addition, about 20% of the ads show between-group divisiveness higher than 0.5, indicating severe divisivenessbetween ideological groups for those ads.
The first axis of divisiveness that we explored was reporting. Wesurveyed respondents regarding: 1) Whether they would report n Microtargeting Socially Divisive Ads:A Case Study of Russia-Linked Ad Campaigns on Facebook FAT* ’19, January 29–31, 2019, Atlanta, GA, USA Measure (Group) Reporting Approval False ClaimsMean Stdev. Mean Stdev. Mean Stdev.
Within-group divisiveness
Liberals 0.87 0.47 0.92 0.36 0.66 0.69Conservatives 0.90 0.43 0.98 0.31 0.86 0.63
Between-group divisiveness
Political 0.24 0.18 0.34 0.24 0.17 0.14
Table 2: Divisiveness measures of the high impact ads. . . . . . . . . . Reported Proportion . . . . . . P r opo r i t ono f A d s (a) Reported Ads % Reported Responses It’s sexually inappropriateIt’s violent or prohibited contentIt’s offensiveIt’s misleading or a scamI disagree with itIt’s a false news storyIts a spamSomething else (b) Reasons of Inappropriateness
Figure 5: Distribution of the high impact ads on the (a) pro-portion of reported ads in our dataset, (b) reasons of inap-propriateness. the ad shown? , and 2) If they would, why do they find the adinappropriate? Answer choices given, drawn directly from Face-book’s reporting interface [11], were: sexually inappropriate , violent , offensive , misleading , disagree , false news , spam , and something else .Figure 5 shows the reporting responses for the high impact IRAads. For over 73% of these ads, at least 20% of the respondentsresponded that they would have reported the ads. We observe thatthe majority of the ads were reported on the grounds of beingoffensive (25%), violent (15%), and misleading (15%). Additionally, asubstantial proportion (9%) of the reported responses belonged tothe something else category. In such cases, the respondents enteredfree-text to explain their reason for inappropriateness. Out of the 61responses that we received in the free-text box, the pre-dominantreasons were that the ad incites racism (20%), and that the ad createsdivide (5%) in the society.Next, to examine ideological divisiveness, we find that the meanwithin-group divisiveness is 0.87 (stdev = 0.47) for liberals and0.90 (stdev = 0.43) for conservatives. Both of these within-groupdivisiveness measures being less than 1, suggests that the likelihoodwith which individuals within the same ideological group agreeabout reporting an ad is higher than that when compared againstindividuals across ideological groups.Figure 6 (a, b) shows the distribution of the reporting acrossideological groups. We find significant differences in terms of thereporting behavior across political ideologies. Defining a medianthreshold for divisiveness, we find that in over 50 percent of the ads, Specifically, we asked “Some social media platforms allow you to report content byclicking "report". Would you report this ad (e.g., Mark it as inappropriate or offensive)”With answer choices “Yes”, “No”, “I don’t know”. Ad . . . . . . M ean G r oup R epo r t ed P r opo r t i on LiberalsConservativesOverall (a) Ideological differences . . . . . . . . . . between-group divisiveness . . . . . . P r opo r t i ono f A d s Mean (b) Between-group divisiveness
Figure 6: Distribution of reporting across ideological groups.(a) shows the distribution o proportion of the ads being re-ported by either of political ideology, with x-axis containingeach of the high impact ads, (b) plots the between-group di-visiveness for the high impact ads.
Reported by both liberals and conservatives
TAG YOUR PHOTOS WITH
Reported predominantly by liberals.
Joinustolearnmore!Whyaren’twhitehoodsandwhitesupremacistpropagandaillegalhereinAmerica?WhyareGermansashamed of their bigotry, while America is proud of it? Black America( @black BlacklivessssWe simply can’t allow Muslims to wear burga, otherwise everybody who wants to commit a crime or terror attack wouldwear this ugly rug and hide his or hers identity behing it. The risk is too highl Burga and other face covering cloth shouldbe banned from wearing in public!Five police officers were killed in an organized attack during the protest in Dallas this Blue Lives MatterBlack intelligence is one of the most highly feared things in this country.Parasite is an organism that lives in or on another organism and benefits by deriving nutrients at the host’s expense. About20 million parasites live in the United States illegally. They exploit Americans and give nothing in return. Isn’t it time to getrid of parasites that are destroying our country?
