Arming the public with artificial intelligence to counter social bots

“Botometer is based on Random Forest classifiers (Breiman, 2001).”

“…sequence-to-sequence models for automatic conversations (Li et al., 2015, 2016), deep reinforcement learning for controlling account behaviors (Mnih et al., 2015; He et al., 2016; Serban et al., 2017), and generative adversarial networks for creating fake profile images and content (Goodfellow…”

“In the past four years, Botometer has constantly increased its basin of adoption, and has provided data for a number of influential studies, including by Vosoughi et al. (2018) and the Pew Research Center (Wo-jcik et al., 2018).”

“Deep-learning approaches can also be used to detect sentiment online (Socher et al., 2013; Tang et al., 2014; Severyn and Moschitti, 2015), allowing the next generation of social bots to be more perceptive about human emotions.”

“Some of these methodologies have already demonstrated a capability to surpass human performance on various tasks, such as face recognition and game playing (Mnih et al., 2015; Taigman et al., 2014).”

“Media literacy efforts may be critical in this respect (Lazer et al., 2018).”

“Machine learning is playing a growing role in decision making on issues of broad societal importance (Munoz et al., 2016; Barocas and Selbst, 2016).”

“Algorithmic interpretability would make it easier to spot such biases, and in fact a recent European Union regulation grants users the right to ask for an explanation of an algorithmic decision that was made about them (Goodman and Flaxman, 2016).”

“Recent developments of deep-learning models create opportunities to employ sequence-to-sequence models for automatic conversations (Li et al., 2015, 2016), deep reinforcement learning for controlling account behaviors (Mnih et al., 2015; He et al., 2016; Serban et al., 2017), and generative…”

“Another research challenge that could be relevant for the dismantling of coordinated bot attacks is the detection of promoted campaigns at their early stages (Ferrara et al., 2016b; Varol et al., 2017b).”

“To calibrate Botometer, we employed Platt’s scaling (Niculescu-Mizil and Caruana, 2005), a logistic regression model trained on classifier outputs.”

“Deep-learning approaches can also be used to detect sentiment online (Socher et al., 2013; Tang et al., 2014; Severyn and Moschitti, 2015), allowing the next generation of social bots to be more perceptive about human emotions.”

“Unlike other social engineering attacks, such as spear phishing (Jagatic et al., 2007), bots can achieve scalability through automation.”

“Health is one domain of particular concern (Allem and Ferrara, 2018), where we have observed social bots influencing debates about vaccination policies (Ferrara et al., 2016a; Broniatowski et al., 2018) and smoking (Allem and Ferrara, 2016; Allem et al.”

“A study based on 14 million tweets posted during and after the 2016 U.S. presidential election revealed that bots played a key role in the spread of low-credibility content (Shao et al., 2018a,b).”

“Our analysis (Bessi and Ferrara, 2016) suggests that retweets were a vulnerable mode of information spreading: there was no significant difference in the amount of retweets that humans generated by resharing content produced by other humans or by bots.”

“A recent report 2 revealed that reality television stars, professional athletes, comedians, TED speakers, pastors, and fashion models all purchased fake followers on Twitter at least once.”

“Botometer is based on a supervised machine learning approach (Davis et al., 2016; Varol et al., 2017a).”

“Recent developments of deep-learning models create opportunities to employ sequence-to-sequence models for automatic conversations (Li et al., 2015, 2016), deep reinforcement learning for controlling account behaviors (Mnih et al., 2015; He et al., 2016; Serban et al., 2017), and generative…”

“Deep-learning approaches can also be used to detect sentiment online (Socher et al., 2013; Tang et al., 2014; Severyn and Moschitti, 2015), allowing the next generation of social bots to be more perceptive about human emotions.”

“The spread of fake news online is another area in which the effect of bots is believed to be relevant (Lazer et al., 2018; Bessi et al., 2015).”

“The figure also shows reliability diagrams for raw and calibrated scores (DeGroot and Fienberg, 1983).”

“During the 2010 U.S. midterm elections, primitive social bots were found to support some candidates and attack their opponents (Metaxas and Mustafaraj, 2012), injecting thousands of tweets pointing to websites with fake news (Ratkiewicz et al., 2011).”

“Our analysis highlighted the presence of Twitter bots during the 2017 French Presidential Election (Ferrara, 2017).”

