1 Introduction
In the recent era, metaverse technology is rapidly emerging and there are a lot of potential applications that can benefit from these developments, such as healthcare, industry, business, and so on. While there is no single agreed-upon definition of a metaverse [
35], the metaverse is a convergence of physical, augmented, and virtual reality and provides a powerfully immersive experience to users by allowing them to seamlessly interact with the real and virtual (computer-simulated) world. The term “metaverse” is a combination of two terms, i.e., “meta,” which means transcending, and “universe,” which refers to the physical universe and the current virtual world. This is the basic definition of the term metaverse, nevertheless, the literature shows that it does not have a unified definition [
35].
Metaverse allows the creation of shared virtual space by connecting all virtual worlds through the Internet, where digital avatars (i.e., users) can communicate and interact with each other similar to the physical world. Key backbone technologies in the metaverse include artificial intelligence (AI) and extended reality (XR) that leverage different technological developments such as virtual reality (VR), augmented reality (AR), and mixed reality (MR). In addition, similar to the current Internet, metaverse will leverage other concomitant technologies such as information and communication technologies (ICT), 5G, and 6G, but metaverse will provide a qualitatively different experience to its users by enabling real-life-like 3-D experiences through the incorporation aforementioned technologies.
Metaverse allows for the digitalization of traditional brick-and-mortar institutions and businesses—it will be possible to develop virtual markets, digital lands, and digital infrastructure, which can be bought and sold using blockchain and non-fungible tokens (NFTs), which are non-interchangeable units of data stored on a blockchain. Metaverse can be a game-changer in terms of the impact of its potential applications due to the greater immersion, involvement, and personalization possible due to AI-XR. This is the prime reason various corporations have shown great interest in the idealization of the metaverse and are making big bets on developing their own AI-XR-based metaverse ecosystems.
Metaverse is receiving increasing traction from numerous major tech companies worldwide such as Facebook (which is recently rebranded with the name “Meta”), Microsoft, Google, and Amazon. In addition, the widespread adoption of the metaverse is evident in the infusion of billions of dollars of investments by these companies in an attempt to achieve great technological transformation. However, despite such huge traction of the metaverse and its potential to transform existing ecosystems like healthcare, there are numerous challenges associated with the use of AI in the metaverse that may hinder their seamless adoption by end-users in the longer term. In addition, in the backdrop of recent technologically induced social issues, there is a palpable lack of trust and confidence in such technologies.
Since technology can be used both ways (i.e., for good and harm), it is vital that governments, corporations, and society at large seriously consider ethical and moral issues. There are many ethical questions about privacy, security, transparency, accountability, democracy, freedom of speech, and anonymity that technology alone cannot answer. Some specific concerns related to how AI-XR-based metaverse applications will impact humanity are: (1) How will AI-XR-based metaverse applications impact and promote human values and human rights? How will AI-XR-based metaverse promote social welfare and not cause harm to society at large; (3) How can we regulate critical applications of metaverse like healthcare? (4) How do we align the commercial and technical imperatives of AI-XR metaverse applications with human values and promote a moral vision and character development? (5) How do we ensure that AI-XR metaverse application developers do not exploit or manipulate their users? Keeping in mind the aforementioned questions, in this article, we present an analysis of the security, privacy, and trust issues associated with the use of AI-XR in metaverse applications.
Contributions of this Article: To the best of our knowledge, this article is the first attempt towards analyzing the challenges associated with the use of different AI techniques in AI-XR metaverse applications. The comparison of this article with existing survey and review articles that are focused on analyzing security and privacy aspects of AI-XR metaverse applications is presented in Table
1. In summary, the following are the salient contributions of this article:
(1)
We highlight various issues that arise with the use of AI in metaverse applications that mainly include security, privacy, and trustworthiness.
(2)
We present a taxonomy of different potential solutions that can be used to realize secure, robust, safe, and trustworthy AI-XR applications.
(3)
We identify various ML-based use cases across different layers of metaverse architecture and highlight several ML-associated vulnerabilities at each layer.