Reported predominantly by conservatives.
Come and march with us on 16 April. Stand with Baltimore. Let’s make change! Freddie Gray Anniversary MarchClick Watch More to join us! Let’s fight against police brutality together! donotshoot.us Donotshoot.us Don’t ShootThe USA is exactly the place where cops can’t care less about people’s civil rights. They are cynical toward the rule of lawand disrespectful of the rights of fellow citizens. Details: http://donotshoot.us/We Muslims of the United States are subject to Islamophobia from the media where regularly STOP SCAPEGOATING MUS-LIMS!People, our race is in danger! Together we are an invincible power. Just say your word! Join us! Black Pride
Table 3: Example ads on the basis of reporting behavior bythe respondents from two political ideologies. liberals and conservatives completely disagreed with each other (eg.conservatives showed more than their median reported proportionand liberals showed less than their median reported proportion,and the vice versa). Table 3 shows a few examples of the ads whichshowed the greatest differences in the reporting behavior by therespondents of two political ideologies. These ads typically mentionpolitically-charged topics. For example, immigration — “TAG YOURPHOTOS WITH
AT* ’19, January 29–31, 2019, Atlanta, GA, USAFilipe N. Ribeiro*, Koustuv Saha*, Mahmoudreza Babaei, Lucas Henrique, Johnnatan Messias, Fabricio Benevenuto, Oana Goga,Krishna P. Gummadi, and Elissa M. Redmiles . . . . . . . . . . Proportion Responses . . . . . . P r opo r i t ono f A d s ApprovalDisapproval (a) CDF: Proportion Responses Ad . . . . . . . . . P r opo r t i on R e s pon s e s ApprovalDisapprovalNeither (b) Approval per ad . . . . . . between-group divisiveness . . . . . . P r opo r t i ono f A d s Mean (c) Between group divisiveness Ad − . − . − . − . . . . . . . M ean G r oup A pp r o v a l S c o r e LiberalsConservativesOverall (d) Ideological Differences
Figure 7: Distribution of the ads on approval and disap-proval: (a&b) overall, (c&d) across ideological groups. (a&c)plot the cumulative distribution functions (cdfs), (b&d) plotthe differences in approval in each ad, where x-axis consistsof all the ads.
As another characterization of people’s reactions to the ads, weasked respondents in a second survey whether they approve ordisapprove of a particular ad, and how strongly they approve ordisapprove. These questions in the survey were constructed basedon questions about political preference that have been extensivelypre-tested by Pew Research for previous surveys about politicalpolarization [6]. We find that 87% of the adds were approved and 63%of the ads were disapproved by at least 20% respondents (see Figure 7(a)). To quantify the received responses, we assigned an approvalscore on a 5 point scale with values of -2 (strong disapproval), -1(weak disapproval), 0 (neither approve or disapprove), +1 (weakapproval), and +2 (strong approval). While computing the meanapproval score for a group, we dropped the 0 responses to ensurethat a mean approval score close to 0 corresponds to similar weightsfrom approval and disapproval. Table 4 lists some example ads alongwith their approval tendencies by the two ideological groups withinour dataset.Figures 7 (c&d) show the relationship between respondents’ideology and approval of ad content. We observe that the mean Specifically, we asked “Do you approve or disapprove of what the ad says or im-plies?” Answer choices: Approve; Disapprove; Neither; There is nothing in this adto approve or disapprove of; I don’t know. Followed by a measure of strength “Doyou [approve/disapprove] very strongly, or not so strongly?” if the prior question wasanswered with approve or disapprove.
