Identifying Instagram Crowdturfers with Machine Learning

Researchers in Italy and Iran claim to have formulated the first machine learning system capable of recognizing the ‘crowdturfing’ activity of human (rather than automated) influencer accounts on the Instagram platform. Crowdturfers are real people who perform ‘profile building’ services to platforms which sell such activity on a wholesale basis.

The new method claims an accuracy score of around 95%, and uses semi-supervised learning in Natural Language Processing (NLP) systems.

The authors claim that to the best of their knowledge, their system represents the first crowdturfing (CT) detector system that can reliably hone in on non-bot accounts that are engaged in fake, paid profile engagement and boosting.

To accomplish this, the authors purchased 1293 crowdturfing profiles from 11 CT platform providers in order to obtain data to train their CT detector. Since Instagram has a number of effective anti-bot measures in place, the researchers note, those seeking to exploit the platform’s enormous user base for commercial purposes have turned to paying genuinely influential Instagrammers to ‘engage strategically’ with ‘client’ accounts, mostly by sharing comments, or through activity related to comments on posts.

Having trained the model, the authors then set it loose to analyze the engagement profiles of 20 ‘mega-influencers’, each with over 1 million followers, concluding that ‘more than 20% of their engagement was artificial’.

The paper is titled Are We All in a Truman Show? Spotting Instagram Crowdturfing through Self-Training, and comes from five researchers across the University of Padova in Italy, and Iran’s Imam Reza University.

Breaching the Instagram TOS

Unlike Twitter, favored by social media researchers due to its commitment to aiding research, Instagram not only provides no API or updated data dumps to help researchers, but prohibits machine-driven browsing in its Terms of Service. Therefore the researchers’ first task was to gain an exemption from their guiding Institutional Review Board, justified by prior works that used a similar approach to investigate ‘underground activities’.

The crowdturfing services were purchased for fresh Instagram accounts created by the researchers for their purposes, all of which were deleted after the experiment, obviating the involvement of ‘legitimate’ users. Neither the influencer accounts studied nor the CT platform services are named.

Another ethical hurdle was that the researchers could not request consent of the influencers being studied, due to the Hawthorne effect (i.e. it might have changed the influencers’ behavior), and this exemption was also granted by the IRB.

Finally, since Instagram allows ‘manual collection’ of data, the researchers compromised on their breach of the TOS by setting their automated scraping tools to ‘human speed’, which necessitated a data-gathering phase of five months.

Humans for Sale

The researchers purchased 100 ‘fake follower’ profiles from each of 11 (unnamed) providers.

The paper states*:

‘All the providers we selected ensure to deliver followers who interact with the target profiles by liking and commenting on their posts to boost their engagement rate.

‘These CT profiles are identified as high quality followers and usually cost more than “base” fake profiles. The reliability of these providers is supported by famous [review] platforms like TrustPilot.’

The average cost of buying an Instagram influencer, the paper notes, is not that high, at approximately $3 for 100 ‘high quality’ followers. The authors note:

‘Most providers deliver the followers within a few hours. They offer a drop protection, which means that the number of followers the customer purchases will either remain stable over time or new followers will be delivered to replenish the lost ones.’

The researchers report that some of their fresh Instagram accounts suffered a loss of 15-20% of CT followers after one month, but that in certain cases they gained more than expected. For the most expensive CT provider (CT-10, in the table above), only three followers were lost after one month.

The paper notes that the followed/following ratio becomes more ‘authentic’ the more you pay to the CT provider, with the second-most expensive provider offering a ratio that’s very close to a standard user’s baseline.

One characteristic of a CT Instagram account is that its profile will rarely be set to ‘private’ (a fact that enabled data to be drawn from the purchased fake followers, since most of the analyses centered on profiles and related comments), though this should not be seen as a reliable ‘signal’ in this regard.

‘People joining these platforms are interested in generating a minimum amount of posts that make them reliable, except few cases (CT-4, CT-10). The low-quality profiles show a very high imbalance in followers and following, and the average number of posts is close to 0, far below the CT profiles.’

Data

The researchers collected data through an implementation of the browser-automating framework Selenium. The resulting dataset includes profile information from 1293 CT and 1307 non-CT users.

This admittedly low sample quantity made it feasible to set Selenium to a credibly human speed over a rational period of time. Additionally, the authors note, the representative/interpretive power of semi-supervised learning techniques accommodates smaller datasets very well. Having experimented, for the purposes of thoroughness, with a fully-supervised model, the researchers conclude:

‘[The] results in the semi-supervised mode do not differ significantly from those in a supervised way. This suggests that CT profiles share very similar [characteristics], and that the algorithm can converge [through a small amount of] labeled data.’

The authors gathered all available data from the source code of the ‘compromised’ users’ profile pages, including details generally obscured when rendered, such as the #videos element.

They then pre-processed the data features by removing those with zero or low variance, and finally converted any categorical or non-numeric data into strictly numeric or Boolean features.

Method and Explorations

Besides, Selenium, technologies used across the experiments include: a version of SpaCy implemented with a transformer-based pipeline; a scikit learn self-training classifier; and the Instaloader framework.

There is no customary ‘results’ section in the new paper, since it deals with an objective (i.e., automated inference of corrupt Instagram accounts) that veers away from the central locus of interest to date (i.e., automated inference of automated bot activity on Instagram), meaning that there is no like-for-like prior work against which to compare it.

The researchers adopted a wide range of methods on the available purchased users, (which they feel comfortable describing as ‘fake’ rather than just ‘non-CT’, since these genuine accounts are conducting non-organic, paid engagement activities), across a range of NLP-related technologies.

Among the facets studied were language analysis (which, in the CT world, nearly always defaults to English, though CT platforms offer geo-located non-English followers too); comment counts (where fake users stick very close to the frequency of real users, for fear of detection); and common words analysis:

The paper notes that the prevalence of the word ‘dokter’ (see image above) in fake accounts seems to relate to a specific internal campaign:

‘“Dokter” [appeared] in 1069 distinct comments. By further investigating the accounts spamming [this] word, we found a small portion of what seems to be a botnet whose objective is to spam “Instagram doctors” accounts. All these doctors’ profiles have a WhatsApp business link that, once clicked, starts a chat with a message to complete.’

As far as the researchers can deduce, this strange artifact may be a remnant of a large botnet that they stumbled across while seeking activities from real Instagram users.

In total the researchers collected 603,007 comments from posts across 248,388 unique Instagram users, of which, the authors estimate, 55,719 were crowdturfing accounts.

The paper notes with interest the dominance of female-themed topics in the gathered data. Having used GPU-PDMM (a technique developed for the obligatorily short posts on Twitter) to extract 12,830 suitable comments from an available corpus of 121,822 comments, the algorithm found that in considering content from 12 males and 8 females, the majority of comments deal with female-related topics.

The researchers conclude:

‘[While] Instagram and the research community focused a lot on detecting bots and automated accounts, we believe more studies should be conducted on CT activities, which negatively impact influencer marketing, the Instagram platform, and most of its users.’

 

* Researchers’ quoted TrustPilot URL omitted.

First published 28th June 2022.