Malicious Intent Research Articles

Traversable Ledger for Responsible Data Sharing and Access Control in Health Research

Healthcare institutions and health registries often store patients’ health data. In order to ensure privacy, sensitive medical information is stored separately from the identifying information of the patient. Generally, institutions anonymize medical information while sharing it for external use. However, internal users may also use it for identifying inaccuracies or missing information. Even though internal users may be legally permitted to access sensitive medical information, such access may lead to the identification of the patient, which can be vulnerable to patient privacy. Ensuring the accountability and responsibility of the internal users may lead to tractability in case of adversarial access with malicious intentions. Therefore, a secure system must be developed for the storage and retrieval of health data. To this end, in this paper, we propose a ledger-based system that cryptographically ensures that all access to health data must be logged into a ledger. Nevertheless, the ledger entries must be protected against adversarial access, too. At the same time, the ledger must be traversable by the patients as well as internal users. To address these needs, we propose techniques for the construction of a ledger to permit both internal users and patients to securely traverse and view only the entries to which they are linked.

PERSONAL DATA BREACH ON THE INTERNET: A CASE STUDY ON GOOGLE FONT

Web pages have maintained their popularity from the moment the internet entered our lives becoming a social media catalogue for every sector. Websites facilitated and accelerated many processes such as reaching target audiences, advertising, or sales. Thus, the presence of every sector in the social environment was ensured. With the development of information technology, design opportunities have also developed and the visuality and attractiveness of web pages have gradually increased. Video and text effects are at the top of the design possibilities. Apart from the attractive possibilities of these developing design possibilities, they have also been used for malicious purposes such as stealing or damaging information. This study addresses how the use of Google Fonts conflicts with the European Union's General Data Protection Regulation (GDPR) and the ways to solve this problem. The GDPR has introduced strict rules on the protection and processing of personal data. However, Google Fonts, which is widely used by web developers and designers, sends users' IP addresses to Google's servers without explicitly stating how this data is processed. This is contrary to the GDPR principles of transparency and data minimization. This article elaborates on the privacy implications of using Google Fonts as well as the GDPR violations. As a solution, this study introduces alternatives such as local font hosting, open-source font libraries, and associated best practices. It also emphasizes the significance of the adoption of privacy-oriented design principles by web developers and designers and discusses the potential of these approaches to achieve GDPR compliance. In terms of theoretical and practical perspective, this study aims to provide a roadmap for harmonizing the use of Google Fonts and similar services with applicable privacy-related legislation.

Can LLMs deeply detect complex malicious queries? A framework for jailbreaking via obfuscating intent

Abstract This paper delves into a possible security flaw in large language models (LLMs), particularly in their capacity to identify malicious intent within intricate or ambiguous inquiries. We have discovered that LLMs might overlook the malicious nature of highly veiled requests, even without alterations to the malevolent text in those queries, thus exposing a significant weakness in their content analysis systems. To be specific, we pinpoint and scrutinize two aspects of this vulnerability: (i) LLMs’ diminished capability to perceive maliciousness when parsing extremely obscured queries, and (ii) LLMs’ inability to discern malicious intent in queries that have been intentionally altered to increase their ambiguity by modifying the malevolent content itself. To illustrate and tackle this problem, we propose a theoretical framework and analytical strategy, and introduce a novel black-box jailbreak attack technique called IntentObfuscator. This technique exploits the identified vulnerability by concealing the genuine intentions behind user prompts, thereby compelling LLMs to inadvertently produce restricted content and circumvent their inherent content safety protocols. We elaborate on two specific applications within this framework: ”Obscure Intention” and ”Create Ambiguity,” which skillfully manipulate the complexity and ambiguity of queries to effectively dodge the detection of malicious intent. We empirically confirm the efficacy of the IntentObfuscator approach across various models, including ChatGPT-3.5, ChatGPT-4, Qwen, and Baichuan, achieving an average jailbreak success rate of 69.21%. Remarkably, our tests on ChatGPT-3.5, boasting 100 million weekly active users, yielded an impressive success rate of 83.65%. Additionally, we verify our approach across a range of sensitive content categories, including graphic violence, racism, sexism, political sensitivity, cybersecurity threats, and criminal techniques, further highlighting the considerable impact of our findings on refining ”Red Team” tactics against LLM content security frameworks.