Approved by both the liberals and the conservatives
Show up, fight racism and take a stand for equality. Monday, May 2 at 4 PM at Erie County Holding Center Justice ForIndia: Not 1 More!Alton Sterling, an innocent 37-year-old Black male, was outrageously executed by two Baton Justice For Alton SterlingDid you see this? Damn... We lost count of how many mentally ill citizens were murdered during encounters with violentcops. Here is another woman suffering from mental illness. Full story: http://bit.ly/10rglhkJoin Us! Support The Police!Darkness cannot drive out darkness; only light can do that. Hate cannot drive out hate; only love can do that. Martin LutherKing, Jr.
Disapproved by both the liberals and the conservatives
No wonder white boys don’t get shot when they’re arrested’:Anti-immigration is the only salvation!O000ps seems like someone screwed up! Salute our brave and smart cops who mistake man’s member for a deadly weapon.Follow US and stay WOKE!It’s ok they’re women so they’ll only find the kitchenThis man beat up police officer who tased his wife. Do you agree with the man who defended his woman?
Approved by the liberals and disapproved by the conservatives
Two years have passed since August 11, 2014, the date. when 25-year-old Ezell Ford was mur Justice For Ezell Ford AndDonnell ThompsonWe don’t want to honor racism, slavery and hatred. This is what Confederate Heritage is. Not My Heritage RallySay it loud: I’m black and I’m proud!We Muslims of the United States are subject to Islamophobia from the media where regularly STOP SCAPEGOATING MUS-LIMS!Click to Learn More! Everybody knows that Islam is against terrorism but not everyone believes this fact! Islam does notsupport terrorism under any circumstances. Terrorism goes against every principle in Islam. In fact if a Muslim engages interrorism, he is not following Islam and so he is not a Muslim! America, stop insulting peaceful citizens, stop taking all ofus as criminals, we don’t deserve such attitude!
Approved by the conservatives and disapproved by the liberals
Heritage not hate y’all! Our flag has nothing to do with racism! The Federal Government shouldn’t be able to dictate whatwe can and cannot do. Go follow Confederate page
Table 4: Example ads on the basis of the approval behaviorby the respondents from two political ideologies. within-group divisiveness for liberals is 0.92 (stdev = 0.36) and 0.98(stdev = 0.31) for conservatives (Table 2). Both the within-groupdivisiveness values being lower than 1, suggests that the likelihoodwith which individuals within the same ideological group wouldagree about approving an ad is higher than that when comparedagainst individuals across ideological groups. The divisiveness inapproval responses is further confirmed by the between-groupdivisiveness measure which ranges between 0 and 1 (mean = 0.34)across the high impact ads.
To examine whether the high impact IRA ads contained any falseclaims, in another survey we asked the respondents if they couldidentify any false claims present in the ads. We find that 89%(433 out of 485) of the high impact ads were identified to haveat least one false claim, and about 45% of the ads contained falseclaims according to 10% of the respondents. Figure 8 (a) shows thecumulative distribution of the ads with the number of respondentswho identified at least one false claim in them.Next, as in the other two content analyses, we examined whetherrespondents’ ideology related to their perception of the presence offalse claims (Figure 8). Both the within-group divisiveness valuesbeing lower than 1, suggests that the likelihood with which individ-uals within the same ideological group would agree about findingfalse claim in an ad is higher than that when compared againstindividuals across ideological groups. Table 5 shows a sample of Specifically, we asked respondents to “Please copy and paste any phrases or sentencesin the advertisement that you think contain a “factual claim”. That is, something thatsomeone could verify as True or False. If you cannot identify any claims, please type“No Claims” in the first box.” We then asked them to label the phrases they had identifiedas “True”, “False” or “not sure whether they are True or False”. n Microtargeting Socially Divisive Ads:A Case Study of Russia-Linked Ad Campaigns on Facebook FAT* ’19, January 29–31, 2019, Atlanta, GA, USA
False Claims identified by both liberals and conservatives
Bernie Sanders has proven himself to be the best candidate in every way. He is fair and strong and he is the only one fightingfor the black community. For more than 40 years he is advocating against any injustice and stays true to his moral values.Vote for Bernie!Illegal immigrants are not only flooding our country with drugs and sinking our economy, they are a major threat to Amer-icas integrity. They don’t care about American laws, history and heritage. They just want our money and social benefits.Considering the fact that they multiply like rabbits, soon we will all wear sombreros. We need to stop this invasion.When you live in Texas you know that you are the chosen one!It is time to wake up and see the truth. Cops are not our friends and government doesn’t care about you. Share this if you’reawake!.It’s ok they’re women so they’ll only find the kitchen
False claims identified by liberals.
