Study Finds That Twitter Is Still Unable To Keep Up With The Influx Of Junk Accounts

Researchers from the University of Iowa developed an artificial intelligence (AI) engine that they claim can identify abusive Twitter applications months before the service itself identifies them.

Twitter has scrambled to rein in the bots and trolls polluting its platform since the world learned of state-sponsored campaigns to spread misinformation on social media and sway the 2016 election. In regards to the larger problem of automated accounts on Twitter designed to spread spam and scams, inflate follower counts, and manipulate trending topics, a new study reveals that the company is not keeping up with the deluge of trash and abuse.

The authors of the paper state that, using a machine-learning approach they developed, they can identify abusive accounts in far greater volumes and much more quickly than Twitter, often flagging the accounts months before Twitter banned them.

Inundating the Zone

Zubair Shafiq, a computer science professor at the University of Iowa, and his graduate student Shehroze Farooqi identified more than 167,000 apps using Twitter’s API to automate bot accounts that spread tens of millions of tweets promoting spam, malware links, and astroturfing campaigns. They write that Twitter waited more than 60 percent of the time for these apps to send more than 100 tweets before identifying them as abusive, whereas the researchers’ detection method flagged the vast majority of malicious apps after only a few tweets. For approximately forty percent of the apps examined, Twitter appeared to take more than a month longer than the study’s method to detect abusive tweeting by an app. They estimate that this delay allows abusive apps to generate tens of millions of tweets per month before being banned.

Shafiq explains, “We demonstrate that many of these malicious apps remain undetected by Twitter’s fraud-detection algorithms, sometimes for months, and cause a great deal of damage before Twitter discovers them and removes them.” In May, the research will be presented at the Web Conference in San Francisco. “They have stated that they are now taking this issue seriously and implementing numerous countermeasures. The conclusion is that these countermeasures had little effect on the applications responsible for millions upon millions of abusive tweets.”

We discovered a method that detects them better than Twitter – ZUBAIR SHAFIQ, UNIVERSITY OF IOWA

The researchers say they have been sharing their findings with Twitter for over a year, but the company has not requested additional information about their methodology or data. When WIRED reached out to Twitter, the company acknowledged the study’s objectives but disputed its findings, arguing that the Iowa researchers lacked a comprehensive understanding of how Twitter combats abusive accounts. “Research-based solely on publicly available information about Twitter accounts and tweets rarely paints an accurate or comprehensive picture of the measures we take to enforce our developer policies,” a spokesperson wrote.

Twitter has, to its credit, taken aggressive measures to prevent the most organized disinformation trolls from exploiting its platform. In a report published last week, the social media company stated that it had banned over 4,000 politically motivated disinformation accounts from Russia, 3,300 from Iran, and over 750 from Venezuela. In a statement to WIRED, Twitter noted that it is also implementing new restrictions on how abusive applications gain access to Twitter’s API. The company claims that it has banned 162,000 abusive applications in the last half of 2018.

However, according to the Iowa researchers, their findings indicate that abusive Twitter applications are still prevalent. At WIRED’s request, Shafiq and Farooqi ran their machine-learning model on tweets from the last two weeks of January 2019 and immediately discovered 325 apps they deemed abusive that Twitter had not yet banned, some with explicitly spammy names such as EarnCash_ and La App de Escortas.

Given the enormous impact of the automated tools, the researchers focused solely on identifying toxic tweets produced by third-party applications. Occasionally, malicious applications controlled accounts that spammers or con artists themselves had created. In other instances, they hacked the accounts of users who had been duped into installing the applications, or who had done so in exchange for incentives such as an increase in fake followers.

Tweet Dreck

Among the 1.5 billion tweets the researchers began with, 457,000 third-party applications were represented; Twitter makes only 1% of all tweets available through a research-focused API. The two then used this information to train their machine-learning model for detecting abusive apps. They recorded which accounts each application posted to, as well as the age of the accounts, the timing of tweets, the number of usernames, hashtags, and links included in the tweets, and the ratio of retweets to original tweets. Importantly, they observed which accounts were eventually banned by Twitter during the 16-month observation period, essentially using these bans to identify abusive accounts.

