Article Preview
Top1 Introduction
Instagram is one of the most popular social media platforms globally, with over one billion monthly active users (Tankovska, 2021b). More and more public services are using it, and companies are using it to promote and sell their products. However, in OSNs, everything can be bought, even likes and subscribers. The race for subscribers can lead some influencers to cheat, attracted by the lure of financial gain and the desire to work with marks. Unfortunately, this breaks the trust (Castellini et al., 2017). Many people try to gain fame by exploiting this mechanism, by buying fake profiles on the online black market (Aggarwal & Kumaraguru, 2015). likes.io1 is another example of a company that offers to buy fake followers. The customer can buy 100 followers generated by bots for only 4$. As of January 2021, footballer Cristiano Ronaldo tops the list of the most popular Instagram accounts. With nearly 258.7 million followers, he is the most followed user on Instagram (Tankovska, 2021a).
This paper's solution for fake detection in online social networks is based on supervised machine learning techniques used for various artificial intelligence applications, such as smart cities (Koumetio Tekouabou et al., 2020; Mohammadi & Al-Fuqaha, 2018), finance (Bahrammirzaee, 2010; Ghoddusi et al., 2019), marketing (Li et al., 2018; Wirth, 2018), and computer vision (Khan & Al-Habsi, 2020; Nyalala et al., 2019).
It is better to have a precise model, but it is also better to get an explainable model, especially for making efficient and transparent decisions (Shirataki & Yamaguchi, 2017). The need for the explicability of an algorithm can be seen through the following example: Let's imagine a situation in which the fake account detection system suspends a user. Although it is based on a ML algorithm, if the algorithm is very complex and the decision making is difficult to explain, security advisors would be unable to justify such a decision. This example presents one of the possible scenarios in which misunderstanding how the algorithm works is a real challenge in the field of artificial intelligence.
1.1 Research Questions
This article aims to study how to detect fake profiles on online social networks using machine learning classification (Igual & Seguí, 2017; Kubat, 2017), also giving the background reason for the decision taken by the predictive system. We ask the following research questions that we try to address in this paper:
Question 1: How to reference a legitimate account from a fake account?
Question 2: Can we justify the decision of a prediction system for more transparency?
Question 3: What are the characteristics of a fake profile to easily identify it on Instagram?
Question 4: Is it possible to improve existing labeled datasets to generate more examples of real and fake accounts without the need for data collection and annotation steps?
1.2 Proposed Solution
Our objective in this paper is to identify an approach to distinguish between legitimate and fake profiles on instagram based solely on publicly available profile data. This article uses different ml algorithms such as ensemble methods, SVM, and decision tree. In parallel to these algorithms, we used the min-max normalization technique to improve accuracy and also balance the data by the smote algorithm. In addition, the majority of these decision support systems are sophisticated black boxes, which signifies that the reasoning and internal mechanisms are masked from the individual, and even experts cannot completely comprehend its justification of their predictions. Consequently, trusting in machine learning systems require interpretability (carvalho et al., 2019; kaur et al., 2020). Our approach has achieved high accuracy for random forest and xgboost.