Motion-based user identification across XR and metaverse applications by deep classification and similarity learning

Abstract

This paper examines the generalization capacity of two state-of-the-art classification and similarity learning models in reliably identifying users from their motion patterns across diverse eXtended Reality (XR) applications. We introduce a novel dataset comprising motion data from 49 users in five XR applications: four XR games with distinct task and action profiles, and one social XR application without predefined tasks. Using this dataset, we evaluate both models’ identification performance and, in particular, their ability to generalize across applications. Our results show that while the models can accurately identify individuals within the same application, their cross-application performance remains limited. Accordingly, recent approaches to biometric motion-based verification and identification exhibit low generalization capacity. While the results suggest that current risks of unintended or privacy-critical user identification in XR and Metaverse contexts are limited, they also indicate that these risks are likely to grow rapidly as model generalization improves. To support reproducibility and encourage further research on motion-based user identification in typical Metaverse use cases, we release our cross-application XR motion dataset and accompanying code publicly.

1 Introduction

Users of Virtual Reality (VR), Augmented Reality (AR), and Mixed Reality (MR)—collectively referred to as eXtended Reality (XR)—inherently share their motion data with the system, a prerequisite for providing immersive and responsive experiences. Recent research has shown that such motion data—e.g., tracking information from Head-Mounted Displays (HMDs) and hand controllers—contains distinctive patterns that allow modern machine learning models to identify users solely from their movements (Rogers et al., 2015; Miller R. et al., 2020; Rack et al., 2023; Nair et al., 2023a; Baldoni et al., 2025). These capabilities open up two major use cases: as authentication mechanisms, they can grant secure access to systems and resources; and as identity assurance methods, they can foster trust among collaborators in social XR settings. At the same time, they pose emerging privacy risks, as the same techniques can enable unintended user identification in contexts where anonymity is desired—an issue that has raised growing concerns about privacy in the Metaverse (Nair et al., 2024a; Nair et al., 2024b; Miller M. R. et al., 2020), where motion data are often shared across distributed social XR systems.

Preliminary work has already demonstrated remarkably high accuracy in motion-based user identification (Rack et al., 2024a; Rogers et al., 2015; Miller R. et al., 2020; Rack et al., 2023; Nair et al., 2023a). However, these results were obtained from motion patterns generated within the same applications, leaving open the question of how well such models generalize across different XR experiences. This raises an important issue: how critical are the actual risks of unintended user identification today, given the diversity of application types envisioned for a future Metaverse? Concretely, can individuals be identified across different XR applications—a capability that would have profound implications for user privacy? For instance, numerous players of the Beat Saber VR game publicly upload replays of their sessions, including full motion-tracking data, to online leaderboards such as Beat Leader¹ to compare performances. A malicious actor could potentially exploit such data using motion-based identification models to build behavioral profiles, undermining anonymity across other parts of the Metaverse.

Addressing the privacy threats posed by state-of-the-art, data-driven user identification across XR applications requires datasets that capture the essential variance in human motion. However, existing open datasets lack sufficient numbers of participants and do not cover a diverse range of applications exhibiting the required variability in movement patterns (see Table 1). To address this gap, we collected a new dataset comprising motion data from 49 users who engaged in five distinct VR applications specifically selected to maximize movement diversity. The chosen applications range from rhythm games such as Beat Saber and Synth Riders, which involve precise, repetitive actions, to open-ended experiences like Half-Life: Alyx and Superhot VR, where motion is less constrained and highly varied. In addition, we recorded dyadic, multimodal conversations between participants in a Social VR scenario. This setting captures natural, unscripted movements and represents a key use case for future Metaverse environments, complementing the more structured gameplay data.

Contributions

Our results demonstrate that both prominent state-of-the-art models still exhibit substantial limitations in their ability to generalize user identification across applications. Consequently, the current privacy risks posed by such approaches appear moderate under present conditions. However, our cross-application dataset—together with future comparable developments—provides a valuable foundation for advancing model generalization research while simultaneously enabling systematic assessment and monitoring of the associated privacy threat vector. The dataset and the complete training and evaluation code are publicly available in our Identification Across XR Applications repository.

2 Related work

Motion-based user identification extends biometric analysis to encompass human motions to create a representation of an individual’s unique physiological and/or behavioral characteristics. In this context, we focus on typical XR systems that track the head and at least one hand. Following the terminology defined by Jain et al. (2011), biometric user identification systems can be used for either identification or verification tasks. Verification involves confirming or denying a user’s claimed identity, such as when logging into an account. Identification involves determining a user’s identity from a set of known identities, which is typically relevant for applications such as content personalization or advertising. Given our focus on the identification setting, we present related work on motion-based user identification in XR and summarize publicly available datasets.

2.1 Motion-based user identification in XR

This section provides an overview of current motion-based user identification approaches and, in particular, distinguishes between non-pretrainable and pretrainable methods. Non-pretrainable methods are represented by classification learning models that are directly trained on data from the users they should identify at a later point. Rogers et al. (2015) were the first to use motion data to identify 20 individuals. Miller R. et al. (2020) showed that this approach can be scaled to work with 511 users. Rack et al. (2023) showed that deep learning methods can also be used for identification, and Nair et al. (2023a) demonstrated that it is possible to identify up to 50,000 users with 94.33% accuracy using 100 s of motion data. Baldoni et al. (2025) have shown that users can be identified by combining motion data with traffic data. They trained a separate model for each application and demonstrated user identification within a single application with accuracies exceeding 90% in some cases. They also briefly examined whether users could be identified across two applications and reported an accuracy of around 30% in this scenario. This aspect was not further explored in their study. A severe barrier to the real-world applicability of such non-pretrainable methods is that onboarding is expensive since models require retraining for each new user, which takes significant time and resources. Classification learning models also cannot indicate an ‘unknown’ user, as they will always predict a user from the training dataset.

In contrast, pretrainable methods, such as feature-distance and similarity learning, can immediately be used to onboard and identify users without needing expensive retraining. These methods work by producing some distance metric that will be small for motion sequences of the same users and large for different (or unknown) users. Li et al. (2016) demonstrated that feature-distance methods could verify individuals with 95.57% accuracy based on head movements while listening to music. Miller et al. (2021) showed that similarity learning could potentially identify individuals across different types of VR systems. Rack et al. (2024a) showed that similarity learning can indeed be pre-trained and immediately onboard new users, positioning motion-based user identification as a viable solution for real-world applications. Subsequently, Rack et al. (2024c) showed that similarity learning shows comparative performance to a feature-distance model on uniform ball-throwing motion sequences but significantly outperforms complex handwritten in-air signatures. In our paper, we evaluate both non-pretrainable and pretrainable approaches using a state-of-the-art transformer architecture.