“Recently, Kudugunta and Ferrara (2018) proposed a deep neural network approach based on the contextual long short-term memory (LSTM) architecture that exploits content and metadata to detect bots at the tweet level.”

“Similar cases of political manipulation were reported in other countries (Stella et al., 2018).”

“will at essentially no cost, human detectors cannot. Later eorts therefore mostly leveraged machine learning methods. Approaches based on supervised machine learning algorithms are the most common (Subrahmanian et al., 2016). Supervised approaches depend and often start with the collection of an extensive dataset, wit…”

“A more extreme example of these kinds of bots are identity thieves: they copy usernames, profile information, and pictures of other accounts and use them as their own, making only small changes (Cresci et al., 2015).”

“In 2011, Lee et al. (2011) identified thousands of social bots.”

“While supervised methods have proven to be effective in many cases, they do not perform well at detecting coordinated social bots that post human-generated content (Cresci et al., 2017; Grimme et al., 2018; Chen and Subramanian, 2018).”

“Early attempts combined multiple features via Euclidean distance and then applied clustering or network community algorithms (Ahmed and Abu-laish, 2013; Miller et al., 2014).”

“Despite all the nefarious uses of social bots, there are great examples of bots providing valuable services to social media users (Mønsted et al., 2017).”

“which it is classied as human. This threshold approach is useful for scientists counting bots in a collection (Suarez-Serrato et al., 2016; Haustein et al., 2016; Gorwa, 2017; Stella et al., 2018; Shao et al., 2018a), but it is not intuitive for humans evaluating a single account. A calibrated classier C0 is one whose…”

“The API, whose users tend to be more advanced, also provides uncalibrated scores.”

“During the 2010 U.S. midterm elections, primitive social bots were found to support some candidates and attack their opponents (Metaxas and Mustafaraj, 2012), injecting thousands of tweets pointing to websites with fake news (Ratkiewicz et al., 2011).”

“While supervised methods have proven to be effective in many cases, they do not perform well at detecting coordinated social bots that post human-generated content (Cresci et al., 2017; Grimme et al., 2018; Chen and Subramanian, 2018).”

“A recent report 2 revealed that reality television stars, professional athletes, comedians, TED speakers, pastors, and fashion models all purchased fake followers on Twitter at least once.”

“This threshold approach is useful for scientists counting bots in a collection (Su´arez-Serrato et al., 2016; Haustein et al., 2016; Gorwa, 2017; Stella et al., 2018; Shao et al., 2018a), but it is not intuitive for humans evaluating a single account.”

“In particular, over the past 10 years we have witnessed an explosion of social bots (Lee et al., 2011; Boshmaf et al., 2013), a presence that doesn’t show signs of decline.”

“Early attempts combined multiple features via Euclidean distance and then applied clustering or network community algorithms (Ahmed and Abulaish, 2013; Miller et al., 2014).”

“…for automatic conversations (Li et al., 2015, 2016), deep reinforcement learning for controlling account behaviors (Mnih et al., 2015; He et al., 2016; Serban et al., 2017), and generative adversarial networks for creating fake profile images and content (Goodfellow et al., 2014; Zhao et al., 2017).”

“Cresci et al. (2016) encoded tweet type or content as strings of symbols, then considered the longest common substring to detect coordination.”

“Abokhodair et al. (2015) dissected a botnet that tried to misdirect the online discussion during the Syrian civil war in 2012.”

“An early attempt by Wang et al. (2012) involved building a crowdsourcing social bot detection platform.”

“duce adequate ts. This is likely due to the shape of the P(S jHuman) curve, with its mode near, but not at, zero. Ultimately we chose a lesser-known family of curves for our t: Bernstein polynomials (Babu et al., 2002). Implementation is straightforward as computing values of a polynomial curve is simple, even one wit…”

“Abokhodair et al. (2015) dissected a botnet that tried to misdirect the online discussion during the Syrian civil war in 2012.”

“After the Boston marathon bombing, for example, bots started to automatically retweet indiscriminately, without verifying the credibility of the posts being retweeted or that of their sources (Gupta et al., 2013), unintentionally amplifying the diffusion of misinformation.”

“Health is one domain of particular concern (Allem and Ferrara, 2018), where we have observed social bots influencing debates about vaccination policies (Ferrara et al., 2016a; Broniatowski et al., 2018) and smoking (Allem and Ferrara, 2016; Allem et al., 2017).”