(4)
We present a case study to highlight the real threat of AI-based vulnerabilities by considering a prospective metaverse application design scenario.
(5)
We elaborate upon various open issues that require further development.
Organization of this Article: The rest of the article is organized as follows: Section
2 presents relevant background. The discussion of challenges related to security, safety, privacy, and trust is presented in Section
3. The taxonomy of different vulnerabilities associated with the use of ML in AI-XR metaverse applications is discussed in Section
4. Different potential solutions that can ensure security, privacy, safety, and trust in ML applications are discussed in Section
5. Various open issues that require further research attention are highlighted in Section
6. We conclude the article in Section
7. The organization of the article is depicted in Figure
1.
3 Challenges in Implementing Secure and Trustworthy AI-XR Metaverse Applications
Despite the significant potential of different AI-XR metaverse applications, there are various challenges related to security, privacy, and lack of trust that can hinder their wide adoption. A few such challenges include privacy breaches, security invasion, user profiling, unfair AI outcomes, and so on. These challenges may directly or indirectly put the users’ safety at risk and can influence social acceptability [
81]. Moreover, as discussed above, metaverse is the integration of different modern technologies such as AI, blockchain, and 5G/6G, therefore, it is likely that the inherent issues associated with these technologies get translated into the metaverse. In this section, we describe different challenges that can hinder the secure, safe, robust, and trustworthy employment of AI-XR metaverse applications. Specifically, we characterized and discuss these challenges in two dimensions, i.e., challenges associated with the use of AI techniques including ML/DL-based methods and XR-related challenges in the metaverse. We will start by first discussing AI-related challenges.
3.1 AI-associated Challenges in AI-XR Metaverse
Modern AI techniques that include ML/DL-based models suffer from different vulnerabilities that hinder the smooth, safe, secure, and trustworthy use of these methods in critical applications such as healthcare, autonomous vehicles, and AI-XR metaverse applications. Below, we briefly discuss various such challenges.
3.1.1 Privacy Issues.
Ensuring the privacy of the end-users will be a major challenge in AI-XR metaverse applications. As these applications are designed to monitor and collect users’ data at an unprecedented fine-grained level, in a bid to create a replica of the digital world [
45], there is a greater danger and risk of privacy breaches [
150]. For example, to create an immersive virtual scene in the metaverse, data from different sensors will be collected and analyzed using AI models, e.g., facial expressions, brain wave patterns, hand movements, eye movements, biometric, and speech data [
150]. This raises obvious concerns regarding the privacy of users and opens a new horizon for digital crimes [
45]. Users’ sensitive information including daily routine activities, personal logs, and schedules will be stored on a server, which ultimately becomes a critical privacy challenge in a publicly distributed network. Such data include body movements, voice, reflexes, and even more critical data that include subconscious and unconscious responses such as eye movements and physiological signals. Features such as eye tracking are readily accessible using commercially available products such as the HTC Vive Pro Eye and Pico Neo VR headsets even though the XR expert Louis Rosenberg recommends banning such features and data collection in non-health-related applications for ethical reasons. In general, the data is collected through on-device sensors at the user site, processed at their devices or nearby local server, and logged as storage in the cloud. Considering the above-mentioned procedure, some malicious attacks can be encountered, which are classified as (i) data collection, (ii) data storage, (iii) data usage, and (iv) user profiling. Furthermore, as AR technology depends on the precise localization of users in the physical world, modern smartphones use Lidar sensors and
Simultaneous Localization and Mapping (SLAM) algorithms to precisely locate the user in the real world. This opens up the possibility of privacy violations, as sensitive information may be exploited for anti-social purposes. Some important concerns related to the privacy of data are described next.
Data Collection. Generally, the data in AI-XR metaverse applications is collected through users’ input either by voice command or textual input and multi-modal sensors including camera, microphone, textual commands, gesture sensors, and wearable devices. These devices are expected to frequently collect personal information such as daily routine activities, voice and biometric data, and personal preferences including shopping items, TV shows, and preferred food choices. In addition, the private data will be used for the creation of avatars for a digital representation of a real human in the metaverse, which also raises privacy issues. For example, the built-in location sensor in the Oculus headset can be used for tracking users’ presence in the real environment with a precise accuracy [
150].