2.2 Existing VR motion tracking datasets

We list previously published VR motion tracking datasets in Table 1 and compare them based on the number of users and VR applications. Thus, we differentiate between datasets that feature either specific or nonspecific movements. This distinction is also reflected in the number of possible activities that can be carried out in the different datasets. VR applications with specific movements typically allow for a limited range of activities. In contrast, VR applications with unspecific movements tend to support a broader variety of possible activities. However, the exact number of activities is often only an estimate, as activities may overlap and are sometimes difficult to distinguish clearly.

In datasets with nonspecific movements, participants are not required to follow predefined actions. The tasks are only loosely defined in terms of timing and speed, resulting in a wide variety of activities that users can perform within the VR application. For example, Rack et al. (2023) create a dataset from 71 users playing Half-Life: Alyx. Moore et al. (2023) created a dataset that involved 45 users performing assembly tasks, which they recently expanded to include 108 total participants (Moore et al., 2024). Wen et al. (2024) introduced the VR.net dataset, which included 16 users playing various VR games; however, this dataset is no longer publicly available, and only five of these users participated in the same two VR applications. Thiel and Steed (2022) presented a dataset of 20 participants who played seven VR applications, including Beat Saber, Half-Life: Alyx, Pistol Whip, and Clash of Chefs. Participants were free to choose which applications to play, resulting in each application being played by 2–14 users. Only one participant played four different VR applications, while the others played three or fewer, which were not always the same.

On the other hand, the VR application often dictates specific movements or relies on an external source, typically at a predetermined time. This usually limits the range of possible activities in the corresponding VR applications. Several datasets have been created to capture these movements. Liebers et al. (2024), Liebers et al. (2023), and Liebers et al. (2021) developed several distinct datasets, including one focusing on 16 users performing bowling and archery tasks, another featuring 15 users playing Beat Saber, and a third capturing 16 users interacting with various interfaces, such as buttons and sliders. Miller et al. (2021) contributed a dataset involving 41 users throwing a ball using different devices. Nair et al. (2023b) created the BOXRR-23 dataset, consisting of 100,000 users playing Beat Saber and Tilt Brush. Baldoni et al. (2025) created a dataset with 60 users, where one group of 30 played Beat Saber and Cooking Simulator. The other 30 users played Medal of Honor: Above and Beyond and Forklift Simulator. Lastly, Rack et al. (2024c) provided a dataset where 48 users each entered 80 different writing sequences.

As seen in Table 1, most datasets are restricted to single VR applications. Only three datasets (Wen et al., 2024; Thiel and Steed, 2022; Baldoni et al., 2025) capture users interacting with multiple VR applications. However, only a small number of users were recorded using the same set of VR applications, which prohibits a comprehensive evaluation of cross-application identification. We contribute to the landscape of existing datasets by publishing our own dataset consisting of 49 users using five different VR applications.

3 Concept: Model and application choices

Following the related work, we identified (1) similarity learning and (2) classification learning as the two state-of-the-art machine learning approaches selected for our comparison. We use both transformer-based and Recurrent Neural Network (RNN) architectures. This design builds on prior work (Rack et al., 2024c; Liebers et al., 2021; Nair et al., 2023a), which has demonstrated that deep learning approaches are highly effective for user identification. Previous work has shown that Convolutional Neural Networks (CNNs), RNNs, and simpler methods such as feature distance metrics or MultiLayer Perceptrons (MLPs) are less effective for user identification based on motion data (Rack et al., 2023; 2024c; Liebers et al., 2021). In contrast, both selected machine learning approaches, similarity and classification learning, allow for extensive architecture and hyperparameter optimization, which is described in the upcoming Section 5.

3.1 Similarity learning (pretrained)

The similarity learning is a pretrainable deep learning method and is previously employed in related studies (Rack et al., 2024c; Miller et al., 2021; Rack et al., 2024a). This approach involves learning to map input data into an embedding space where distances between embeddings with the same label are minimized.

Embeddings are multi-dimensional vectors, and distances between them can be computed using various methods. The model is trained so that embeddings with identical labels are close together in the embedding space, while embeddings with different labels are far apart. A key advantage of embeddings is that they allow identification of users who were not present during training, as each user receives a distinctive embedding based on their movements.

3.2 Classification (non-pretrained)

The classification model is a non-pretrained deep learning method. This method, previously employed in related studies (Rogers et al., 2015; Miller R. et al., 2020; Rack et al., 2023; Nair et al., 2023a; Baldoni et al., 2025), has demonstrated strong performance in identifying users. It requires that all users be present during training, as it cannot generalize to unseen users. The model outputs a vector with one dimension per user, and the identity with the highest activation is selected as the prediction. Although this method lacks flexibility in handling new users, it enables us to examine how well a model can handle high intra-class variance, especially when trained on motion data from diverse VR applications.

3.3 Model training

The selected two classification and similarity models are trained on all VR applications (see Figure 1; Subsection 3.4) in a multi-domain training setup to test whether they can learn consistent user representations despite differing interaction styles, motion patterns, and environmental constraints. In contrast to the narrower evaluations of previous work, such as Baldoni et al. (2025), which only followed a domain generalization setup by training the models on one application and testing on another, we perform a more comprehensive analysis of model behavior in different applications.

FIGURE 1

3.4 VR applications

The main requirement for the choice of applications used to create our cross-application dataset and the model evaluation is a sufficient variety of application- and task-specific motion patterns these applications require or involve. We selected the following five VR applications, also depicted in Figure 1:

1. Synth Riders is a rhythm game where players hold two virtual balls and follow a path to dance in sync with the music. After a brief tutorial in the game, all participants play the same set of songs in the same sequence. The game includes only a few different activities; users primarily stand and move their hands back and forth to hit the balls in the game.