ClintonsaidtheUnitedStatesneedstoconfrontthe’systematicracism’initslawenforcementefforts.We’resickofpoliticiansorganizing and leading the systematical propaganda against our police. It is unfair and vilely to accuse our heroes of everysin and crime. In fact, the efficiency of our cops resulted in a decrease of the average amount of crimes, especially in largecities. Law-abiding citizens should never fear cops, but criminals do. And that’s why Hillary is on the criminals’ side. Joinour rally on July, 23th in New York City, it’s time to show Clinton that we will never let her become our next President!It might sound like a cliche but "get a job" is a really good advice for young liberals protesting against everything in theworld. Old man Ronald knew what he was talking about! Our college students should have an experience of paying taxesbefore standing for illegal immigrants’ rights. They should rise their own children before standing for gay parenthood. It’sno secret most active liberal’s supporters are people about 20-25 years- old while most conservatives are older. Well; as theysay; wisdom comes with ages.HisfailedmedicalreformandunbelievablenationaldebtisenoughtoputObamabehindbars.butthat’snotall.Hisgreatest"accomplishment" is flooding America with countless criminals and giving them all an absolute omnipotence. Thanks toBarack Hussein Obama we have at least one big terror attack each year; not to mention illegals raging out and poisoningour country with drugs. For what he did to America Obama should rot in prison for the rest of his life.Border Patrol agents in South Texas arrested an illegal alien from Honduras that had previously been deported and convictedof Rape Second Degree. Thanks to Obama’s and Hillary’s policy, illegals come here because they wait for amnesty promised.The wrong course had been chosen by the American government; but all those politicians are too far from the border to seewho actually sneaks through it illegally. Rapists, drug dealers, human traffickers; and others. The percent of innocent poorfamilies searching for a better life is too small to become an argument for amnesty and Texas warm welcome.Anti-immigration is the only salvation!
False claims identified by conservatives.
Don’t Shoot is a community site where you can find recent videos about outrageous police misconduct, really valuable onesbut underrepresented by mass media. We provide you with first-hand stories and diverse videos. Join us! Click Learn more!We don’t want to honor racism, slavery and hatred. This is what Confederate Heritage is. Not My Heritage RallyThe USA is exactly the place where cops can’t care less about people’s civil rights. They are cynical toward the rule of lawand disrespectful of the rights of fellow citizens. Details: http://donotshoot.us/Police are beyond out of control, help us make this viral! Follow our account in order to spread the truth!Join us to study your blackness and get the power from your roots. Stay woke and natural! Nefertiti’s Community
Table 5: Example ads on the basis of false claims identifiedby the respondents from two political ideologies. Identifiedfalse claims are highlighted in pink. . . . . . . Proportion of Respondents . . . . . . P r opo r i t ono f A d s Mean (a) Proportion of FCs Identified Ad . . . . . . . . . M ean G r oup F C P r opo r t i on LiberalsConservativesOverall (b) Political Ideology
Figure 8: Distribution of the ads on false claims (FCs): (a)overall (as a cumulative density function), (b) across ideo-logical groups (where each ad is plotted on the x -axis). ads and false claims identified by respondents from each ideologicalgroup (liberal, conservative). This section focuses on peoples’ perceptions of the content of the485 IRA ads we identified as high impact. To assess these percep-tions along three axes – likelihood of being reported, approvaland disapproval, and the presence of false claims – we conducted
Figure 9: Top 20 attributes based on the number of advertise-ments they appeared. three U.S. census-representative surveys. Our analysis of the per-ceptions queried in these surveys shows that ideological opinionsof individuals influence their perceptions of these ads. We find thatmany of these ads were severely divisive, and generated stronglyvaried opinions across the two ideological groups of liberals andconservatives (see Figure 6, 7, 8).
