Artificial intelligence tools in the ongoing fight against bullying and cyberbullying: a multidisciplinary approach

Abstract

The problem of bullying and cyberbullying is growing and is a phenomenon that negatively influences our society: for this reason, it requires advanced solutions aimed at prevention. Starting from a previous project called BullyBuster which paved the way for the application of artificial intelligence (AI) in this area, exploiting interdisciplinary skills to develop algorithms capable of identifying and mitigating bullying behaviours, in this paper, we present the follow-up, called BullyBuster 2 (BB2). BB2 aims to broaden the spectrum of intervention to include adult populations, reflecting the universal nature of bullying across all age demographics. Furthermore, it updates and improves the original project's methodologies with new psychological insights and broader AI applications, ensuring a more inclusive anti-bullying strategy. Furthermore, BB2 aims to form a large research consortium to set a new standard for interdisciplinary collaborations in preventive and response strategies against such social challenges.

Keywords

• eol>bullying
• detection
• cyberbullying
• artificial intelligence
• interdisciplinarity

1. Introduction

bullying in both physical and virtual contexts and for this reason it has been included in the Global Top 100 list The worrying rise in bullying and cyberbullying in re- of AI projects addressing the 17 United Nations Stratecent years has brought attention to the urgent need for gic Development Goals by the International Research eficient prevention measures and solutions [ 1]. In fact, Center for Artificial Intelligence under the auspices of such violent behaviours not only compromise individual UNESCO1. feelings of security and dignity, but they also have signif- This paper introduces “BullyBuster 2 – the ongoicant psychological efects that may reverberate across ing fight against bullying and cyberbullying with the society. Bullying has changed throughout time and is help of artificial intelligence for the human wellbeing” now prevalent in a variety of adult demographics and project (BB2, hereinafter), the follow-up to the Bullysettings, which has made it clear that creative and flexible Buster project. Funded by the European Union - NextGensolutions are required. erationEU and within the PRIN 2022 PNRR, it again in

In this context, the “BullyBuster - A framework for volves four multidisciplinary research groups belonging bullying and cyberbullying action detection by computer to four universities in Southern Italy (University of Bari vision and artificial intelligence methods and algorithms” Aldo Moro, University of Cagliari, University of Foggia, project [2](BB1 in the following) was proposed: funded University of Naples Federico II) with the aim of combinunder the tender relating to Projects of Relevant National ing artificial intelligence, technology, law and psychology Interest (PRIN) in 2017 and recently concluded, it marked skills to develop a comprehensive framework. a pioneering efort by interdisciplinary teams from four BB2’s goal is to advance bullying and cyberbullying universities in Southern Italy to tackle this threat through detection capabilities by incorporating advanced models cutting-edge artificial intelligence tools [ 3 ]. This initia- that take into account a broader range of victim profiles tive sought to combine psychological insights with tech- and refining its detection algorithms to accommodate nological advances to create tools capable of identifying new forms of interpersonal aggression and IT misconduct. BB2 also aims to broaden the objectives of the previous

1https://ircai.org/top100/entry/bullybuster-a-framework-forbullying-and-cyberbullying-action-detection-by-computer-visionand-artificial-intelligence-methods-and-algorithms/

2.1. Legal, juridical and psychological aspects

2.2. Multimedia contents artifacts

2. The BullyBuster framework

In this section, we explore the key elements of Bully

Buster’s architectural framework (Fig. 2), which includes the tools created in BB1 as well as the enhancements planned for BB2. The architecture of the system is structured into five primary modules: (1) a manipulated multimedia content detection module, to combat the spread of malicious and fake multimedia content, as deepfakes; (2) a physical violence detection module, to monitor and identify potential bullying incidents in the physical domain; (3) a stress detection module, able to assesses the emotional state of individuals by analyzing the timing and rhythm of keystrokes and other behavioral biometrics; (4) a verbal abuse detection module, focused on the analysis of textual communications to detect verbal signs of bullying or cyberbullying; (5) a internet addiction detection module, to detect psychological malaise by automatic behavioral analyses. When combined, these components provide a complete system that is suited to address the complex problem of bullying in both physical and digital scenarios. In order to address the critical issues of privacy and data protection, this solutions have been examined from a legal and juridical perspective. Below, we describe the individual modules of the BullyBuster framework.