2. Superhot VR is a first-person shooter with a unique mechanic: time only moves forward when the player moves, meaning the faster the player moves, the faster time progresses. All participants start at the beginning of the game and go through the integrated tutorial. The game includes many activities the user can perform, such as dodging enemies, shooting with a weapon, punching the enemies, or walking around.

3. Beat Saber is a rhythm game where players use laser swords to slash boxes in the correct direction, following the rhythm of the music. Participants begin with a short tutorial in the game and then play a set of songs in a predetermined order. This VR application has only a few different activities the user can perform: the user slashes the boxes by swinging their hands back and forth, mainly staying in place with minimal movement.

4. Half-Life: Alyx is a first-person shooter where players move freely around the world, solving puzzles and shooting monsters. Participants start later in Chapter 1 when they are given a weapon for the first time and receive a brief introduction to its use. In this game, users have much freedom and thus can perform many different activities through various navigational options. They can shoot, grab objects, solve puzzles through different controller movements, and move around the room directly or via teleportation.

5. Social VR is represented by a diadic multi-modal conversation in a virtual environment between participants. Here, each individual interacts with an experimenter physically placed in a separate room. This Wizard-of-Oz-like setting was used to better control the flow and progress of the interaction. Conversations occur through the VR headset, which includes a built-in microphone and speakers. The topics are not predetermined, but most discussions center around previously played games. These conversations are not recorded. In this VR application, users are free to move as they wish, with no instructions from the application. However, they can move around the virtual room using only physical movement; teleportation is turned off to prevent them from getting too close to each other. During a conversation, there are many different activities that a user can perform, such as deictic gestures, beat gestures, iconic gestures (Kendon, 2004; McNeill, 1992), and many others.

With these five VR applications, we have a sufficient variety of application- and task-specific motions, which we have summarized for each application in Table 2. Two of these—Synth Riders and Beat Saber—feature largely predefined and repetitive movements, while the other two—Half-Life: Alyx and Superhot VR—involve more unconstrained and diverse motion. Additionally, to explore identification performance in a Social VR scenario, we included a fifth application: a Social VR application. This gives us a dataset with a high degree of motion diversity. We will reveal clear distinctions in motion characteristics between the selected applications with preliminary statistical analyzes.

4 Cross-application dataset collection

This section describes the dataset creation process involving the five VR applications. It outlines the data collection procedure and provides details about the participants involved. Participants were recruited through our university’s participant recruitment system. We submitted an ethics application for the study, and the institution’s ethics committee approved it.

4.1 Procedure

The overall procedure of the data collection process is illustrated in Figure 2. First participants were informed about the procedure and the various VR applications. They then consented to the data being used and published in anonymized form. At the start, demographic data were collected for each participant, including height, weight, general VR experience, and VR games previously played. Participants were then given a brief introduction to the VR setup and, before each VR application, an introduction to the VR application and its controls. Each participant used the VR applications in sequence, spending 10–15 min on each, with the option to remove the headset and take short breaks between each VR application. The participants stood during all the VR applications. We chose the fixed sequence of VR applications following Figure 2 to ensure consistency and comparability of the recorded movement data between users in the respective VR applications. The chosen sequence aimed to create an enjoyable experience by alternating between different genres, ensuring variety and minimizing the need for breaks. We selected the Social VR scenario as the final application to facilitate discussion, using the previously played games as conversation starters.

FIGURE 2

4.2 Recording and implementation

Participants used the HP Reverb G2 headset in a 10-square-meter area within our lab. We used Python and the OpenVR (Bruns, 2025) library to log the current application and continuously collect the tracking data, which includes the positions (x, y, z) and rotations (x, y, z, w) of the left and right controllers and the headset. Additionally, we recorded the participants’ field of view using OBS (Open Broadcaster Software) (OBS Project, 2024), enabling us to analyze their actions throughout the sessions. The software for the Social VR environment is based on Latoschik et al. (2019); Merz et al. (2024). It was developed using the Unity game engine (version 2020.3.21f1), with Photon’s PUN2 architecture providing the networking infrastructure. All data is provided in the repository, along with further documentation detailing the individual data points.

4.3 Sample description

The dataset consists of 49 participants, 27 males and 22 females. The participants’ height ranges from 161 cm to 191 cm, with an average height of 175 cm. Age ranges from 19 to 54 years, with an average age of 27. General VR experience ranges from 0 to 200 h, with an average of 17 h 38 participants have never played any VR games, seven have played one VR game, one participant has played two VR games, and three participants have played all the VR games included in the study.

5 Evaluation

This section describes the evaluation we applied. Here, we detail how we preprocessed our dataset, the architecture, and the training.

5.1 Dataset analyzes

We analyzed the dataset using various methods to identify potential differences between the VR applications. This allowed us to determine whether the dataset captures a wide range of movement patterns across applications. First, we examined user movements by calculating the average distance between the HMD and each controller in 1-min intervals across all participants. Second, we analyzed vertical head orientation by computing the rotation angle around the X-axis based on head rotation data. This indicated whether participants tended to look upward or downward. For each participant and application, we calculated the mean and standard deviation of these angles. We selected these metrics because they offer an application-agnostic, interpretable characterization of user behavior: together, they capture both gross physical movement (HMD–controller distance) and visual/attentional demands (vertical head orientation), which are central aspects of how different VR applications are used. Depending on the type of dependent variables, we applied a repeated-measures MANOVA to the three movement-distance metrics and separate repeated-measures ANOVAs to the head-orientation metrics to assess statistical differences between applications. If the MANOVA yielded a significant result, we conducted follow-up univariate ANOVAs on the dependent variables included in the MANOVA. For post hoc comparisons, we used paired t-tests with Bonferroni correction. All tests were conducted with . We performed all analyzes in Python using the scipy library.

5.2 Preprocessing

We preprocess the raw data used for training and evaluation to remove irrelevant information (i.e., noise), which has been shown to improve training performance (Rack et al., 2022). This preprocessing involves several steps based on tools developed by Rack et al. (2024b). Following this, we resample the motion data to 30 FPS and compute additional features, including the distance between the HMD and the controllers, as well as the angle between the controllers, to enrich the data with more informative signals. Subsequently, the data is converted into a Body-Relative-Velocity (BRV) encoding. This encoding method eliminates irrelevant information, such as the user’s position or orientation within the scene, to prevent overfitting and ensure that models focus on the relevant identification signals. To achieve this, we transform the motion sequences into a Body-Relative (BR) encoding, where each frame’s positions and rotations are referenced to the local coordinate system of the HMD. This process also removes the HMD’s position, which is consistently fixed at its local coordinate system’s origin (0,0,0). We then calculate the first derivative between frames (BRV) to determine the positional and angular accelerations from the BR data. After these preprocessing steps, the input sequence consists of 18 features per frame: (rot-x, rot-y, rot-z, rot-w) for the HMD and (pos-x, pos-y, pos-z, rot-x, rot-y, rot-z, rot-w) for each controller (left and right).