Next, we focus on understanding how the target formula is createdby advertisers and the role that Facebook interface plays on that.
The Facebook ads platform provides three approaches for advertis-ers to target people [3, 23], briefly described next.
Personally Identifiable Information (PII) targeting is the form inwhich advertisers provide personal information about users suchas name, phone number, and email address so that Facebook candirectly place the ads to them. This kind of targeting does not appearin the IRA dataset.
Look-alike audience target . For this targeting option, advertisersprovide to Facebook a list of users similar to that one in the PIIor a list of people who liked the advertiser Facebook page. Then,Facebook attempt to target a similar audience to the group in thisspecific list. Only 1 .
1% of the high impact ads used this option.
Attribute-based targeting allows the advertiser to create a targetformula based on a wide range of elements that include user ba-sic demographics (i.e. gender, age, location, language), advanceddemographics (i.e. political leaning, income level, ‘Parents withchildren preschoolers’), interests (i.e. newspapers, religion, politics),and behaviors (i.e. ‘Business Travelers’ or ‘New Vehicle buyers’).Recent work showed that the number of possible interests provideby Facebook is greater than 240,000 [23]. Facebook allows one toinclude or exclude users with each of those attributes and combinemultiple attributes as part of a target formula. The vast majority ofthe high impact ads, 895 out of the 905, used this option to elaboratea formula. We found that 78% of the ads used 2 or more interestsand behaviors in their formula, creating very complex formulaswith up to 39 distinct attributes.
AT* ’19, January 29–31, 2019, Atlanta, GA, USAFilipe N. Ribeiro*, Koustuv Saha*, Mahmoudreza Babaei, Lucas Henrique, Johnnatan Messias, Fabricio Benevenuto, Oana Goga,Krishna P. Gummadi, and Elissa M. Redmiles
Figure 10: Cumulative Distribution Function (CDF) for thenumber of suggestions.
Figure 9 shows the top attributes that appear in the ads targetformula based on the number of times they appeared in different ads.There were 497 distinct attributes and the most present attributesinterest were African-American history and African-American CivilRights Movement (1954-68), appearing in 295 (32%) ads. We cannote a prevalence of attributes related to African-American andHispanic Population, with interests like Mexico, ‘Hispanidad’ and‘Latin hip hop’. Next, we investigate aspects of the Facebook adsplatform design that might have favored the IRA ads to massivelyexplore this particular targeting strategy.
Facebook provides a tool for advertisers that, given a target attribute,it presents a list of other attributes that target people with similardemographic aspects [23]. For example, in the list of suggestedtargeting interests for ‘Townhall.com’, a page with an audiencein which 79 .
5% of the users are very conservative users accordingto Facebook, there are other pages with similar bias towards veryconservative users, i.e. ‘The Daily Caller’ (67 . . . appear as suggestionsfor at least one of the others in the formula. For ad ID 1840 , wewere able to find 9 out of 10 of the interests using the interestsuggestion feature. This provides evidence that this feature mayhave been a key element used by the IRA campaign to choose thetarget audience. In this Section we show that the vast majority of the IRA ads useattribute-based targeting, containing complex target formula thatincludes interest and behavioral attributes that are likely suggestedby Facebook. Next, we investigate the extent to which these formu-las allowed advertisers to reach demographic biased audiences.
We start by describing our methodology to reproduce the IRAqueries (without running the ad) and gather the demographicsof the of the targeted users.