Data Storage. The ultimate aim of developing personalized AI-XR metaverse applications is to aid human beings and ease their daily routine activities. These applications will contain multiple sensing devices that generate a substantial amount of data. Whereas these devices will be resource-constrained with small storage units, which leads to the tradeoff between data generation and storage at the edge level. To address these shortcomings, these devices upload their data along with the corresponding logs to online local or global data centers. Though, the data storage tradeoff is resolved by connecting with the servers. However, it also raises privacy concerns regarding the access permissions and data protection of the consumers. Also, if the communication channel is hacked by an attacker, then the data can be manipulated to get the intended outcomes. The literature suggests that adversaries can extract information regarding the actual data even if the communication is encrypted and can track the location of the users by realizing different attacks such as advanced inference attacks [
152] and differential attacks [
153].
Data Usage and Consent. Continuous data collection through multi-modal sensors including cameras, microphones, and other sensors will be closely involved in the daily routine activities of metaverse users. These sensors collect continuous data regardless of privacy awareness, which is logged over the local or global server(s). Consequently, this procedure raises legal questions regarding the users’ consent and the kind of data that is collected and shared. Moreover, metaverse service providers can also utilize this data to optimize their inference models and to make them robust and more personalized. However, it also raises privacy concerns such as the collection, disclosure, and sharing of the data without the explicit or implicit consent of users.
User Profiling. Similar to the current social media (in which users are considered as the product), everything will be a product alike in the metaverse. Metaverse will act as a meta-platform for different entities (such as users, developers, content creators, businesses, etc.), and this raises questions about data collection and its utilization for user profiling [
40]. Also, the provisioning of the metaverse services requires that the users should be uniquely identifiable in the metaverse. For this purpose, VR headsets/glasses or any other such device can likely be used for illegally tracking users in real life [
127]. Moreover, such devices can be attacked by malicious actors and can be exploited to track users’ real locations for possible digital and real-world crimes. Guzman et al. [
38] presented a data-centric perspective to avoid unprecedented privacy challenges related to data collection and its usage in MR.
3.1.2 Lack of Trustworthiness.
According to the definition of Trust, expressed by Lee and See [
80], in the perspective of automated systems, “
Trust is an attitude that an agent will help achieve an individual’s goals in a situation characterized by uncertainty and vulnerability.” Metaverse is a data-driven technology in which service providers will use different automated tools to assist human beings as a recommendation engine in various domains, e.g., shopping recommendations, movie recommendations, and even recommendations regarding their health. The efficacy of such recommendation systems is highly dependent on the collection of personalized data for intelligent decision-making, e.g., an AI model is trained using the collected fine-grained health data to suggest more accurate recommendations regarding the well-being of users. Despite the huge potential of such features of AI-XR metaverse applications, it also raises many questions about which parameters have influenced the underlying AI model in producing a particular decision and the trustworthiness of such predictions and decisions. Next, we discuss such challenges across three dimensions: (i) Truthful AI, (ii) Transparency, and (iii) Explainable and Interpretable AI.
Need for Truthful AI. The evolution in technology also brings threats among them. Over the past few years, AI-based personal assistants including Siri, Alexa, Astro, and the like, have managed to gain wide social acceptance, and these devices are being used by consumers daily for automating household routine tasks. Currently, misinformation, falsehood, or malfunction in AI-based text or speech analyses and command execution are not considered a matter of concern. However, it is expected that AI-enabled intelligent systems with linguistic capabilities will be a major feature of AI-XR metaverse applications. In this regard, it will be quite challenging to enforce the criterion of truthfulness in AI-based systems to ensure the safe selection of statements and behavior according to the social norms of users while interacting with human society.