5.3 Architecture and hyperparameter search

To optimize identification performance, we investigated various hybrid architectures combining Transformers with Gated Recurrent Units (GRUs) and Long Short-Term Memory (LSTM) networks. Specifically, we evaluated models where a Transformer precedes a GRU or LSTM, as well as configurations in which a GRU or LSTM is followed by a Transformer. Each architecture was tested with varying model sizes and layer counts for the Transformer, GRU, and LSTM components. The architecture that achieved the best performance on our dataset was selected and is shown in Figure 3. Input sequences are first processed by a positional encoder, which adds the frame position to each motion data frame. The data is first passed through a linear layer to increase the feature dimensionality, and then forwarded to the transformer. The output of the transformer is followed by a GRU, which produces the final predictions by taking the values of the last hidden state. The classification learning model includes an additional linear layer after the GRU, as the hidden size of the GRU would otherwise be too small. The models are implemented in Python using PyTorch Lightning and PyTorch Metric Learning (Musgrave, 2025). The complete source code is published in the identification across XR applications repository.

FIGURE 3

Our architecture includes a wide range of hyperparameters, each of which can significantly influence identification performance. As is common in machine learning, optimal hyperparameter configurations are not known a priori and must be determined through a hyperparameter search. We initially defined a broad search space for this purpose, as detailed in Table 3. We refined the search space over time by expanding or narrowing parameter ranges in response to intermediate validation outcomes and launched new sweeps accordingly. In total, we conducted over 1,000 runs using the training and selected the configuration that achieved the highest validation accuracy. The code repository documentation provides a complete overview and explanation of all hyperparameters.

5.4 Similarity learning model

For the similarity learning model, we use the architecture described in Subsection 5.3, which uses cosine similarity between embeddings as a similarity metric. We adopt this distance metric as it has demonstrated good performance in prior studies (Rack et al., 2024c). We split our dataset of 49 users into three disjoint sets, each containing motion data from all VR applications but involving different users. The first 23 users were used for training, the next nine for validation, and the final 17 for testing to evaluate overall performance.

To evaluate the similarity learning model, we do not directly predict which user is being identified. To identify a user, we compare different embeddings and measure how far apart they are. If the distance between two embeddings is small, it means they likely belong to the same user. If the distance is large, it’s likely a different user. To perform this comparison, we take an embedding (called a query embedding) that we want to identify and compare it with many reference embeddings (which come from known users). We then check which reference embeddings are the closest to the query embedding. We use majority voting among the 20 nearest reference embeddings to assign a user identity to the query embedding. No distance threshold is applied to reject uncertain matches; instead, a user is always assigned, even if the nearest embeddings are relatively distant. This follows the closed-set evaluation setup, where each query belongs to one of the enrolled users, ensuring consistency with the classification learning model.

From the test dataset, we calculated different embeddings for each user and each VR application based on the different movement sequences. Each sequence in our final model, as shown in Table 3, consists of a window size of 450 frames, corresponding to a time span of 15 s at 30 FPS. During training and evaluation, a stride of 50 was used, resulting in an overlap of 400 frames between consecutive windows. For different tests, we do not use all reference embeddings from all VR applications but only from certain ones in order to test how the person is identified if you only have specific movement data from them. For example, we can calculate how accurately the model can identify a person in Synth Riders based on their gameplay data from Beat Saber. In this case, we use the embeddings from Beat Saber as the reference embeddings and compare them to the query embeddings from Synth Riders to identify users based on the distance between the embeddings.

We use this method to compute embeddings from one VR application to serve as reference data. We then evaluate how accurately the model can identify individuals using query embeddings obtained from either the same or a different VR application. For the reference embeddings of each application, we selected every 150th user embedding distributed across the entire playtime. For the query embeddings of each application, we included all embeddings. To quantify performance, we compute several metrics. First, we calculate the standard accuracy by checking whether the nearest reference embedding corresponds to the correct identity of the query embedding. Additionally, we compute sequence-based accuracies. For this, we consider a sequence of query embeddings obtained over time from user movement and examine which user is most frequently matched by the nearest reference embeddings across the sequence. We then verify whether this majority identity is correct. Note that the number of available samples per user and application decreases as the sequence length increases.

5.4.1 Comparing embedding spaces across VR applications using GOPA

To better understand how user motion is encoded by the similarity learning model, we investigate the structure of its embedding spaces across different VR applications. Specifically, we aim to examine whether the spatial arrangement of user embeddings exhibits similar relative orientations across these applications. If the underlying structure of each embedding space is preserved—up to an orthogonal transformation (rotation or reflection)—then this would indicate that the model captures consistent patterns of user motion, independent of the specific application context. This assumption implies that the main difference between embedding spaces lies in their orientation rather than their intrinsic structure. To test this hypothesis, we restrict ourselves to a basic alignment via rotation and reflection only, excluding scaling and translation. We apply Generalized Orthogonal Procrustes Analysis (GOPA) to align the embedding spaces and assess how well user embeddings from different applications can be brought into a common space. If the embeddings align well after rotation and reflection, it suggests that the model exhibits structural consistency across applications and that users occupy similar relative positions in each space. This analysis provides insight into the geometric properties of the learned embeddings and the degree to which the model is invariant to application-specific transformations.

We then analyze the resulting rotation matrices to examine differences in the embedding spaces across applications. To prepare for GOPA, we compute a representative embedding for each user and application by aggregating all of the user’s embeddings within one VR application into a single point. This central embedding is defined as the unit-norm mean direction (extrinsic spherical mean) that maximizes the average cosine similarity to all other embeddings of that user within the same application. This approach ensures that the central embedding captures the user’s overall movement signature in that application. After aggregation, we obtain one embedding per user per application. Since the test dataset includes 17 users and five VR applications, we construct five embedding spaces, each consisting of 17 user-specific central embeddings.