Before launching an advertisement in Facebook, the advertiser canget the estimated audience (i.e., the number of monthly active users)likely to match the target formula. Our methodology consists ofusing the Facebook Marketing API to reproduce the targetingformula of all high impact IRA ads and get the demographics ofthe population that matches each targeting formula, without run-ning any ad. This methodology has been extensively used recentlyfor different purposes, including inferring news outlets politicalleaning [20], study migration [26] and gender bias [13] across coun-tries, and for public health awareness [21] and lifestyle diseasesurveillance [5]. For our analysis, we considered seven demographiccategories: political leaning, race, gender, education level, income,location (in terms of states), and age. As a baseline for comparison,we also gathered the demographic distribution of the United StatesFacebook population.Only 11% of the used attributes that appear in the IRA ads tar-geting formulas are not available for targeting anymore due tochanges in the Facebook Marketing API. In most of these cases,we reproduced the ad target formula without the missing attribute,especially when the attribute looks redundant with the others inthe formula. We did not reproduce only 6 targeting formulas. To assess the audience bias of each of the demographic aspects thatwe considered, we computed the differences between the fractionof the population with a demographic aspect and the same fractionof the population in the baseline distribution (i.e. the U.S. Facebookpopulation), namely the bias score . For instance, if the percentageof African-Americans in the audience of a particular ad is 40%, the bias score for this dimension in the ad is 0.25 as the percentage ofAfrican-American in the U.S. Facebook population is nearly 15.5%(0 . − . developers.facebook.com/docs/marketing-apis n Microtargeting Socially Divisive Ads:A Case Study of Russia-Linked Ad Campaigns on Facebook FAT* ’19, January 29–31, 2019, Atlanta, GA, USA Figure 11: Bias in demographic dimensions. Each violin rep-resents the bias score for all high impact ads in a particu-lar demographic dimension. The median is represented by awhite dot in the center line of the violin. 50% of the data ispresent between the two thick lines around the center
Group Report Approval False ClaimsLiberals -0.17*** 0.41*** -Conservatives -0.15*** 0.32*** -
Table 6: Pearson’s r correlation between targeting and theideological divisiveness for the high impact ads (*** p < . , no statistical significance in the case of false claims). percentage of ads with bias score superior to 0.15 is 52% for African-American and 41% for Liberals. Our dataset suggests the presence ofthose ads that target extremely biased populations of conservatives,Liberals, Hispanic, and especially African-Americans. The targetaudiences for the IRA ads are slightly biased towards women andyoung adults (18-34 years), which are omitted from Figure 11 dueto space constraints. Next, we investigate if the advertisers target the ads towards audi-ences that are less likely to identify their inappropriateness due totheir ideological perception bias. Additionally, we examine if theads directed to biased audiences could leverage the already existingsocietal divisiveness to further amplify it among the masses.To understand these nuances of targeted advertising, in this sec-tion, we focus on the relationship between the targeted populationand the ideological divisiveness in reporting, approval, and falseclaim identifying behaviors for the ads. Table 6 reports the correla-tion values between the targeted population and the tendency ofthe population to report, approve, and identify false claims.
Reporting.
We observe a negative correlation in the case ofreporting for both Liberals and Conservatives (also see Figure 12 (a)).This suggests that the targeted population has a lower tendency toreport than the non-targeted one. This is also evident per Figure 12(b), where we find that the reporting by the targeted populationcarries way lower likelihood than the reporting by the overall (ornon-targeted) population. . . . . . . Liberals . . . . . . C on s e r v a t i v e s Cns. ReportedLib. Reported (a) Reporting: Targeted Ad . . . . . . . . . P r opo r t i on R epo r t ed Non-TargetedTargeted (b) Reporting: Bias . . . . . . Liberals . . . . . . C on s e r v a t i v e s Cns. ApprovedLib. Approved (c) Approval: Targeted Ad − . − . . . . . A pp r o v a l S c o r e Non-TargetedTargeted (d) Approval: Bias . . . . . . Liberals . . . . . . C on s e r v a t i v e s Cns. Find FCsLib. Find FCs (e) False Claims: Targeted Ad . . . . . . . . . . P r opo r t i on F C s Non-TargetedTargeted (f) False Claims: Bias
Figure 12: Relationship between targeting and the responsesby ideological groups. (a,c,e) show the proportion of popu-lation targeted and their tendency of response. Each circlerepresents an ad, and their size is proportionate with the be-tween group disputability for that ad. (b,d,f) compares themean responses of the targeted ads with their hypotheti-cal non-targeted counterpart (i.e., overall responses), whereeach ad is represented on the x -axis Approval.