“Another research challenge that could be relevant for the dismantling of coordinated bot attacks is the detection of promoted campaigns at their early stages (Ferrara et al., 2016b; Varol et al., 2017b).”

“As AI systems employ more sophisticated learning methodologies, model complexity increases and systems become vulnerable to overlooked biases (Kirkpatrick, 2016).”

“They can even interact with other users by engaging in conversations, commenting on posts, and answering questions (Hwang et al., 2012).”

“A recent report 2 revealed that reality television stars, professional athletes, comedians, TED speakers, pastors, and fashion models all purchased fake followers on Twitter at least once.”

“While more traditional nefarious entities, like malware, attack vulnerabilities of hardware and software, social bots exploit human vulnerabilities, such as our tendencies to pay attention to what appears to be popular and to trust social contacts (Jun et al., 2017).”

“Health is one domain of particular concern (Allem and Ferrara, 2018), where we have observed social bots influencing debates about vaccination policies (Ferrara et al., 2016a; Broniatowski et al., 2018) and smoking (Allem and Ferrara, 2016; Allem et al., 2017).”

“Since the support of our distribution is necessarily the unit interval, the use of KDE in this case leads to biased estimates near the interval boundaries (Leblanc, 2012).”

“Some bots, for example, search the social networks for popular accounts, follow them, and ask to be followed back (Aiello et al., 2012).”

“Not surprisingly, hybrid approaches that mix human and automated activity have been devised to bypass bot and bot-coordination detection systems (Grimme et al., 2018).”

“They can generate content automatically and interact with human users, often posing as, or imitating, humans (Ferrara et al., 2016a).”

“Some bots, for example, search the social networks for popular accounts, follow them, and ask to be followed back (Aiello et al., 2012).”

“abels usually come from human annotation (Varol et al., 2017a), automated methods (e.g., those based on honey pots described earlier (Lee et al., 2011)), or botnets that display suspicious behaviors (Echeverria and Zhou, 2017a,b). A critical issue with existing datasets is the lack of ground truth. There is no objecti…”

“This threshold approach is useful for scientists counting bots in a collection (Suárez-Serrato et al., 2016; Haustein et al., 2016; Gorwa, 2017; Stella et al., 2018; Shao et al., 2018a), but it is not intuitive for humans evaluating a single account.”

“Health is one domain of particular concern (Allem and Ferrara, 2018), where we have observed social bots influencing debates about vaccination policies (Ferrara et al., 2016a; Broniatowski et al., 2018) and smoking (Allem and Ferrara, 2016; Allem et al., 2017).”

“Cresci et al. (2019) proposed the use of evolutionary algorithms to improve social bot skills.”

“…models for automatic conversations (Li et al., 2015, 2016), deep reinforcement learning for controlling account behaviors (Mnih et al., 2015; He et al., 2016; Serban et al., 2017), and generative adversarial networks for creating fake profile images and content (Goodfellow et al., 2014; Zhao…”

“While supervised methods have proven to be effective in many cases, they do not perform well at detecting coordinated social bots that post human-generated content (Cresci et al., 2017; Grimme et al., 2018; Chen and Subramanian, 2018).”

“The deceptive strategies employed on all these types of channels share striking similarities (Varol and Uluturk, 2018).”

“…have been considered, but in general six broad categories of features have been identified as relevant for discriminating between human and bot accounts (Varol et al., 2018): user metadata, friend metadata, retweet/mention network structure, content and language, sentiment, and temporal features.”

“Future efforts should aim for a pro-active approach to mitigate the effects of social bots by engaging with them early (Cresci et al., 2018).”

“Our recent analysis points to the existence of many distinct types of social bots, such as spam bots, mention bots, cyborgs, and many others (Mitter et al., 2014).”

“abels usually come from human annotation (Varol et al., 2017a), automated methods (e.g., those based on honey pots described earlier (Lee et al., 2011)), or botnets that display suspicious behaviors (Echeverria and Zhou, 2017a,b). A critical issue with existing datasets is the lack of ground truth. There is no objecti…”

“…suspension of abusive bots that spread false or misleading political content has been criticized by some politicians as censorship (Kang and Frenkel, 2018; Ty-nan, 2018); even research on bot detection has been targeted by partisan attacks and labeled as suppression of free speech (Mervis, 2014).”