We then use GOPA to align these five embedding spaces. For each space, GOPA finds an optimal orthogonal transformation (rotation or reflection) that minimizes the total squared distance between embeddings of the same user across different applications. The resulting transforms satisfy ; therefore they preserve inner products, vector norms, and angles, and hence both Euclidean distances and cosine similarity (equivalently, cosine distance). After alignment, all five embedding spaces are transformed into a shared embedding space where embeddings corresponding to the same user across different VR applications are positioned as closely together as possible. Since GOPA uses only orthogonal transformations (rotations and reflections), cosine similarity is preserved—there is no scaling, translation, or distortion. The resulting orthogonal matrices can also be applied to all other embeddings from the similarity learning model, allowing for consistent cross-application comparison.

5.5 Classification learning model

For the classification learning model, the use of the architecture described in Subsection 5.3 and shown in Figure 3. For this, we split the dataset of 49 users such that, for each user and VR application, 45% of the data is used for training, 20% for validation, and 35% for testing. Each sequence in our final model, as shown in Table 3, consists of a window size of 600 frames, corresponding to a time span of 20 s at 30 FPS. During training and evaluation, a stride of 100 was used, resulting in an overlap of 500 frames between consecutive windows.

We compute several metrics to evaluate the model’s ability to identify users on the test set. First, we calculate accuracy by measuring how well users are identified within each individual game. Additionally, we compute a sequence accuracy metric. Here, we consider sequences of consecutive predictions based on temporally adjacent movement data, noting that the number of samples available per user decreases with longer observation periods. The final prediction is made by selecting the user most frequently predicted within the sequence.

6 Results

In this section, we present the results of our data analysis and the outcomes of the machine learning models. For the similarity learning model in particular, we conducted several evaluation tests, as this type of model allows performance to be assessed in various ways using a single trained instance, as described in Subsection 5.4. In contrast, the classification learning model only allows us to test how well users can be identified within the same VR application. It does not support evaluating cross-application accuracy without training a separate model for each application.

6.1 Dataset analyses

Table 4 summarizes the average movement distances and head pitch angles across the five VR applications. Head movement distances are highest in Superhot VR, approximately twice as large as those in Beat Saber, Synth Riders, and Half-Life: Alyx, and significantly higher than in Social VR application. Controller movement distances are most pronounced in Beat Saber, followed by Synth Riders, Superhot VR, and Half-Life: Alyx, with the lowest values observed in Social VR application. A repeated-measures multivariate analysis of variance was conducted with MANOVA as within-subject factor and HMD and controller movement distances as dependent variables revealed a significant multivariate effect of application, Wilks’ , , . Follow-up one-way repeated measures ANOVAs indicated significant differences in movement distances between applications: head movement distance: , ; left controller movement distance: , ; right controller movement distance: , . Post-hoc tests indicated significant differences in head and controller movement distances between all applications, except for head movements between Beat Saber, Synth Riders, and Half-Life: Alyx, and for right controller movements between Superhot VR and Synth Riders.

Regarding head pitch, the average pitch angle remained close to neutral in Synth Riders, Beat Saber, and Social VR application, indicating a level gaze. In contrast, participants tended to look upward more in Superhot VR, with the highest mean pitch angle observed in Half-Life: Alyx. The standard deviations of head pitch were relatively consistent across applications, ranging from 4.71 to 6.40°. A one-way repeated measures ANOVA indicated significant differences in head pitch angles across applications, , . Post-hoc tests indicated that all differences were significant, except between Superhot VR and Beat Saber, as well as between Synth Riders and the Social VR application. Due to space constraints, detailed post hoc test statistics are provided in the supplementary material available via GitLab.

6.2 Similarity learning

As described in Subsection 5.5, we trained and evaluated a pretrainable similarity learning model using the dataset. The final model was trained for 8 min on a system equipped with an NVIDIA GeForce RTX 4090 GPU, an Intel Core i9-13900K CPU, and 128 GB of RAM. Training was conducted over 22 epochs, each lasting approximately 14 s, with 68 batches per epoch and 926 samples per batch. This corresponds to an average processing time of roughly 0.0002 s per sample.

6.2.1 Overall performance

To evaluate the model’s overall performance in user identification, we used data from all users and VR applications as a reference. This approach aimed to assess the general capability of the similarity learning model. We compute the accuracy directly from the nearest embeddings and do not yet consider a larger time window of the user’s interactions, as is done with sequence accuracy. On average, the model achieved an accuracy of 78.5% across all users and VR applications. Figure 4 shows the mean accuracy for each VR application, computed using only query embeddings from the respective application. The results indicate minor variations in accuracy between applications. Specifically, user identification accuracy ranged from 67,7% in Half-Life: Alyx to 86,0% in Beat Saber.

FIGURE 4

6.2.2 Different application in reference and query

In Subsection 5.4, we present results when both references and queries are taken from either the same or different applications. In this step, we identify the user whose reference embedding is closest to the query embedding. When references and queries are taken from the same application (represented by the diagonal on the left heatmap in Figure 5), user identification accuracy ranges from 72.3% to 88.0%, with an average of 83.1% and a mean standard deviation of 6.1%. In contrast, when references and queries are taken from different applications (all off-diagonal entries on the left heatmap in Figure 5), accuracy drops to an average of 18.0%, with a range between 10.5% and 22.6%. The corresponding standard deviations range between 9.0% and 20.0%, with an average of 15.1%.

FIGURE 5

6.2.3 Sequence accuracy

As described in Subsection 5.4, we compute the sequence accuracy. As shown in the middle heatmap in Figure 5, accuracy increases significantly when the user is observed for 10 min. When references and queries originate from the same application (represented by the diagonal), accuracy reaches 100%. When references and queries stem from different applications (i.e., all elements off the diagonal), accuracy ranges from 9.0% to 57.7%, with an average of 30.8%. The heatmap shows that when both the reference and query come from the applications Superhot VR and Half-Life: Alyx, accuracy exceeds 50%. To compare the sequence accuracy with the classification-learning model, we also evaluate the sequence accuracy for 2:30 min. If we use all Applications as a reference and the separate applications as a query, we achieve an accuracy of 100%. When references and queries stem from the same or different applications, accuracy ranges from 9.30% to 100%, with an average of 41.3% and a mean standard deviation of 21.3%.