Bangladeshi False message detection using the Tree and characterized by procrastination classifiers

False message is much moreexotericdue to the rapid use of social media. The ability to detect False message is recognized among the riskiest types of manipulation, as it is formed with the malicious purpose of misleading consumers. Many researchers have already suggested several False message detection systems based on social context and diffusion. In this paper, we have presented a system that can expose False message related to the subject of Bangladeshi digital message content. Moreover, machine learning classifiers named Extra tree and Procrastination have been selected. For feature extractions, we have used Term-Frequency-Inverse Document Frequency (TF-IDF) and countvectorizer simultaneously, and countvectorizer separately. After doing two types of experiments, the Extra tree classifier gives the best result for the first experiment and the Procrastination classifier gives the average result for the first experiment,where both classifiers give the average result for the second experiment.

Advancements in Deepfake Detection : A Review of Emerging Techniques and Technologies

This review paper provides a comprehensive analysis of the current state of deepfake detection technologies, driven by the growing concerns over the misuse of synthetic media for malicious purposes, such as misinformation, identity theft, and privacy invasion. The motivation behind this work stems from the increasing sophistication of deepfake generation methods, making it challenging to differentiate between real and manipulated content. While numerous detection techniques have been proposed, they often face limitations in scalability, generalization across different types of deepfakes, and robustness against adversarial attacks. The aim of this paper is to critically assess existing deepfake detection approaches, highlighting their strengths and weaknesses. The objectives include categorizing detection methods, evaluating their performance in diverse contexts, identifying the limitations of current technologies, and proposing future research directions to enhance detection efficacy and adaptability.

Comparative effects of ranolazine and trimetazidine on cardiac structure and exercise capacity in a healthy rat model: a new doping agent?

Abstract Background/Introduction Anti-anginal drugs that have been shown to improve myocardial energy use include ranolazine and trimetazidine [1]. An inquiry is warranted into their ability to impact cardiac functions and exercise capacity. Purpose This study aims to evaluate the effects of ranolazine; and trimetazidine as a positive control group on cardiac structure and exercise capacity in a rodent model in comparison with the control group. Methods In a randomized, controlled trial, 24 Sprague-Dawley rats were divided into three groups: control, ranolazine, and trimetazidine. Rats were given daily doses of ranolazine (50 mg/kg/1 ml) [2], trimetazidine (10 mg/kg/1 ml) [3], or saline (control) via gastric gavage for 30 days. At the beginning of the trial, as well as on the 15th and 30th days, the rats were weighed, placed on rotarod, and assessed using transthoracic echocardiography (Figure 1). The rats underwent 21 days of forced swimming testing. According to the guidelines for rat exercise testing [4], signs of exhaustion, lack of coordinated movements, and failure to surface for three seconds following submersion in the water were designated as grounds for terminating the test. Weight, rotarod durations, M-mode and Doppler measurements, and forced swimming test durations were recorded on the 0th, 15th, and 30th days. Results No discernible differences were found between the groups in the initial evaluations. By using echocardiography, the left ventricular masses expanded in all three, indicating the emergence of an athlete's heart. In comparison to the control group, the ranolazine and trimetazidine groups showed considerable increases in their ability to exercise. In particular, the swimming times of the trimetazidine group and the ranolazine group improved to a mean of 1889 seconds by the 21st session, 1700 seconds in the trimetazidine group, and 1548 seconds in the control group (Figure 2), respectively, indicating enhanced cardiac performance. Conclusion(s) Ranolazine exhibits significant potential to enhance exercise capacity and cardiac efficiency in rats, surpassing the effects observed with trimetazidine. Trimetazidine is already registered as a doping agent in WADA [5]. These findings highlight the potential use of ranolazine as a performance enhancing agent in healthy subjects for a malicious purpose of creating unfair advantages in professional sports. Further studies in human subjects are needed.Figure 1Figure 2

Discerning Transuranic Sub-Concealments in Nuclear Waste Barrels Using γ-Ray Spectrometry: A Non-Invasive Assay Strategy To Combat the Threat of Nuclear Diversion.