The term "deepfake" refers to any way of creating fake multimedia material representing one or more individuals [ 6 ]. The modification may specifically target the individual’s identity, speech, or expression. Although this technology is of considerable appeal to some sectors of society, such as the film and video game industries, the harmful applications are vast and particularly concerning. In fact, deepfakes may position a person in non-real humiliating or compromising situations, such as scenes of sex or violence, or force her/him to say something she/he never said. Deepfake can thus be used to insult, stalk, propagate revenge pornography, and abuse or cyberbully someone. In BB1, we exploited the complementarity of several individual video deepfake detectors with appropriate fusion rules to increase the generalization ability of modern deepfake detection systems [ 7 ]. The goal of BB2 is to extend the deepfake video detection capabilities and add voice cloning detection to make multimedia content analysis complete.

2.3. Anomaly detection of events

The battle against bullying may be greatly aided by hav

ing a visual understanding of human behaviour through video analysis. Behavioural "indicators," whether in the form of victimisation or bullying, might include physical aggression, seclusion, or other physical patterns like encircling. In BB1, systems have been created that can use one or more surveillance cameras to monitor groups [ 10 ]. The goal was to develop a system that automatiof people who are not individually identifiable [ 8 ]. The cally identified aggressive behaviour in texts, especially data is then processed to report "anomalous" events like in Italian-language comments. To do this, the language violent or panicked episodes based on behavioural mod- patterns used in aggressive comments were analyzed, els that have been suitably codified. In BB2, the crowd noting the frequent use of vulgar and negative language, analytics model will specialize in bullying contexts by insults, and ofensive words. In addition, many of these generating synthetic data: obtaining contextualised train- comments began with "no/not," which is used to contraing data proved to be a challenge during the creation of dict or deny a statement. Based on these observations, the crowd analyzer in BB1, made worse by the start of several features were developed for feature engineering, the COVID-19 epidemic, which made it impossible to including the number of negative words, the number of mimic actual bullying scenarios. In addition, a substan- "no/not," use of capitalisation, positive/negative comment tial amount of video is needed for deep-learning models. weight calculated with WordNet and SentiWordNet, use In BB2, we aim to generate a suficient amount of data of the second person, presence of threats and bullying and to include the typical behaviours of bullies and vic- terms [ 10 ]. These features were used to create several tims, highlighted in the psychological analysis, through dictionaries of terms and a dataset of aggressive and nonthe synthetic creation of this data. aggressive comments. Several shallow learning models were tested to evaluate the efectiveness of these features 2.4. Verbale abuse models in detecting verbal aggression [ 10 ]. In the context of the BB2 project, this phenomenon Cyberbullying is a widespread phenomenon involving can also be analysed by considering diferent age groups, the use of digital technologies to perpetrate aggressive especially the preponderance of adults. Adults may rebehavior, threats, or harassment toward other individu- ceive significant criticism, fueling a cycle of online negals. This bullying occurs primarily through online plat- ativity and conflict that can also be reflected in real life. forms, such as social media, instant messaging, and fo- BB2 aims to address this problem by implementing and rums. It can take many forms, including name-calling, improving existing text models, with particular attendefamation, social exclusion, and disclosure of private tion to the users’ ages. While BB1 primarily addressed information. Cyberbullying can have serious emotional, adolescents, BB2 includes an analysis of adult language, psychological, and social consequences on victims and recognizing that verbal aggression and communication poses a significant challenge for parents, educators, and dynamics are not limited exclusively to the younger genmental health professionals in ensuring a safe and re- eration. BB2’s approach is based on analyzing textual spectful online environment for all users. One of the dynamics found in social media comments, examining most predominant vectors is textual language [ 9 ]. how diferent users interact and communicate. This ap

In the BB1 project, UNIBA focused on identifying ver- proach allows for the identification of common linguistic bal aggression, a crucial aspect of cyberbullying that pri- and behavioral patterns. The categories used are simimarily involves texts and comments on online platforms larly taken by assigning the label of cyberaggression or non-cyberaggression. Through this in-depth analysis, BB2 aims to provide users, social platforms, and mental health professionals with a better understanding of how verbal aggression occurs online and the factors that inlfuence it. This can enable the development of targeted interventions to prevent and address cyberbullying, improve online safety, and promote a healthier and more respectful virtual environment for all users.