6.2.4 Top-3 user identification

Since cross-application accuracy for identifying the correct user is relatively low, we investigated whether our model could at least narrow down the prediction to the top three candidates. The majority voting approach we used not only predicts the most likely user but also returns a ranked list of the top candidates. Rather than focusing solely on the top one prediction, we evaluated whether the correct user appeared among the top three candidates with the highest number of similar reference embeddings. The right heatmap in Figure 5 shows that the 10 min sequence accuracy reaches 100% when both the reference and query samples originate from the same application (i.e., along the diagonal), as measured by whether the correct user appears among the top three candidates. When reference and query samples come from different applications (i.e., the off-diagonal elements), accuracy ranges between 29.1% and 76.4%, with an average of 56.0%.

6.2.5 Comparing embedding spaces across VR applications using GOPA

Building on Subsection 5.4.1, we estimated an orthogonal transformation (rotation or reflection) for each VR application by aligning user-specific central embeddings in the test dataset. Because these transformations are estimated on the same test users used for evaluation, the GOPA-based alignment should be interpreted as a post hoc diagnostic upper bound rather than as evidence of out-of-sample generalization. Figure 6 shows a 2D Uniform Manifold Approximation and Projection (UMAP) of the original 480-dimensional embedding space. In the left plot, embeddings appear broadly scattered across both domains and users, consistent with the findings from Subsection 6.2.2. After applying the application-specific orthogonal transformations, the right plot reveals well-separated, user-specific clusters. To quantify the alignment quality, we measured the average pairwise cosine similarity between users within each VR application. In the original embedding space, this similarity is 0.798. After transforming the embeddings into the shared space using the learned orthogonal matrices, the similarity increases to 0.963. Here, a similarity of 1 indicates identical embeddings, 0 corresponds to orthogonal vectors, and to opposites. We further evaluated the alignment by computing the mean cosine similarity between corresponding user-centric embeddings in each rotated space and the shared embedding space. These values range from 0.979 to 0.988, suggesting close alignment across all applications. To analyze the nature of the transformations, we examined the orthogonal matrices derived for each VR application. The effective orientation change—quantified as the mean angular deviation across dimensions—ranges from to . In the 480-dimensional space, these values represent aggregated reorientations (including possible reflections) rather than interpretable 3D rotations. Additionally, pairwise similarities between the orthogonal matrices range from 0.003 to 0.014, indicating that each transformation encodes a distinct reorientation.

FIGURE 6

To evaluate the practical impact of these transformations, we applied the corresponding orthogonal matrix to each test embedding based on its VR application and assessed nearest-neighbor identification in the shared embedding space. Since the orthogonal matrices are learned directly from the test embeddings, the resulting identification accuracies likewise represent an optimistic upper bound under idealized alignment conditions rather than a deployable generalization procedure. Figure 7 shows the resulting accuracy heatmap, using different applications as reference and query sources (as described in Subsection 6.2.2). When reference and query embeddings come from the same application (diagonal), accuracy averages 82.9%. Cross-application accuracy (off-diagonal) averages 52.3%. For 10-min sequences, within-application accuracy increases to 100.0%, and cross-application accuracy to 94.3%.

FIGURE 7

6.3 Classification learning

As described in Subsection 5.5, we trained and test a non-pretrainable classification model using the dataset. The model was trained for 15 min on a system equipped with an NVIDIA GeForce RTX 4090 GPU, an Intel Core i9-13900K CPU, and 128 GB of RAM. Training was conducted over 44 epochs, each lasting approximately 21 s, with 96 batches per epoch and 315 samples per batch. This corresponds to an average processing time of roughly 0.0008 s per sample.

The model achieved a maximum validation accuracy of 46.7%, indicating its capability to identify individuals correctly. Using this final model, we assessed its generalization performance on the test data, achieving an accuracy of 43.2%. To further evaluate the model, we computed the sequence accuracy over a 2:30 min window. Due to the separation of training, validation, and test sets, the number of contiguous sequences was limited, resulting in a sequence accuracy of 46.2%. We also evaluated user identification performance across different applications by separately measuring overall accuracy and 2:30 min sequence accuracy on the test dataset. The model achieved 46.5% accuracy and 49.6% sequence accuracy on Synth Riders, 41.9% accuracy and 45.3% sequence accuracy on Superhot VR, 68.7% accuracy and 73.0% sequence accuracy on Beat Saber, 27.8% accuracy and 30.7% sequence accuracy on Half-Life: Alyx, and 28.4% accuracy and 29.4% sequence accuracy in the Social VR application.

7 Discussion

In this work, we investigated the capabilities of two prominent state-of-the-art machine learning models to identify individuals based on their motion patterns across different XR applications.

Our results suggest that the ability of recent models, whether pretrainable or non-pretrainable, is yet limited when identifying individuals across different applications. However, our dataset includes a wide range of VR applications with specific and unspecific movements, thus encompassing many different activities. This diversity provides a valuable opportunity for researchers to train and evaluate new machine learning models. It enables a detailed analysis of how future machine learning approaches might perform within individual VR applications and across multiple applications.

7.1 Within-app identification

As shown in prior work, users can be reliably identified within the same application (Rack et al., 2024a; Rogers et al., 2015; Miller R. et al., 2020; Rack et al., 2023; Nair et al., 2023a). Our similarity learning model confirms this finding: although it was trained on all data from the application, it can accurately identify new users when prior data from the same application is available. Both rhythm games yielded high identification scores, likely due to minimal activity variation, as users usually performed similar motions. Identification accuracy in Superhot VR was comparable, which is reasonable since user activities, such as moving or shooting, remained relatively consistent. Accuracy in the Social VR application was also comparable, as the range of movements in this setting is not overly complex. In contrast, Half-Life: Alyx presents a broader range of possible activities, making the identification task more challenging for the model. Figure 4 initially suggests that the similarity learning model performs worse than previous approaches. This is likely due, on the one hand, to a higher variance in the data, and on the other hand, to the fact that we are directly evaluating the output—i.e., the nearest embedding. However, as shown in the middle heatmap in Figure 5, when a longer sequence is considered, the accuracy increases significantly, reaching 100%.