Transparency Issues. In general practices, developing AI-based systems is about training sophisticated algorithms on large-scale data to learn an efficient and generalizable model that can be deployed in the real-world environment. However, it is a well-known fact that the performance of these models is directly proportional to the transparency of training data, i.e., AI models will perform as well as their training data. Numerous factors such as input data riddled with poor cleansing, the selection of inherently biased data, underfitting, and overfitting influence the performance of these models and can result in fairness and accountability issues (discussed later in this section). Unlike typical application development, there are no quality assurance tools available to spot bugs and evaluate the bias factor in the training data. For instance, if it was known at which stage the model is going to infer at a perfect scale, then there would not be a need to perform training on such large-scale data. This process is all about the hit-and-trial procedure, which is a quite challenging task to identify the right approximations with better data, hyperparameters, and configuration settings.
Explainable and Interpretable AI. The rapid adoption of AI-based applications in human society has also grown the complexity of the systems, which ultimately requires system understandability to make them legitimate and trustworthy. In a critical human-facing technology such as the AI-XR metaverse, interpretable and explainable AI models are required to answer questions about accountability and transparency of their decisions and outcomes: for example, how the employed model reached the decision and which factors influenced the models to make that decision [
8]. Such questions are particularly important for human-centric applications where the potential impact of AI will be limited if it is not able to provide accurate and transparent AI predictions. Therefore, the key objective of these models is to develop a relationship of trust between human users and AI. Also, in Reference [
88], explainability is described as an inherently human-centric property. However, one of the main challenges in developing explainable methods is the tradeoff between achieving the modesty of an algorithm and ensuring the discretion of sensitive user data. In addition, it is a challenging task to identify the right information while generating a simple yet useful explanation for users. It is worth noting that the terms interpretability and explainability are closely related and are often used interchangeably in machine learning literature, however, these terms are different in practice. Interpretability of the models is defined as the extent to which its outcomes are predictable, i.e., for a given change in the input or model parameter(s), the interpretability enables us to predict the respective change in its output. On the counter side, explainability deals with the explanation of internal processes of the ML/DL models in a human-understandable way.
In the literature, various studies have been presented that aim to understand and identify the requirements for XAI from a human perspective. For instance, Liao et al. [
87] developed an algorithm-informed questionnaire to be asked by various AI developers during semi-structured interviews to identify the gap between the current XAI work and practices to develop user-centered XAI systems. The study concluded that the current research in XAI lacks an understanding of real-world users’ needs and requirements for enhancing transparency in AI. Furthermore, the authors suggested a way forward to reduce this gap that includes addressing users’ needs and reducing the gap between algorithmic output and developing human-understandable explanations. Similarly, the authors in Reference [
88], presented an overview of existing XAI literature from a human-centered perspective. Specifically, they considered the following questions:
What is the role of human-computer interaction (HCI)techniques in shaping XAI research? Furthermore, they highlighted three roles that the HCI technique plays in navigating, assessing, and expanding XAI research that includes: (1) Understanding all AI stakeholders’ needs for diverse explainability; (2) Identifying limitations by considering the gap between the explainability and actionable understanding; and (3) Performing theoretical analysis of human explanations, behavioral, and cognitive processes to develop novel computational frameworks to foster enhanced compatibility between humans and XAI. On a similar note, T. Miller emphasized the need to leverage and incorporate insights from various social sciences including psychology, philosophy, and cognitive science to define explanations for XAI [
95]. Considering these fields, the author also provided a detailed overview of existing literature related to human explanation along with providing insights about how these existing methods can be incorporated into XAI systems.
3.1.3 ML Security Issues.
Despite the state-of-the-art performance of modern AI techniques including DL-based systems, it has been shown that these models are highly vulnerable to carefully crafted adversarial perturbations known as adversarial ML attacks [
136]. The threat of these attacks has been already demonstrated for many critical applications such as healthcare, autonomous vehicles, and so on. On a similar note, AI-XR metaverse applications are essentially critical, as they involve humans, and ensuring their safety from any harm is profoundly important. On the counter side, the existence of these challenges raises many concerns about the safety, security, and robustness of AI-based metaverse applications, thus hindering their practical deployment. As it is equally important that any AI-based system should be equally trusted by all stakeholders involved including service providers, developers, and end users. These challenges are detailed later (Section
4).