Insufficient inventory control arising from inadequate monitoring procedures can lead to vulnerabilities in nuclear security. In addition, insider threats, by either malicious intent or negligence, can pose a substantial risk by exploiting such deficiencies to perform unlawful actions, such as theft or diversion, which may lead to compromised nuclear security. Interim storage barrels, intended for temporary containment of low-density nuclear waste, require special attention in this regard. In certain scenarios, the inadequacy of the existing waste barrel assay methodologies to identify and correctly measure any undesirable transuranic subconcealment present inside may pose a risk of nuclear diversion. The present work aims to establish an absolute waste barrel assay procedure to mitigate such risks by offering a robust method for the detection and assay of transuranic subconcealments in nuclear waste barrels using high-resolution γ-ray spectrometry. Challenges related to the varying amounts of attenuation experienced by the γ-rays within the subconcealments, low-density matrix, and the steel barrel wall have been addressed by an iterative photopeak efficiency transfer approach. The methodology has been verified for the assay of seven mock-up samples mimicking nuclear diversion attempts using waste barrels, for which conventional barrel scanning would be inadequate and tomographic scanning would be highly time-consuming. Using this methodology, Pu and 241Am isotopes have been assayed within <10% of the expected value for the majority, with the measurement uncertainty of <10%. The present method, which is simple and noninvasive, can identify inconsistencies with the labeled inventory and detect potential diversion attempts.

Toward the Adoption of Explainable Pre-Trained Large Language Models for Classifying Human-Written and AI-Generated Sentences

Pre-trained large language models have demonstrated impressive text generation capabilities, including understanding, writing, and performing many tasks in natural language. Moreover, with time and improvements in training and text generation techniques, these models are proving efficient at generating increasingly human-like content. However, they can also be modified to generate persuasive, contextual content weaponized for malicious purposes, including disinformation and novel social engineering attacks. In this paper, we present a study on identifying human- and AI-generated content using different models. Precisely, we fine-tune different models belonging to the BERT family, an open-source version of the GPT model, ELECTRA, and XLNet, and then perform a text classification task using two different labeled datasets—the first one consisting of 25,000 sentences generated by both AI and humans and the second comprising 22,929 abstracts that are ChatGPT-generated and written by humans. Furthermore, we perform an additional phase where we submit 20 sentences generated by ChatGPT and 20 sentences randomly extracted from Wikipedia to our fine-tuned models to verify the efficiency and robustness of the latter. In order to understand the prediction of the models, we performed an explainability phase using two sentences: one generated by the AI and one written by a human. We leveraged the integrated gradients and token importance techniques, analyzing the words and subwords of the two sentences. As a result of the first experiment, we achieved an average accuracy of 99%, precision of 98%, recall of 99%, and F1-score of 99%. For the second experiment, we reached an average accuracy of 51%, precision of 50%, recall of 52%, and F1-score of 51%.

Design Insights for Developing Deepfakes Awareness Games for Adolescents

Around a decade ago, significant advancements in the deep learning domain like Generative Adversarial Networks (GANs) started to facilitate the creation of new synthetical content. The development of deepfake algorithms gained momentum due to the increase in computing power and open-source available resources which facilitated broader access to advanced deep learning methods. While initially deepfakes were created by creative pursuits in various media and marketing campaigns, they started to be misused driven by malicious purposes in various social media manipulation campaigns. These campaigns imply spreading disinformation and misinformation for discrediting individuals and creating confusion in relation to specific topics and events and are meant to altering users’ beliefs and behaviours. Such facts raised important ethical and societal concerns. Although deepfakes still represent a recent research area, a large body of studies is dedicated to, on the one hand, generation, and detection of deepfakes, and on the other hand, to understanding their implications and consequences for society. At the same time, governmental and practitioner efforts are devoted to containing their use and impact on society by proposing various strategies and programs. Nevertheless, both academic and societal efforts are in an incipient stage and often adopt a generalist perspective. This represents a crucial point in relation to the fact that many users in various social media platforms like TikTok is represented by adolescents that are known to be unaware and more vulnerable to media content in general, and deepfakes content, in particular. This represents the knowledge gap that this research aims to tackle by synthesizing design insights as requirements and guidelines that can be used when developing and deploying deepfakes awareness games for producing and/or enhancing awareness of adolescents. Accordingly, a systematic literature review is conducted and merged with previous experience in this domain by adopting a transdisciplinary research approach that merges methods and techniques from the deepfakes, AI, social media, cyber security, and gamification domains.