“Such coordination cannot be captured by our feature-based approach (Cresci et al. 2017; Yang et al. 2019).”

“By leveraging more than 1,200 features of an account, Botometer evaluates possible bots based on their profile characteristics, social network structure, the content they produce, their sentiment expressions, and the timings of their actions [35].”

“Finally, there is a continued arms race between bot-detection algorithms and bot creators [21,33].”

“Property Semantic Neighbor varol-icwsm [27] pronbots [31] ✓ celebrity [31] ✓ gilani-17 [13] ✓ cresci-rtbust [19] ✓ cresci-stock [7] ✓ midterm-18 [32] ✓ botwiki [32] ✓ verified [32] ✓ PAN-19 3 ✓ caverlee [15] ✓ ✓ cresci-17 [6] ✓ ✓ TwiBot-20 ✓ ✓ ✓ are also encouraged to download and use the full…”

“Even though many new bot detection methods are developed every year—outperforming Botometer in some cases—[15, 16, 18], we focus on Botometer as it is by far the most used tool in studies analyzing bots on Twitter [12], especially in social science [5, 7, 10].”

“WecomparetheESCapproach( Botometer-v4 )withthe Botometer-v3 baseline model [48], as well as the Botometer-v3.”

“Although Twitter strengthened their efforts to contain malicious actors in recent years, 1 deceptive bots remain prevalent and display evolving tactics to evade detection [34].”

“models to identify the inauthentic accounts (Yang et al., 2019).”

“e identification of highly-predictive features that discriminate between bots and humans. Other recent surveys propose complementary or alternative taxonomies that are worth considering as well [36], [50], [51] Some openly accessible tools exist to detect bots on platforms like Twitter: (1) Botometer3, also used here,…”

“Second, despite Botometer offers a simple, effective way to detect Twitter bots by relying on individual account features, recent studies highlighted that social bots acting in coordination are harder to identify when inspected individually [33], [62].”

“Most machine learning tools to combat online abuse target the detection of social bots, and mainly use methods that focus on individual accounts (Davis et al. 2016; Varol et al. 2017; Yang et al. 2019; 2020; Sayyadiharikandeh et al. 2020).”

“We used Indiana University’s Botometer, a tool that assigns a bot-score to a Twitter account based on an account’s activity [14, 15].”

“BotOrNot) [8, 40], consisting of both a public Web site (https://botometer.”

“We used the maximum of Botometer’s English and universal scores, both provided in the [0 , 1] range, as our automation score.”

“political-bots (Yang et al. 2019) Data based on politics-oriented bots shared by a Twitter user 61 0”

“Based on machine learning algorithms, Botometer extracts more than 1000 predictive features that capture many suspicious behaviors, mainly through characterizing the account’s profile, friends, social network, temporal activity patterns, language, and sentiments [78,79].”

“The developers, computer scientists at the University of Indiana, de fi ne social bots as ‘ automated or semi-automated accounts designed to impersonate humans ’ (Yang et al., 2019: 48).”

“Social & humanism character shifts to virtual global (Yang, 2019).”

“Inauthentic, coordinated, malicious accounts, whether human-controlled or automated (Ferrara et al., 2016; Shao et al., 2018; Varol et al., 2017; Yang et al., 2019), domestic or state-sponsored (Zannettou et al., 2019), or anywhere in between, have been weaponized to influence public conversations…”

“Researchers from Indiana University has developed several tools to detect misinformation spread by social bots [66, 107].”

“…2013), supervised models of automated behavior (Ferrara, Varol, Davis, Menczer, & Flammini, 2016; Varol, Ferrara, Davis, Menczer, & Flammini, 2017; Yang et al., 2019; Yang, Varol, Hui, & Menczer, 2020; Hui et al., 2019), time series analysis to detect promoted campaigns (Varol, Ferrara, Menczer,…”

“There has been work on exploring network structure, looking for malicious accounts and coordinated inauthentic behavior (Cresci et al., 2017; Yang et al., 2019; Chetan et al., 2019; Pacheco et al., 2020).”

“More recently, several scientific studies have incorporated more annotated samples to support this research such as [27], [44], [45] including some procedures for achieving better level of accuracy by strategically selecting a subset of training samples that better generalize the problem.”