7.2 Cross-app identification

Our similarity learning model demonstrates relatively strong performance when identifying users within a single application. However, cross-application identification—recognizing a user across different contexts—proves significantly more challenging. The average accuracy in this setting is just 18.0%, which is much better than random guessing, which is 5.8% . This is also evident in our classification learning model. Similar to the similarity learning model, it was trained using all applications, but unlike the similarity learning model, it included all users during training. Despite this, the classification learning model still struggles to identify individual users, achieving only 43.5% accuracy. Baldoni et al. (2025) reported similar results with their classification learning model when they trained it on one application and then attempted to identify users in a second application. When we look at the individual accuracy scores from the VR applications, we see that the model has fewer difficulties identifying users in Beat Saber compared to the other applications. This might be because the user movements in Beat Saber were the most consistently distinguishable for the classification learning model. This may have allowed the model to focus on these patterns during training when it had access to data from all VR applications.

A likely reason for the low accuracy of both models is the pronounced difference in movement patterns across applications. Our statistical analyses of the dataset also confirm these differences, revealing clear distinctions in two dimensions across the datasets. In applications such as Beat Saber or Synth Riders, users exhibit more intense movements—as measured by controller travel distance—compared to Social VR application or Superhot VR. Head movements also differ significantly. This variability makes it difficult for the models to learn a consistent representation of user identity across contexts. This is also evident from the analysis of the embedding spaces, as shown in the left panel of Figure 6. Initially, the user embeddings from different application-specific spaces are not aligned. However, this misalignment can be corrected through an orthogonal transformation—specifically, a rotation or reflection—without requiring any scaling or translation. This indicates that the structure of user-specific embeddings is preserved across applications and differs mainly in orientation. The learned orthogonal matrices suggest that all embedding spaces require similar rotational adjustments. This consistency implies that none of the VR applications induces a fundamentally different embedding structure. Thus, the model exhibits rotational variance across applications, which can be effectively addressed through orthogonal alignment.

The similarity learning model appears to capture a shared underlying motion pattern that is consistent across application contexts, up to a global rotation or reflection. Performance results based on the aligned embeddings (Figure 7) confirm the effectiveness of this approach, despite the fact that the transformations were estimated using only one user-specific central embedding per application. After alignment, cross-application identification accuracy improves by more than 50%, reaching 94.3% for 10-min sequences. These improvements demonstrate that embeddings can be successfully mapped into a shared space that preserves user identity across applications. It is important to note, however, that the orthogonal transformations were fitted using data that includes the test dataset. Therefore, the accuracy improvements do not reflect generalization performance. Instead, they show that orthogonal alignment is highly effective in reducing inter-application variation within the embedding space.

However, in order to improve the accuracy of the model and better identify users between VR applications, we can calculate sequence accuracy by observing the user over a longer period of time. For example, in the similarity learning model, accuracy increases significantly—exceeding 50%—when observing users across applications such as Half-Life: Alyx and Superhot VR. This improvement is likely because both VR applications involve more undefined or exploratory movements, leading to a greater overlap in motion patterns and, consequently, more similar embeddings. This also highlights an advantage of the similarity learning approach, in addition to the fact that new users can be identified after training. Since movements are translated into embeddings rather than directly mapped to users, the model can assign similar embeddings to similar movements, even when originating from different applications. As a result, it can still correctly identify the user, despite potential difficulties in distinguishing between the contexts. The classification model also benefits from longer observation, showing a moderate accuracy increase of over 5% after 2:30 min. This time limit reflects the constrained dataset size per user and replay, which reduces the number of test samples and may hinder generalization. To enable comparison, we evaluated sequence accuracy over the same 2:30 min duration using the similarity model. When using the mean over all individual application combinations as both the reference and the query, the average sequence accuracy is comparable to that of the classification model. However, when aggregating all applications as references, the similarity model reaches 100% accuracy. This discrepancy highlights key differences between the models. First, the data splits differ: the classification model trains on more users but with fewer samples per user, increasing inter-user variability at the cost of within-user generalization. Second, the evaluation protocols are fundamentally distinct. The similarity model leverages comparisons across applications, enabling finer discrimination. In contrast, the classification model must predict identity from global representations without contextual cues, making the task inherently harder. Further analysis of the similarity model shows that top-three sequence accuracy continues to improve with a 10-min window, suggesting that the model effectively captures user-specific patterns.

The core challenge is bridging the gap between the different types of movement exhibited by the same user across different applications. These intra-user differences are often greater than inter-user differences within similar applications, further hampering generalization. Increasing architectural complexity—through additional layers or alternative layer types—does not appear to be a sufficient solution. Instead, a fundamental rethinking of modeling approaches may be required. Despite these challenges, the potential for improved cross-application identification is evident, as shown by the top three prediction results from the similarity learning model. Still, it remains an open question whether an ideal model can truly transfer user-specific movement profiles from one application context to another.

Currently and increasingly in the future, the reliable identification of specific individuals across applications based on our models constitutes a relevant privacy threat. While the identification accuracies observed in our experiments remain below perfect recognition, they are consistently and clearly above chance level—particularly when individuals are observed over longer time spans, where performance further improves. Moreover, the substantially higher top-3 accuracy indicates that users can already be narrowed down to a small set of plausible candidates. This creates a practical attack scenario: even if a user intends to remain anonymous within one application, they can be recognized across applications, or at least strongly narrowed down, through distinctive motion patterns. The real-world implications of such reidentification are substantial. Cross application identification can merge separate identities, for example, across VR games and Social VR, thereby revealing user information. If a user is identified across multiple VR contexts, activities, usage times, routines, preferences, social interactions, or sensitive interests can be aggregated over extended periods and attributed to a user without explicit consent. A concrete pathway that is already plausible today arises from publicly shared VR motion data. In games like Beat Saber, users upload gameplay data to platforms like Beat Leader. Such data can be used to train models or to create embeddings that can then be applied in other applications, for instance, in Social VR environments where the same users intentionally remain anonymous or do not want to be linked to other games. For such a scenario, our similarity learning approach can already be used to recognize users across applications or at least narrow them down significantly. At the same time, it should be considered that the model in this work was trained on only 23 users. In realistic settings with substantially more users, the embedding space would likely be more densely populated, which could in turn reduce identification accuracy.