We observe a positive correlation in the case of ap-proval for both Liberals and Conservatives (also see Figure 12 (c)).This suggests that the targeted population has a greater tendency toapprove the ads as compared to the non-targeted population. Thisis also evident per Figure 12 (d), where we find that the approvalscore by the targeted population carries greater score for a majorityof the ads compared to the overall (or non-targeted) population.
AT* ’19, January 29–31, 2019, Atlanta, GA, USAFilipe N. Ribeiro*, Koustuv Saha*, Mahmoudreza Babaei, Lucas Henrique, Johnnatan Messias, Fabricio Benevenuto, Oana Goga,Krishna P. Gummadi, and Elissa M. Redmiles
False claims.
For false claims, we do not find any significantcorrelation between the targeted population and divisiveness. How-ever, per Figure 12 (e&f) we do find that the targeted populationhas a lower tendency to identify false claims.Taken together, we can assume that the ads were “well-targeted”in a way towards that population which was more likely to believe,and approve and subsequently less likely to report or identify falseclaims in them.
Our findings show that the IRA ads reached audiences that are verybiased towards African-Americans and Liberals. More important,we show that ads were overall targeted towards a population thatis more likely to believe, and approve and subsequently less likelyto report or identify false claims in them.
In this paper, we provide an in-depth quantitative and qualitativecharacterization of the Russia-linked ad campaigns on Facebook.Our findings suggest that the Facebook ads platform can be abusedby a new form of attack, that is the use of targeted advertising tocreate social discord. These ads showed to be divisive, were 10 timesmore effective than a typical Facebook ad, were biased especially interms of race and political leaning, and tended to be targeting morethe users who are less likely to identify their inappropriateness. Wealso provide strong evidence that these advertisers have exploredthe Facebook suggestions tool to engineer the targeted populations.While this tool may be helpful in many ways, it needs to becarefully redesigned to avoid that a malicious advertiser reaches soeasily groups of vulnerable people. For example, Facebook recentlypresented its intention to manually inspect ads before they arelaunched [1], aiming to guarantee that ads do not divide or discrim-inate people. Our work suggests that the priority of the candidatesto be manually inspected can be based on their targeting formula.For instance, those ads that target extremely narrowed populations,on the basis of race, political leaning, and other sensitive topicshave greater likelihood of being divisive. Additionally, the ads thatexperience severely high click-through rates could also be flaggedto be quickly inspected.As a final contribution, we have deployed a system (available at ) that displays the adsand their computed information such as the demographics of theirtargeting audiences.
F. Benvenuto and F. Ribeiro acknowledge grants from Capes, CNPq,and Fapemig. E. M. Redmiles acknowledges support from the U.S.National Science Foundation Graduate Research Fellowship Pro-gram under Grant No. DGE 1322106 and from a Facebook Fellow-ship. This research was partly supported by an European ResearchCouncil (ERC) Advanced Grant for the project “Foundations for FairSocial Computing”, funded under the European Union’s Horizon2020 Framework Programme (grant agreement no. 789373). Thisresearch was partly supported by ANR through the grant ANR-17-CE23-0014.
REFERENCES [1] [n. d.]. Reviewing Targeting to Ensure Advertising is Safe and Civil.[2] Athanasios Andreou, Márcio Silva, Fabrício Benevenuto, Oana Goga, PatrickLoiseau, and Alan Mislove. 2019. Measuring the Facebook Advertising Ecosystem.In
NDSS ’19 .[3] Athanasios Andreou, Giridhari Venkatadri, Oana Goga, Krishna P. Gummadi,Patrick Loiseau, and Alan Mislove. 2018. Investigating Ad Transparency Mecha-nisms in Social Media: A Case Study of Facebook’s Explanations. In
NDSS ’18
WebSci ’17 . ACM, Troy, NY, USA.[6] Paul C Beatty and Gordon B Willis. 2007. Research synthesis: The practice ofcognitive interviewing.