Enhancing Cybersecurity: Ransomware Detection - A Proof of Concept Study

Presently, our digital landscape faces a pervasive onslaught of diverse cyber threats, encompassing distributed denial of service (DDoS), phishing, ransomware, and smishing, all orchestrated with malicious intent. Counteracting these malicious incursions poses a formidable challenge, particularly in devising efficacious detection solutions. Ransomware incidents are on the ascent, particularly within critical sectors such as healthcare, finance, and telecommunications. Consequently, this paper introduces a proof of concept (POC) aimed at detecting ransomware activities targeting Internet of Medical Things (IoMT) devices. The primary objective of this paper revolves around the identification and evaluation of factors correlating with ransomware attacks on IoMT and then developing a ransomware detector. The experimental framework involves the utilization of a simulated environment mirroring real-world IoMT devices and networks. The methodology integrates diverse approaches, encompassing data collection from IoMT devices, analysis of ransomware behaviour through the study of encryption patterns, and anomaly detection. The POC assesses the efficacy of these methodologies in detecting and responding to ransomware threats, with the experimentation conducted through hybrid analysis within a controlled laboratory setting. The dataset consists of 13 families with a total malware of 9251 taken from GitHub. For POC, a total of 3459 ransomware dataset samples have been selected. As a result of the experiment, thirteen (13) distinct features have been identified as trigger factors for ransomware attacks. These features are obtained by capturing of the processes initiated by the ransomware samples using a process monitor (procmon). A temporal pattern of the processes which include the file system, API, and access based on their frequency of occurrence were used to develop a ransomware detection model. The proposed approach achieved an accuracy of 99.2% and an error rate of 0.8 %, when evaluated on temporal pattern dataset and by using an enhanced artificial neural network (EANN).

Acoustic detection of Unmanned Aerial Vehicles in flight with respect to the signal-to-noise ratio using machine learning

Nowadays, unmanned Aerial vehicles (UAV) are more and more commonly found in the world for a variety of activities. It has become an important task to detect and classify them in order to manage their potential malicious intentions. Current detection and classification methods mostly rely on feature extraction and the use of artificial neural networks that can be heavily impacted by the signal-to-noise ratio (SNR). The approach proposed aims to solve a two-class problem (presence or absence of an UAV) by extracting information on the content of the acoustic source, and then presenting it as input to a neural network. The features used are strongly inspired by the harmonic sound emitted by UAV. Attention is also given on the SNR of the scene. Indeed, various disturbing sources such as trains, helicopters or footsteps are involved from an open library. These signals are added to our recordings of UAVs in flight by tuning their amplitude to obtain different SNRs. The aim is to establish the performance of the detection system as a function of the SNR of the scene.

Research by the method of an extended systematical literature review E-SLR the problem of ensuring the security of personal data when using OSINT

Objective. The amount of personal data in open sources increases, which makes it possible for third parties to access it using open source intelligence (OSINT) methods, which can be used for malicious purposes. The aim of the work is to identify threats and existing methods and means of ensuring the security of a user's personal data and his reputation when using OSINT by intruders, as well as to identify the main problems in protecting user PD taking into account OSINT. Method. The study uses an extended method of systematic literature review (e-SLR), which is a systematic literature review (SLR) supplemented by responses from ChatGPT, GigaCHAT, YndexGPT neural networks. Result. 41 sources were received for the analysis of the problem, on the basis of which threats to personal data were identified: violation of the confidentiality of personal data and the operation of information systems, targeted attacks using social engineering, password disclosure, espionage; protection tools: data processing before publication, anonymization and depersonalization, limitation of personal data, selection of sites, protection using OSINT, creation of complex passwords, use of protection tools, organizational measures; problems in the development of protection tools: working with big data, unreliability of information and sources, labor-intensiveness of data analysis, technical limitations, bias, ethical and legal aspects. Conclusion. The results were used to develop models for protecting personal data in open sources, methods and means for detecting and preventing violations of their security.

Trustworthiness evaluation of workers in critical facilities using electroencephalography-based acquaintance test

This study emphasized the critical importance of prioritizing workers' trustworthiness in safeguarding critical facilities from the potential harm caused by insiders' betrayal. By noting the limitations of existing physical protection systems in critical facilities, this study showed that effectively characterizing and detecting malicious insider activity can be possible by observing immediate brain physiological reactions to specific stimuli. Based on the finding, a novel approach using electroencephalography (EEG)-based acquaintance tests was introduced to objectively assess ’a level of suspicion between colleagues by measuring their brain responses when he/she is exposed to familiar and unfamiliar stimuli. If a person claims falsely about one's acquaintance, the model assumes he/she has malicious or suspicious intent by violating the reporting obligation. The experiment was designed with a relatively short time of monitoring of less than 2 min and with a simple EEG headband device called MUSE. The experiment is to analyze whether the model could provide reliable prediction of disguised acquaintance avoiding complex preparation process. Averaged N170 peak analysis indicated that MUSE provided adequate signal characteristics to classify acquaintance. Also, a machine learning-based subject-wise classification model showed adequate capability to differentiate the EEG signals of acquaintances from unknowns. The final prediction combined multiple single-trial classification results, correctly detected the participant's acquaintance about 94.1 % of the cases, with similar performance when strangers were presented. The results indicated the possibility of using biosignal to enhance security culture and mitigate insider threats in critical facilities, by providing an indication of behavior that disregards security policies and procedures.