The text also fits within the framework of the Behavioral Biometrics approach, in which touch-related features, called Touch Dynamics, are extracted along with textual output. This could be a starting point for future work [ 11 ].

2.5. Stress detection 2.6. Digital devices addiction

Examining data from mobile devices has emerged as a prominent approach for understanding individuals’ behavior in today’s digital era [17]. However, the abundance of features on these devices, such as web browsing, online gaming, digital photography, GPS navigation, and various social applications, can easily engage users’ attention, potentially causing significant distractions from real-world activities. Unfortunately, excessive engagement with these features can lead to various health issues, particularly psychological challenges, such as lack of self-control, withdrawal symptoms, sleep disruptions, social isolation, depression, and dificulties in maintaining focus. Additionally, users may display symptoms of irritability, restlessness, stress, and mood fluctuations. In response to these concerns, the goal of BB2 is to introduce an innovative module aimed at assessing individuals from a psychological standpoint, based on their interactions with diferent features of their mobile devices.

Biometrics involves the utilization of body measurements

and statistical analysis to extract and quantify human characteristics [ 12 ]. Initially utilized for user authentication and identification purposes [ 13], , this technology is now increasingly utilized across various domains, includ- 3. The interdisciplinary laboratory ing entertainment and personalized user experiences [14].

Within the field of biometrics, there exist two distinct cat- Based on the results of the completed BB1 project and the egories of approaches: i) Physiological biometrics, which ongoing BB2 project, an interlaboratory group named involves the direct measurement of physical attributes "BullyBuster" is being formed with the goal of attracting such as facial features, fingerprints, iris patterns, reti- membership and engagement from individuals who are nal structures, and vocal characteristics; ii) Behavioral particularly interested in solving the issue of bullying and biometrics, which focuses on capturing specific human cyberbullying. The interdisciplinary laboratory brings tobehaviors such as handwriting, typing, or speaking. gether specialists from various fields to develop strategies

Among various forms of behavioral biometrics, for analyzing bullying and cyberbullying behaviours, exkeystroke dynamics has emerged as a highly eficient tending beyond just school-aged individuals. This team and economical technique, easily deployable using read- includes: (1) experts in information security, computer ily available hardware. In recent years, it has seen in- vision, and artificial intelligence, to create and propose creased utilization for user authentication, analyzing the innovative models and methods to be implemented in habitual rhythm patterns exhibited while typing on both appropriate demonstration products; (2) psychologists, physical and virtual keyboards [15]. to understand the behaviour patterns of individual in

In the context of combating cyberbullying, we employ volved in bullying events; (3) sociologists, to contribute keystroke dynamics for user emotion recognition. The by examining the efects of bullying and aggressive beaim is to exploit its potential as a cost-efective and widely haviours in social media and other environments; (4) juaccessible method for emotion recognition, requiring ridical and legal professionals, to explore new legislation only a standard keyboard as hardware [ 16 ]. Furthermore, that can address (cyber)bullying, enabling both preventasince a keystroke recorder can be implemented as hard- tive and punitive technological solutions; (5) economists ware or software, with the latter being inconspicuous, and healthcare professionals, to assess the financial and individuals using the keyboard must be made aware of health impacts of (cyber)bullying, whether stemming the monitoring process. In BB2, keystroke dynamics will from the attacks themselves or from the origins of such be examined in real-world settings, considering both key- behaviours. The laboratory aims to create a genuine supboard inputs and tapping habits to characterize a subject port system and reference services for preventing and based on their mobile device usage behavior. Addition- fighting bullying. ally, this analysis can incorporate touch dynamics on mobile devices.

4. Conclusions

Bullying and cyberbullying are significant societal issues that have an adverse impact on people and communities.

Acknowledgments

This work is supported by the European Union – Next

Generation EU under the PRIN 2022 PNRR project “BullyBuster 2 – the ongoing fight against bullying and cyberbullying with the help of artificial intelligence for the human wellbeing” (CUP: P2022K39K8).