With the resulting machine-learning-trained model, they were able to identify 93% of the applications that Twitter would eventually ban without examining more than the first seven tweets. “In a sense, we rely on Twitter’s eventual classification of malicious apps. However, we discovered a way to detect them better than Twitter “Shafiq says.

Twitter retorted in its statement that the Iowa researchers’ machine-learning model was flawed, as they were unable to accurately identify which applications Twitter had banned for abusive behavior. Since Twitter does not make this information public, the researchers could only speculate based on the applications from which tweets were removed. This may have been the result of a ban, or it may have been the result of users or applications deleting their tweets.

The factors used to train the model in this research are not strongly correlated with whether or not an application violates our policies, according to a statement sent to WIRED by a Google spokesperson.

In contrast, the Iowa researchers note in their paper that they only marked an application as Twitter-banned if at least 90 percent of its tweets were deleted. Less than 30 percent of tweets are removed from popular, benign apps like Twitter for iPhone and Android, according to their findings. If legitimate app users delete their tweets more frequently, “these would be a small minority, these apps would not be used by many people, and I do not expect their results to be affected by this,” says Gianluca Stringhini, a researcher at Boston University who has worked on previous studies of abusive social media apps. Therefore, I would anticipate that their ground truth is reasonably solid.

The researchers refined their definition of abusive apps by crawling websites that advertised fake followers and downloading 14,000 of the applications they advertised. About 6,300 of these had contributed to the 1.5 billion-tweet sample, so they also served as examples of abusive apps for the training data of the machine-learning model.

The rate of false positives was a disadvantage of the Iowa researchers’ method. They admit that approximately 6% of the applications their detection method identifies as malicious are safe. They contend, however, that the false-positive rate is low enough for Twitter to assign humans to review their algorithm’s results and identify errors. Shafiq says, “I don’t believe more than one person would be required to complete this review.” “If you do not aggressively target these applications, they will compromise a great number of additional accounts and tweets and consume a great number of additional man-hours.”

The researchers concur with Twitter’s assessment that the company is moving in the right direction, tightening the screws on spam accounts and, more importantly, abusive applications. In June of 2017, they observed that the company appeared to be more aggressively banning malicious apps. However, according to their findings, Twitter is not utilizing machine learning’s potential to detect app abuse as quickly as it could. Shafiq says, “They’re probably doing some of this right now.” “But insufficient.”

Why Trust Us?

Best Top Reviews Online was established in 2018 to provide our readers with detailed, truthful, and impartial advice on what to buy. We now have millions of monthly users from all over the world and annually evaluate over a thousand products.

The above article was written by the BestTopReviewsOnline team, which consists of some of the most knowledgeable technical experts in the United States. Our team consists of highly regarded writers with vast experience in smartphones, computer components, technology apps, security, and photography, among other fields.

Related Stories

  • All Post
  • Best Picks
  • Explainers
  • How To
  • News
  • Versus
APT Lazarus Aims Mac Malware at Engineers

August 17, 2022

The North Korean APT is conducting a cyberespionage campaign against users of Apple and Intel-based systems using a bogus Coinbase job posting. The North Korean APT Lazarus is up to its old tricks with a cyberespionage campaign aimed at engineers…

Thousands Of Citrix Servers May Be Vulnerable To Attack

December 30, 2022

Many servers remain unpatched, researchers are warning. Numerous Citrix ADC and Gateway servers continue to be susceptible to critical vulnerabilities that were reportedly patched by the company weeks ago, according to experts. Citrix discovered and patched an “Unauthorized access to…

Get more info



Best Products

Buying Guides

Contact Us

About Us

We provide a platform for our customers to rate and review services and products, as well as the stores that sell them. We research and compare the most popular brands and models before narrowing it down to the top ten, providing you with the most comprehensive and reliable buying advice to help you make your decision.

Disclaimer is a participant in the Amazon Services LLC Associates Program, an affiliate advertising program designed to provide a means for sites to earn advertising fees by advertising and linking to As an Amazon Associate I earn from qualifying purchases.


Address & Map

20 S Santa Cruz Ave, Suite 300, Los Gatos, CA 95030, United States

© 2022 Pty. Ltd. All Rights Reserved. Licensing: All third-party trademarks, images, and copyrights used on this page are for comparative advertising, criticism, or review. As this is a public forum where users can express their opinions on specific products and businesses, the opinions expressed do not reflect those of