Roman original or deliberate fake?

A private collection in Cordoba contains a previously unpublished bronze head. It is not a Roman original. Its iconographic incongruences and its intentional fractures suggest a malicious intention in its manufacture. It is possibly a fake that someone intended to introduce in the art market. A second bronze portrait of the same personage belongs to a private collection in Murcia. It has similarly never been made known. It is also a modern sculpture, confirming the non-ancient origin of the bronze in Cordoba. The joint study of the two objects can serve as a methodological example of the procedure to detect modern fakes of bronze portraits with an antique appearance.

Against linkage: A novel generative face anonymization framework with style diversification

AbstractWith the advent of the digital era, millions of facial images are shared online daily, posing severe privacy threats. Generative face anonymization (GFA) methods generate virtual faces to conceal original identities, protecting sensitive information while preserving utility. However, deep learning based user identity linkage (UIL) methods can link similar faces to the same identity and leverage the linked profiles for malicious purposes, including localization and behaviour prediction. These UIL methods pose a significant challenge to the diversity of virtual faces, a challenge that existing GFA methods have not adequately addressed. To address this research gap, we propose Style Diversification‐based Generative Face Anonymization (SD‐GFA), a framework that generates virtual faces with diverse identities and high visual quality. SD‐GFA features an equalized control module to balance input faces and user‐specified keys, a face generation module with a re‐connection strategy for high‐quality synthesis, and a maximum probability simulation module to enhance diversity. Our experiments demonstrate that SD‐GFA effectively mitigates linkage risk by improving the diversity of virtual faces, while also enhancing their utility and visual quality. This study provides a robust solution to enhance the security of anonymized faces shared on the internet.

Enhancing cybersecurity: A comprehensive approach to addressing the growing threat of cybercrime

The rise of technology and internet usage has revolutionized interactions and business practices but has also significantly increased the prevalence of cybercrime. Cybercrimes, including data theft, ransomware, attacks on critical systems, and botnet-driven attacks, have grown more sophisticated and damaging. Recent research by Mphatheni and Maluleke (2022), Cascavilla, Tamburri, and Heuvel (2021), Shah and Chudasama (2021), Dupont and Whelan (2021), Djenna et al. (2023), and Back and LaPrade (2020) offers a detailed analysis of contemporary cybersecurity challenges and emerging solutions. These studies emphasize the need for a comprehensive approach to combating cybercrime. Mphatheni and Maluleke (2022) highlight the absence of a universal definition for cybercrime, which undermines prevention efforts and overlooks the global economic impact, particularly in Africa. Cascavilla, Tamburri, and Heuvel (2021) stress the importance of advanced machine learning techniques and threat intelligence for detecting cybercrimes across various web layers, advocating for improved risk assessment and anonymity measures. Shah and Chudasama (2021) propose a new cybercrime taxonomy to address gaps in current legislation, especially in regions like India where crimes often stem from data issues or malicious intent. Dupont and Whelan (2021) call for greater integration between cybercriminology and cybersecurity, promoting a continuous approach that fosters better collaboration between security and crime control domains. Djenna et al. (2023) introduce an advanced deep learning method for early botnet attack detection, enhancing forensic investigations and threat response. Back and LaPrade (2020) emphasize updating cybersecurity practices in academic institutions using a Situational Crime Prevention (SCP) framework to improve digital security. Overall, a proactive and coordinated strategy, incorporating new technologies, enhanced practices, and updated legislation, is essential for effective cybercrime prevention and response. Collaboration among governments, the private sector, and academia is crucial for developing robust defenses against escalating cyber threats.