3.1.4 Lack of Fairness and Accountability.
Modern AI methods, like advanced DL models, lack fairness and accountability in their decisions [
111]. However, such questions are particularly important for critical applications like AI-XR, in which the model’s decisions can have life-threatening consequences for the end-users. Moreover, AI models are developed using training data, which will be mainly collected from human users in AI-XR metaverse applications for providing immersive experiences. Humans possess certain biases that will be readily reflected in the data they generate, and when this data is used for training AI models, the data bias will be directly translated into the developed AI-based system. As a result, the model will be biased towards certain samples that contain certain features (bias), and its decision will not be fair. On a similar note, the critical nature of AI-XR metaverse applications demands accountable decisions. Consequently, data bias if remained unaddressed can ultimately lead to unintended consequences [
79].
3.1.5 Identity Theft and Authentication Attacks.
Users’/avatars’ identities in the metaverse can be stolen or impersonated illegally, leading to authentication and access control issues in the interconnected virtual worlds. Identity theft in the metaverse will be more dangerous than traditional attacks. The identity of a user once stolen will reveal everything about that person’s digital assets, avatars, and social relationships. The attackers can exploit different vulnerable VR gadgets and other service authentication loopholes to realize identity theft attacks and can steal the victim’s secret keys of digital assets and bank details. It has been reported that about 17 users in the OpenSea NFT marketplace were hacked through a phishing attack and flaws in the smart contract that resulted in a loss of $1.7 million.
1Metaverse will leverage different biometrics and password-driven technologies to authenticate users and their avatars in the virtual worlds. The attacker can evade such authentication systems to impersonate real users’ identities to get control of the whole virtual world. Evading AI-based biometric systems has become easier with the advancements in adversarial ML research. Therefore, AI-empowered speech-and-face-recognition-based biometric systems can be easily attacked to realize impersonation attacks. Once the attacker has access to the metaverse, it can exploit the data generated by the victim’s devices to deceive him, committing a crime in the virtual space. However, the exposure of biological data when used for authentication purposes can also lead to severe consequences [
77]. Moreover, the authentication of social friends of a user using their avatars is much more challenging in the metaverse as compared to real-world identity authentication. In this regard, facial data, voice, and videos can be used to develop an AI-based avatar authentication system, however, the unsolved inherent issues of AI can still hinder its practicality.
3.1.6 The Bias Problem.
Bias refers to a model making certain unethical assumptions about the data. Human bias along with its many aspects has been studied by researchers in many disciplines, including law, psychology, and so forth. In Reference [
99], bias is defined as “the prejudice or inclination of a decision made by an AI system which is in a way considered to be unfair for or against one person or group.” Bias in recommendation systems, advertising algorithms, facial recognition systems, and risk assessment tools has been widely studied in recent years. In metaverse applications, data will be collected from a heterogeneous group of people and sources having their own characteristics, stereotypes, and behaviors, which introduces different biases in the collected data. In References [
93,
102,
135], the authors discuss different kinds of bias based on the sources and the types of bias. On the base of sources, these biases have been divided into further categories: biases caused by data, biases caused by algorithms, and biases caused by user interaction.
3.2 XR-related Challenges in Metaverse
AI-XR metaverse applications are essentially human-centric, and ensuring the security, privacy, security, and robustness of such applications is of utmost importance. It has been envisioned that an entirely new form of digital media will emerge from the use of VR and AR in the modern metaverse (TV, print media, and the web). In recent years, there has been immense discussion regarding the concerns about surveillance capitalism, which is happening on the Internet in different applications. Many large tech organizations providing Internet services, such as Facebook, Google, Microsoft, and Amazon, collect large amounts of data related to the surveillance of their users, which is then used to satisfy the needs of advertisers [
146]. The pioneers of VR and AR such as Jaron Lanier [
78] and Louis Rosenberg
2 have predicted that the concerns about surveillance are expected to rise in metaverse. For example, it has been shown how reconfiguring AR in Pokémon Go (an AR mobile game) drew unexpected audiences to museums and public spaces like trains to fill in the space thus creating a form of virtual trespassing. It has been reported that people were putting their lives in danger to pursue virtual characters. This highlights that safety concerns may arise when such immersive technologies are engineered for gaming and experiences. Below, we discuss the key challenges that are hindering metaverse applications in general and we will later discuss the specific challenges that arise with the use of different AI techniques in metaverse applications (Section
4).