The BullyBuster (BB) framework, developed during two

projects funded under the tender relating to Projects of Relevant National Interest (PRIN), one completed and the other ongoing, combines advanced computer vision, artificial intelligence technology, legal aspects and psychology concepts with a multidisciplinary approach to provide a holistic response to these prevalent problems. In this paper, we present the development objectives of the BullyBuster framework, which include the improvement of tools for the detection of bullying and cyberbullying, the extension of its application to include a target of adults and the creation of an interdisciplinary reference laboratory. tional states through keystroke dynamics, in: Proceedings of the 3rd International Conference on Deep Learning Theory and Applications - Volume 1: DeLTA„ INSTICC, SciTePress, 2022, pp. 207–214. doi:10.5220/0011367300003277. [17] F. S. Rahayu, L. E. Nugroho, R. Ferdiana, D. B. Setyohadi, Research trend on the use of it in digital addiction: An investigation using a systematic literature review, Future Internet 12 (2020) 174.

References

URL: https://www.mdpi.com/2078-2489/14/8/430.

doi:10 .3390/info14080430.

[3] G. Orrù , A. Galli , V. Gattulli , M. Gravina , S. Mar-

WORKSHOP PROCEEDINGS , volume 3486 , CEUR-

WS Team , Redaktion Sun SITE , 2023 , pp. 189 - 194 .

[4] M. Wiertsema , C. Vrijen , R. van der Ploeg, M. Sentse,

of Adolescence 95 ( 2023 ) 34 - 55 . doi: 10 .1002/jad.

[5] M. O. Mantovani , Profili penali del cyberbullismo:

la l. 71 del 2017 , Indice penale ( 2018 ) 475 .

[6] M. S. Rana , M. N. Nobi , B. Murali , A. H. Sung , Deep-

access 10 ( 2022 ) 25494 - 25513 .

[7] S. Concas , S. M. La Cava , G. Orrù , C. Cuccu , J. Gao ,

Sciences 12 ( 2022 ). doi: 10 .3390/app12157365.

[8] G. Orrù , D. Ghiani , M. Pintor , G. L. Marcialis , F. Roli ,

descriptor, in: 2020 25th International Conference

on Pattern Recognition (ICPR) , IEEE, 2021 , pp. 4642 -

[9] S. Joksimovic , R. S. Baker , J. Ocumpaugh , J. M. L.

shop on abusive language online , 2019 , pp. 36 - 45 .

[10] V. Gattulli , D. Impedovo , G. Pirlo , L. Sarcinella,

on social networks ( 2022 ) 644 - 651 . doi: 10 .5220/

[11] V. Gattulli , D. Impedovo , G. Pirlo , F. Volpe , Touch

tication via smartphone , Scientific Reports 13 ( 2023 )

[12] P. S. Teh , A. B. J. Teoh , S. Yue , et al., A survey [1] P. K. Smith , J. Mahdavi , M. Carvalho , S. Fisher, of keystroke dynamics biometrics, The Scientific

S. Russell , N. Tippett , Cyberbullying: its na- World Journal 2013 ( 2013 ).

ture and impact in secondary school pupils , [13] A. Jain , L. Hong , S. Pankanti , Biometric identifica-

Journal of Child Psychology and Psychiatry tion , Communications of the ACM 43 ( 2000 ) 90 - 98 .

49 ( 2008 ) 376 - 385 . doi:https://doi.org/10. [14] R. L. Mandryk , L. E. Nacke , Biometrics in gaming

and entertainment technologies, in: Biometrics in a [2] 1G1. 11O/rjrù ., 14A69 .- G7a6l1li0 , . 2V0 . 07G . a0tt1u8ll4i , 6 .Mx.. Gravina , Data Driven World, Chapman and Hall/CRC, 2016 ,

M. Micheletto , S. Marrone , W. Nocerino , A . Pro- pp. 215 - 248 .

caccino , G. Terrone, D. Curtotti , D. Impedovo , G. L. [15] M. Karnan , M. Akila , N. Krishnaraj , Biometric per-

gies for the detection of (cyber)bullying actions: review , Applied soft computing 11 ( 2011 ) 1565 -

The bullybuster project , Information 14 ( 2023 ). 1573 .

[16] S. Marrone , C. Sansone , Identifying users' emo -