Effects of Health-related Deepfakes on Misperceptions: Moderating Effects of Issue Relevance and Accuracy Motivation

ABSTRACT Among its many applications, artificial intelligence (AI) can be used to manipulate audiovisual content for malicious intents. Such manipulations – which are commonly known as deepfakes – present a significant obstacle to maintaining truth in sectors that serve the public interest, and it is necessary to take proactive fact-checking actions to respond to the threat they pose. The current study conducted an online experiment focusing on the topic of cancer prevention. In this experiment, we further examined the influence of individual differences, such as issue relevance and motns, in health information consumption. When compared to textual disinformation, health-related audiovisual deepfakes were found to have a significant effect on increasing misperceptions but no such effect on fact-checking intentions. We also found that exposure to such deepfakes discouraged individuals with high issue relevance from engaging in fact-checking. Deepfakes were also shown to have a particularly potent effect on increasing misperceptions among individuals with high (illusory) accuracy motivations. These findings underscore the need for increased awareness regarding the detrimental effects of health deepfakes in particular and the urgent importance of further elucidating individual variations to ultimately develop more comprehensive approaches to combat deepfakes.

Improved deep neural network (EnhanceNet) for real-time detection of some publicly prohibited items

ABSTRACT Public safety is a critical concern, typically addressed through security checks at entrances of public places, involving trained officers or X-ray scanning machines to detect prohibited items. However, many places like hospitals, schools, and event centres lack such resources, risking security breaches. Even with X-ray scanners or manual checks, gaps can be exploited by individuals with malicious intent, posing significant security risks. Additionally, traditional methods, relying on manual inspections and conventional image processing techniques, are often inefficient and prone to high error rates. To mitigate these risks, we propose a real-time detection model – EnhanceNet using a customized Scale-Enhanced Pooling Network (SEP-Net) integrated into the YOLOv4. The innovative SEP-Net enhances feature representation and localization accuracy, significantly contributing to the model’s efficacy in detecting prohibited items. We annotated a custom dataset of nine classes and evaluated our models using different input sizes (608 and 416). The 608 input size achieved a mean Average Precision (mAP) of 74.10% with a detection speed of 22.3 Frames per Second (FPS). The 416 input size showed superior performance, achieving a mAP of 76.75% and a detection speed of 27.1 FPS. These demonstrate that our models are accurate and efficient, making them suitable for real-time applications.

Generative AI Increases Cybersecurity Risks for Seniors

We evaluate how generative AI exacerbates the cyber risks faced by senior citizens. We assess the risk that powerful LLMs can easily be misconfigured to serve a malicious purpose, and that platforms such as HackGPT or WormGPT can facilitate low-skilled script kiddies to replicate the effectiveness of high-skilled threat actors. We surveyed 85 seniors and found that the combination of loneliness and low cyber literacy places 87% of them at high risk of being hacked. Our survey further revealed that 67% of seniors have already been exposed to potentially exploitable digital intrusions and only 22% of seniors have sufficient awareness of risks to ask techno-literate for remedial assistance. Our risk analysis suggests that existing attack vectors can be augmented with AI to create highly personalized and believable digital exploits that are extremely difficult for seniors to distinguish from legitimate interactions. Technological advances allow for the replication of familiar voices, live digital reconstruction of faces, personalized targeting, and falsification of records. Once an attack vector is identified, certain generative polymorphic capabilities allow rapid mutation and obfuscation to deliver unique payloads. Both inbound and outbound risks exist. In addition to inbound attempts by individual threat actors, seniors are vulnerable to outbound attacks through poisoned LLMs, such as Threat GPT or PoisonGPT. Generative AI can maliciously alter databases to provide incorrect information or compromised instructions to gullible seniors seeking outbound digital guidance. By analyzing the extent to which senior citizens are at risk of exploitation through new developments in AI, the paper will contribute to the development of effective strategies to safeguard this vulnerable population.

CoAT: Corpus of artificial texts

AbstractWith recent advances in natural language generation, risks associated with the rapid proliferation and misuse of generative language models for malicious purposes steadily increase. Artificial text detection (ATD) has emerged to develop resources and computational methods to mitigate these risks, such as generating fake news and scientific article reviews. This paper introduces corpus of artificial texts (CoAT), a large-scale corpus of human-written and generated texts for the Russian language. CoAT spans six domains and comprises outputs from 13 text generation models (TGMs), which differ in the number of parameters, architectural choices, pre-training objectives, and downstream applications. We detail the data collection methodology, conduct a linguistic analysis of the corpus, and present a detailed analysis of the ATD experiments with widely used artificial text detectors. The results demonstrate that the detectors perform well on the seen TGMs, but fail to generalise to unseen TGMs and domains. We also find it challenging to identify the author of the given text, and human annotators significantly underperform the detectors. We release CoAT, the codebase, two ATD leaderboards, and other materials used in the paper.