3.2.1 Safety Issues.
There are various concerns regarding the mental and physical safety of metaverse users. There are several reported incidents of digital harassment, theft, and bullying in XR applications [
92]. The report on “Immersive and Addictive Technologies” highlights rampant incidents of sexual harassment, cyberbullying, and grooming online [
100]. Ensuring the safety of users is a major challenge for AI-XR metaverse applications because of the fact that such incidents have real damage and harm to users despite being experienced in the virtual world. The avatars generated using recent advancements in AI techniques, in particular, generative models, can appear more realistic in AI-XR metaverse applications and can engage users in promotional conversation, thus providing a false sense of a real human behind the avatar. The avatars in such a promotional content are fueled with more personalized data (such as your vitals, emotions, expressions, etc.) to look more realistic. Also, these sales avatars can pitch products to you more persuasively than any real salesman or even a recommendation system due to their access to rich cyber-physical data about you. The research in deep fake technology and photorealistic avatars is already at the stage where computer-generated content is indistinguishable from the real. Such advancements can be leveraged to realize an adversarial attack on AI-XR metaverse applications to get the intended behavior and outcomes.
3.2.2 Potential Antisocial Aspects.
There are various opinions regarding the antisocial aspects of AI-XR metaverse applications. Many people think that introducing AI-XR metaverse may detract the users (humans) from their purposes and may have a somatic effect. In the literature, it has been shown that extended times online can result in users demonstrating post-VR sadness and detachment from reality. For instance, Aldous Huxley in 1932 wrote in his social science dystopian fiction novel that using technology can lead to self-inflicted harm that can influence people to be diverted from their higher priorities and become more prone to being influenced by other interests. As a result of such a quest for technological utopia, the human psyche and society as a whole are greatly afflicted. Social critics have long argued that various digital media, such as television, the Internet, and the Web, make people docile and less connected to the real world. For example, Jerry Mander in 1978 wrote in his book
Four Arguments for the Elimination of Television that TV removes the sense of reality from people, promotes capitalism, can be used as a scapegoat, and all these three factors work together negatively. The modern technological disruptions including the web and social networking services have created a filter bubble detaching people from the real world and the truth. Due to these reasons, the current era is also referred to as a post-truth era [
51]. We may reasonably expect that alienation from the real world will exacerbate with the increasing adoption of VR, AR, and AI-XR metaverse applications, which aim at changing the human perspective of the world. This argument can be supported by the fact that, in 2018, the World Health Organization formally included “gaming disorder” in its International Classification of Diseases following research that shows that technology can promote addictive behavior in people. Moreover, the literature focused on analyzing the social implications of metaverse argues to understand and identify potential psychological problems that can arise in metaverse [
23].
3.2.3 Ethical Aspects.
Any technological intervention involving humans suffers from some serious ethical issues, especially the one that contains intelligence. The
Institute of Electrical and Electronics Engineers (IEEE) has recently published a report on Ethically Aligned Design that mainly focuses on the Ethics of Autonomous and Intelligent Systems [
126]. This report emphasizes the need of developing ethical
autonomous and intelligent systems (A/IS) that promote human well-being and protect human rights through transparent and accountable A/IS and the prevention of the misuse of AI. This report is the collective effort of hundreds of researchers having diverse backgrounds and expertise in important areas such as governance, technology, civil society, and policy-making. This report has a dedicated section on XR and, interestingly, IEEE also has a Global Initiative on Ethics of XR.