“Specifically, we utilize four datasets: botometer-feedback (Yang et al., 2019), gilani-17 (Gilani et al., 2017), midterm-2018 (Yang et al., 2020), and varol-icwsm (Varol et al., 2017), randomly selecting 285 human accounts from each.”

“The research community offered multiple approaches for identifying automated accounts (Davis et al. 2016; Varol et al. 2017a; Yang et al. 2019; Chavoshi, Hamooni, and Mueen 2016; Subrahmanian et al. 2016; Chen and Subramanian 2018; Kudugunta and Ferrara 2018).”

“• Vendor-19 [55] include a collection of fake followers deriving from several companies.”

“Botometer has been successfully used to evaluate the presence and activity of bots in other discourses (Varol et al., 2017; Yang et al., 2019), but our use here represents a first attempt to understand the presence of bots in the online climate change discourse.”

“Whereas the abundance and multi-dimensionality of crowd-sourced big data lend itself to machine learning techniques in the spatial and temporal domain, machine learning has also been used to assess and improve the data quality of VGI, such as OpenStreetMap [61], and to remove automated bots in social media data, such…”

“When used effectively, ML powerful algorithms create a learning environment for systems, accomplishing tasks such as spotting known/unknown malicious attacks [2,3,6].”

“The Social bot detection problem has received attention from several researchers in various communities such as artificial intelligence [12], [13] and cybersecurity [14].”

“While social media enables the fast propagation of fake news or any other misleading information over the Web, the so-called political social bots take care of amplifying their popularity to catch the eye of virtual communities and to create manually crafted viral trends [7].”

“The only distinction worth mentioning is that we used the Botometer API version v3 that brings new features and a non-linear re-calibration of the model: in line with the associated study’s recommendations [86], we used a threshold of 0 .”

“active accounts with similar tweets, likely to be automated according to Botometer [23].”

“This assumption is used to construct what can be called, depending on the application, a social Yu et al. (2008); Yu et al.”

“To address the concern that the Twitter discourse may be driven by bots, we classified a random sample of 100,000 user profiles in our dataset using the Bot-O-Meter API v414 (Yang et al., 2019).”

“Social media platforms are being manipulated through malicious automated algorithms to sway public opinion for political gains (Ferrara et al., 2016; Yang et al., 2019).”

“Due to the rate limit of the API, we only crawled cascades that have less than 2 , 500 events.”

“Second, because diversity is a property of the audience and not of its level of engagement, it is less susceptible to manipulation if one can detect inauthentic partisan accounts [44, 49, 52, 53].”

“Social bots are a major tool for disseminating misinformation or disrupting online conversations [19, 54].”

“One of the challenges faced in evaluating bot detection approaches is that the ground-truth data is insufficient [17].”

“Many tools and techniques have been developed to detect bot and bot-generated information [37, 90, 166].”

“One of the most popular bot-detection APIs is Botom-eter 11 (Davis et al. 2016; Yang et al. 2019), which provides a robust method to detect the existence of bots in social media.”

“In 2019, authors released the third and currently most up-to-date version of Botometer, featuring more training examples and better detection performance than its predecessors [64].”

“Recently, Yang et al. (2019) performed an extensive review of various kinds of bots and their detection techniques.”

“These artificial agents can be used for spamming, cyberbullying, terrorist activities, gaining fame, spreading wrong information, restricting freedom of speech [1] .”

“Social bots might introduce themselves as individual accounts with locations and photos for their profiles [64,66].”

“[14]reviewed different techniques that use artificial intelligence to combat bots in social media.”

“The Twitter streaming API has been a valuable resource for research on tweets themselves[11, 12, 13], but hides another resource: user profiles, containing self-descriptions, held in the metadata of each tweet.”

“(b) Vendor-19 (� = 0.5).”

“, paid-for or coerced crowds), online legitimacy has proven strikingly easy to manipulate (Yang et al., 2019).”

“The need for delivering human understandable justification in terms of online social bots scoring is along with the earlier research work [50],[51] but also with the AI research responsibility directives [33],[39].”

“This classification is based on various features of the user account profile, online social network structures, historical patterns of activity, and language and sentiments (Yang et al., 2019; Botometer tool, 2019).”

“In line with recent literature (Ferrara, 2020; Yang et al., 2019), we used the top 10 and bottom 10 percentiles of users to set apart likely human users from automated accounts, with Botscores of 0.06 and below for humans and 0.67 and above for bots.”