Social media platforms face an ongoing challenge in combating the proliferation of social bots, automated accounts that are also known to distort public opinion and support the spread of disinformation. Over the years, social bots have evolved greatly, often becoming indistinguishable from real users, and more recently, families of bots have been identified that are powered by Large Language Models to produce content for posting. We suggest an idea to classify social users as bots or not using genetic similarity algorithms. These algorithms provide an adaptive method for analyzing user behavior, allowing for the continuous evolution of detection criteria in response to the ever-changing tactics of social bots. Our proposal involves an initial clustering of social users into distinct macro species based on the similarities of their timelines. Macro species are then classified as either bot or genuine based on genetic characteristics. The preliminary idea we present, once fully developed, will allow existing detection applications based on timeline equality alone to be extended to detect bots. By incorporating new metrics, our approach will systematically classify non-trivial accounts into appropriate categories, efectively peeling back layers to reveal non-obvious species.
The digital age has brought with it an unprecedented proliferation of accounts on social platforms,
resulting in a diverse and complex ecosystem. Of particular note are automated accounts, commonly
known as social bots [
One of the most important strands of research began with the realization that bots, programmed to
pursue specific goals, often operate in a coordinated manner and exhibit similar behaviors. In particular,
one modeling and detection technology that was notably relevant was that based on digital DNA
[
Based on the intuition that accounts of the same type behave similarly, if not exactly the same, we propose a method for classifying bot accounts that may be mistaken for real accounts.
CEUR
ceur-ws.org
The classification approach we propose has been designed for Twitter/X users, but minor adjustments can easily adapt it for use on other social media platforms. Figure 1 shows the scheme of the procedure. 30/08/2024, 18:43 poster_version.drawio.svg D N A E n c o d e r
gSpamBot gGenuine species species
species gSpamBot gGenuine Species Species Species species Pareto
gSpamBot gGenuine species species Species left species species species species Species left species species species species
Digital DNA The first phase involves previ-apLrCoScessing usersSbpeycieenscodinSgpetchieeisr online behavior using Digital DNA. Digital DNA aims to compactly represent the behavior of a social account, using a sequence of characters from an alphabet , such as the following madeSpoefcitehsreeSpcehcaiersacters: Species → plain tweet t3ype = { → retweet }
→ reply
Clustering u~:sGeenertic sSimiilanrity to species After the pre~-: pGerneotic Scimeilasritysing stage, users are grou~p: GeendeticiSinmilatrioty macro species,
: weightedLCS
based on the concept of the LongSgpeamsBtoCtommon Substring (LCS)g[S1p1am]B.oTthe LCS of two or more stgrSipnamgBsot
is the longest string that is ~a substring of all of them. Figu~re 2 shows an example = o3.f25an LC~ S curve,
where the abscissasepecaierse theS = g0.1roup~s Mo=f0.2k5 users, andspetchiees ordSin=1a.0 te is~ tMh =e0.25length of thspeeciLesCS foS r=0.3each~oMf= 0.25
the user groups. Considering a dataset of users paired with their digital DNA sequences, the LCS
computation is performed in linegaGerntuiimnee [
To establish the initial gSpamBot and gGenuine groups, the following idea is applied: the LCS of a species represents the users within it; therefore, a group with a long LCS indicates similar social behavior, possibly an indicator of social bots. On the other side, a group with a short LCS implies a diversity in the social behavior of its members. The construction of the initial gSpamBot group is inspired by the Pareto principle, which seeks to determine the subset of individuals that have the most significant impact on the overall community. Based on the original dataset, users have now been categorized into three groups: gSpamBot, which comprises a considerable number of users exhibiting very similar behavior, strongly indicating that they are social bots; gGenuine, formed by merging species that demonstrate human-like actions; and lastly, species that do not fall into either gSpamBot or gGenuine and are therefore unlabeled at this stage.
Classification of species using genetic similarity To classify the unlabeled species (those colored gray in Figure 1), whose users are not immediately classifiable as bots or not, we propose an algorithm that uses custom genetic similarity metrics. The idea is using a sequence alignment algorithm —a wellestablished technique in bioinformatics— to compare the LCS of each unlabeled species with those of the two primary groups. After this alignment, our algorithm introduces a structured classification process. The process involves two key steps: first, calculating a similarity score based on the alignment of LCS sequences between species; and second, evaluating a new metric that considers the relative similarity of DNA sequences within a species and the size of the population contributing to that similarity. By integrating these procedures, we believe that our approach can efectively label the previously unlabeled species as either gSpamBot or gGenuine, thereby completing the account classification process.
In this short paper, we presented the idea of a new approach to social bot detection that we hope will not only achieve efective classification, but also maintain a transparent decision-making process.
We plan to implement and test our proposed classifier using well-established bot repositories (see the ones published on the site of the OSOME research unit at Indiana University) as well as recently discovered datasets where social bots use Large Language Models (LLMs) to write their posts.
This work is partially supported by project SERICS (PE00000014) under the NRRP MUR program funded by the EU - NGEU; by project re-DESIRE (DissEmination of ScIentific REsults 2.0), funded by IIT-CNR; by project ‘Prebunking: predicting and mitigating coordinated inauthentic behaviors in social media’, funded by Sapienza University of Rome.