3 In this report, various ethical issues related to XR have been highlighted, including users’ preference for virtual life over the real world and complete disengagement with society. In addition, the reports conclude with the following remark regarding XR:
“The nature of XR environments fosters unique legal and ethical challenges that can directly affect users’ privacy, identity, and rights. Society will need to rethink notions of privacy, accessibility, and governance across public and private spaces. New laws or regulations regarding data ownership, free use, universal access, and adaptive accessibility within XR environments may need to be developed.”3.2.4 Regulatory Challenges.
The big tech companies have resisted regulation, decrying the fact that regulation will slow down innovation. However, there are various ethics researchers and social scientists who are arguing for much greater regulation to ensure that consumer rights are protected. In this regard, the EU
General Data Protection Regulation (GDPR) has paved the way for many countries and regions attempting to develop similar regulatory laws to protect Internet users. These regulations mainly emphasize the importance of the non-profiling of users (i.e., limiting the storage of tracking data) and for better transparency (i.e., online services and applications should specify why and what information is being stored). On a similar note, AI-XR metaverse applications are subject to the requirement of being transparent in terms of data collection and utilization and should also be subject to the informed consent of users. Also, the development of such applications requires thoughtful deliberation from a regulatory perspective. For instance, it is worth considering banning non-medical applications to collect vital biomedical statistics due to the high risks of being exploited maliciously. To mitigate the risk of users being manipulated deceptively, metaverse operators may be bounded to transparently declare the staging of virtual products and experiences in the metaverse. Rosenberg, one of the pioneers of VR and AR, has already started to argue about the need for regulation for metaverse applications [
118]. For instance, he suggested leveraging the arguments regarding the regulation of social media for developing a legal and philosophical basis for metaverse regulation. As the metaverse can be deemed as an evolutionary expansion of similar services. Rosenberg argued that the only solution to eliminate ethical and privacy-related concerns associated with metaverse is to shift from an advertising-based to a subscription-based business model in which users pay a subscription fee for accessing the metaverse platform. This eliminates the service providers’ need to monitor their user base to a greater extent, however, this is not a feasible solution, as it is difficult to say whether or not users will pay for a safer metaverse.
6 Open Research Issues
In this section, we highlight various open research issues that are particularly associated with the use of ML/DL models in different AI-XR metaverse applications.
6.1 Developing Generalizable Adversarial Defense Methods
Over the past few years, substantial research attention has been devoted to adversarial ML. However, the literature highlights that the attention devoted to developing adversarially robust ML/DL models is significantly less as compared to developing novel attack methodologies [
109]. In the literature, different defensive techniques have been proposed to withstand adversarial ML attacks; however, each method only works in a specific setting and fails to withstand unseen and powerful attacks (consequently, fails to generalize across a wider class of attacks). However, the literature focused on adversarial ML shows that the diversity and severity of these attacks are increasing with each passing day. Therefore, the development of hybrid and universal defensive techniques is the need of the hour. In addition, it is required that defense techniques should be developed while considering evolvable and adaptable adversaries (who can adapt their capabilities to break defense strategy). The threat of adversarial ML can be a major hurdle in the development of secure, safe, robust, and trustworthy AI-XR metaverse applications, and if it remains unaddressed, can cause unintended severe consequences to users and society. It is highly recommended to consider these aspects while developing ML/DL-empowered human-centric applications like the AI-XR metaverse. Moreover, the worst-case robustness test can be performed from an adversarial lens considering different attack surfaces in individual AI-XR metaverse application architecture.
6.2 Investigating Robustness of Privacy-preserving Methods
looseness-1 As discussed above, AI-XR metaverse applications will collect fine-grained data that may include personal attributes to provide personalized services (empowered by ML/DL models). The models trained with such data can be inferred to reconstruct privacy-related information that can be exploited to get intended outcomes and incentives. Although different privacy-preserving ML techniques proposed in the literature have been shown to be quite successful in preserving data privacy, the literature demonstrates that meaningful information can still be inferred even in the presence of an appropriate privacy-preserving method. For example, it has been demonstrated that homomorphic encryption (one of the widely used encryption techniques) is vulnerable to model extraction attacks [
116]. Similarly, Boenisch et al. [
18] showed that sensitive information can be reconstructed from the shared parameters in FL. This suggests that the investigation of vulnerabilities and limitations of existing privacy-preserving methods can be a good step toward developing robust privacy-preserving methods. Ideally, it is required that the ML/DL models should be developed in such a way that they are by design privacy-aware, i.e., they should not be able to learn any privacy-related features from the data that could be compromised upon model inferences.