Public Opinion Quarterly
71, 2 (2007), 287–311.[7] Carlos Castillo, Mohammed El-Haddad, Jürgen Pfeffer, and Matt Stempeck. 2014.Characterizing the life cycle of online news stories using social media reactions.In
CSCW ’14 . ACM, 211–223.[8] Amit Datta, Michael Carl Tschantz, and Anupam Datta. 2015. Automated Experi-ments on Ad Privacy Settings.
PoPETs
Scientific reports
Public opinion quarterly
80, S1 (2016), 298–320.[13] David Garcia, Yonas Mitike Kassa, Angel Cuevas, Manuel Cebrian, Esteban Moro,Iyad Rahwan, and Ruben Cuevas. 2017. Facebook’s gender divide. arXiv preprintarXiv:1710.03705 (2017).[14] Kiran Garimella, Gianmarco De Francisci Morales, Aristides Gionis, and MichaelMathioudakis. 2018. Political Discourse on Social Media: Echo Chambers, Gate-keepers, and the Price of Bipartisanship. In
WWW ’18 (WWW ’18) . 913–922.[15] Pedro Henrique Calais Guerra, Wagner Meira Jr, Claire Cardie, and Robert Klein-berg. 2013. A Measure of Polarization on Social Media Networks Based onCommunity Boundaries.. In
ICWSM .[16] Young Mie Kim, Jordan Hsu, David Neiman, Colin Kou, Levi Bankston, Soo YunKim, Richard Heinrich, Robyn Baragwanath, and Garvesh Raskutti. 2018. TheStealth Media? Groups and Targets behind Divisive Issue Campaigns on Facebook.
Political Communication (2018).[17] Aleksandra Korolova. 2011. The science of fake news.
Journal of Privacy andConfidentiality
3, 1 (2011), 27–49.[18] David MJ Lazer, Matthew A Baum, Yochai Benkler, Adam J Berinsky, Kelly MGreenhill, Filippo Menczer, Miriam J Metzger, Brendan Nyhan, Gordon Penny-cook, David Rothschild, et al. 2018. The science of fake news.
Science
ASONAM’18 .[20] Filipe N. Ribeiro, Lucas Henrique, Fabricio Benevenuto, Abhijnan Chakraborty,Juhi Kulshrestha, Mahmoudreza Babaei, and Krishna P. Gummadi. 2018. MediaBias Monitor: Quantifying Biases of Social Media News Outlets at Large-Scale.In
ICWSM’18 . Stanford, USA.[21] Koustuv Saha, Ingmar Weber, Michael L Birnbaum, and Munmun De Choudhury.2017. Characterizing Awareness of Schizophrenia Among Facebook Users byLeveraging Facebook Advertisement Estimates.
J. Med. Internet Res.
19, 5 (2017).[22] Eva Sharma, Koustuv Saha, Sindhu Kiranmai Ernala, Sucheta Ghoshal, and Mun-mun De Choudhury. 2017. Analyzing Ideological Discourse on Social Media: ACase Study of the Abortion Debate. In
CSS ’17 . ACM, 3.[23] Till Speicher, Muhammad Ali, Giridhari Venkatadri, Filipe N. Ribeiro, GeorgeArvanitakis, Fabricio Benevenuto, Krishna P. Gummadi, Patrick Loiseau, andAlan Mislove. 2018. On the Potential for Discrimination in Online TargetedAdvertising. In
FAT*’18 .[24] Giridhari Venkatadri, Athanasios Andreou, Yabing Liu, Alan Mislove, Krishna PGummadi, Patrick Loiseau, and Oana Goga. 2018. Privacy Risks with Facebook’sPII-based Targeting: Auditing a Data Broker’s Advertising Interface. In
IEEES&P19 . 221–239.[25] Soroush Vosoughi, Deb Roy, and Sinan Aral. 2018. The spread of true and falsenews online.
Science