6.3 Developing Generalizable Explainable and Interpretable Techniques
Another major limitation of DL models hindering their trustworthy applications in critical applications like the AI-XR metaverse is the lack of explainability and interpretability. This can also be exploited by adversarial agents to craft adversarial perturbations to realize attacks on different AI-XR metaverse applications. Although significant research interest has been devoted to the development of novel techniques to explain and interpret DL models, the literature shows that their application is limited to a certain data type or application [
114]. While the AI-XR metaverse applications will have a complex architecture that will simultaneously use multi-modal data for making different intelligent decisions, existing explainability and interpretability techniques cannot be directly used for explaining and interpreting ML/DL-driven decisions. More work is required to create methods that can be generalized across different data types, models, and applications.
6.4 Developing Ethical Data Analysis Pipelines
Current ML/DL models are not capable of considering different ethical norms that are necessary for human-centric applications like AI-XR metaverse applications. On the contrary, these considerations are yet very important to maximize potential benefits and minimize associated harms to ensure safe, robust, and fair data analysis. In AI-XR metaverse applications, different ML/DL models will be trained using massively large data collected by humans and their interactions with the real and virtual universe. While there is no guarantee that the AI decision will be ethically committed, because the data used for model training might contain data bias that will eventually result in biased decisions. Moreover, the outcomes of ML/DL models will be just a reflection of human behavior (including moral failures even if they are not intentionally committed). Therefore, to increase the trust of different stakeholders involved in AI-XR metaverse applications (particularly, end-users) and to provide them with a sense of safety, fairness, and accountability, it is highly desirable to develop novel techniques to ensure fair and ethical data analysis empowered by various ML/DL techniques.
6.5 Pushing AI on Edge: Embedded ML
One of the feasible approaches to preserve the privacy of AI-XR metaverse users will be to deploy ML/DL models on their smart devices, e.g., smartphones, AR/VR gadgets, tablets, and so on. By doing so, models can be developed and inferred on their devices without requiring to transmit data to a central cloud. We envision that various AI-XR metaverse applications will potentially adopt embedded ML or edge-enabled ML due to the proliferation of different enabling gadgets and smart devices. However, numerous challenges related to underlying hardware computing capabilities will arise when sufficiently large ML/DL models will be deployed on resource-constrained devices. Also, the literature argues that the research on enabling edge AI is at its early stages of development [
108]. Therefore, it is worth investigating the feasibility and potential of deploying M/DL models on embedded devices to ultimately develop secure, private, and robust systems to provide personalized services in AI-XR metaverse applications. We refer interested readers to a recent survey on analyzing the notion of edge-enabled metaverse applications for a more comprehensive discussion on the topic and various challenges [
163].
6.6 AI-XR Metaverse-specific Security Solutions
The future AI-XR metaverse will have a complex structure and will be a combination of various enabling (complex) technologies (that possess their associated challenges related to privacy and security, e.g., adversarial ML). Moreover, the massive connectivity of numerous entities (users, services providers, organizations, etc.) along with the decentralization will even worsen the enormity of security and privacy in AI-XR metaverse applications. Individual vulnerabilities associated with each technology can be exploited to realize a more powerful attack to halt or get control of some services or the entire metaverse. If such vulnerabilities are left unaddressed, then they will eventually lead to novel challenges, thus making it challenging to ensure the secure, safe, and robust operation of metaverse services. Therefore, it is very crucial to understand such challenges and develop customized defense solutions to protect AI-XR metaverse applications and services in general.