Taken together, our findings, both the observed accuracies and the plausible application scenarios, demonstrate that cross-application identification constitutes a relevant privacy risk and should be addressed accordingly. Our dataset with 49 users and its wide spectrum of movement styles, ranging from highly structured to completely unstructured, provides a valuable basis for systematically studying these risks and for evaluating future countermeasures.

8 Future work

The state-of-the-art models used were developed with previous datasets from individual VR applications. Enhancing these models could improve their reliability and accuracy through several avenues of future work. One approach involves refining the model architectures by incorporating strategies from other areas of machine learning and adapting them to user identification. For instance, Domain-Adversarial Neural Networks (DANNs) could be explored, where an additional application classifier is trained to reverse the learned feature representations and contribute to the loss function, thereby encouraging the model to focus on user-specific rather than application-specific patterns. Alternatively, further refinements to the loss function could guide the model to better generalize motion patterns across different applications. Additionally, improvements in data preprocessing could further enhance model performance. Currently, our models rely primarily on motion data for user identification. Integrating additional contextual information, such as details about the XR application or types of movements performed, could increase user identification accuracy, potentially improving cross-application performance. A further promising direction is to investigate GOPA-based alignment in a generalization-safe setting. In this work, GOPA was used as a post hoc diagnostic tool to reveal the potential structure of a shared embedding space across applications. Future work should develop and evaluate protocols that learn orthogonal transformations only on training/validation users and then apply them to unseen test users and applications, in order to assess the stability, robustness, and practical feasibility of GOPA as a cross-application alignment method. Another valuable direction is to evaluate cross-application generalization by training models on a subset of applications (e.g., four) and then measuring identification performance on a completely unseen fifth application. This setup simulates a realistic scenario of deploying a model to a new application, revealing how well user identity transfers to previously unseen application contexts. It would also be useful to take a closer look at sequence-level accuracy by systematically analyzing how cross-application identification performance between different VR applications improves as sequence length increases.

When considering the use of additional datasets for user identification across various applications, it is crucial to address privacy concerns. In future work, the dataset can also be used to test current approaches for anonymizing motion data and potentially improve these methods.

9 Limitations

Our work exhibits the following limitations that should be considered when interpreting the results. Firstly, our dataset narrow range of ethnicities may not fully capture the diversity of motion profiles across different cultures and populations. Including a more diverse participant pool would provide a more comprehensive understanding of these variations. Secondly, the fixed order of application use during data recording may introduce ordering effects. For example, participants may become more confident with each application, resulting in learning effects, or they may experience fatigue—particularly in later applications such as Social VR—which could alter motion signatures compared to earlier sessions. This potential bias should be acknowledged and discussed in future analyses. Thirdly, while we present the dataset and initial results, these findings were obtained using current models. They should not be considered definitive when determining whether individuals can be reliably identified across VR applications. Fourthly, the similarity learning experiments were conducted on a test set comprising only 17 individuals; as a result, the outcomes may be sensitive to individual variations and should be interpreted with caution. An important limitation is the comparability between the similarity learning model and the classification learning model. Due to substantial differences in dataset structure and evaluation protocols, a direct comparison between the two models is inherently difficult. Finally, the GOPA analysis provides only an internal, post hoc view of the embedding space. Since the rotation was fitted on and applied to the entire test set, the reported alignments should be interpreted as a diagnostic upper bound, not as a deployable, generalizing solution.

10 Conclusion

This paper explored the capacity of state-of-the-art similarity learning and classification learning model to identify users based on their motion data across different VR applications. For that, we developed and released a novel dataset comprising motion data from 49 users across five distinct VR applications, allowing us to assess identification performance both within and across applications. Our results demonstrate that while models can achieve high accuracy in identifying users within a single application, cross-application identification remains significantly more challenging. While cross-application identification is not yet highly accurate, it already outperforms random chance, raising privacy concerns regarding future advancements. As motion-based identification models continue to evolve, the provided dataset will be valuable to benchmark new models and assess the growing risk of identification in XR environments.

References

• 1

  BaldoniS.BenhamadiS.ChiariottiF.ZorziM.BattistiF. (2025). “Movement- and traffic-based user identification in commercial virtual reality applications: threats and opportunities,” in 2025 IEEE Conference Virtual Reality and 3D User Interfaces (VR), Saint Malo, France, 08-12 March 2025 (IEEE), 72–81. 10.1109/VR59515.2025.00032

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  BrunsC. (2025). Pyopenvr. Available online at: https://github.com/cmbruns/pyopenvr.

  • Google Scholar
  • View reference in article

• 3

  JainA. K.RossA. A.NandakumarK. (2011). Introduction to biometrics. Boston, MA: Springer US. 10.1007/978-0-387-77326-1

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  KendonA. (2004). Gesture: visible action as utterance. 10.1017/CBO9780511807572714

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  LatoschikM. E.KernF.StauffertJ.-P.BartlA.BotschM.LugrinJ.-L. (2019). Not alone here?! scalability and user experience of embodied ambient crowds in distributed social virtual reality. IEEE Trans. Vis. Comput. Graph.25, 2134–2144. 10.1109/TVCG.2019.2899250

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  LiS.AshokA.ZhangY.XuC.LindqvistJ.GruteserM. (2016). “Whose move is it anyway? Authenticating smart wearable devices using unique head movement patterns,” in 2016 IEEE International Conference on Pervasive Computing and Communications (PerCom), Sydney, Australia, 14-19 March 2016 (IEEE), 1–9. 10.1109/PERCOM.2016.7456514

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  LiebersJ.AbdelazizM.MeckeL.SaadA.AudaJ.GruenefeldU.et al (2021). “Understanding user identification in virtual reality through behavioral biometrics and the effect of body normalization,” in Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, Yokohama Japan (New York, NY: Association for Computing Machinery), 1–11. 10.1145/3411764.3445528

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  LiebersJ.BurschikC.GruenefeldU.SchneegassS. (2023). “Exploring the stability of behavioral biometrics in virtual reality in a remote field study: towards implicit and continuous user identification through body movements: towards implicit and continuous user identification through body movements,” in 29th ACM Symposium on Virtual Reality Software and Technology, Christchurch New Zealand (New York, NY: Association for Computing Machinery), 1–12. 10.1145/3611659.3615696

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  LiebersJ.BrockelS.GruenefeldU.SchneegassS. (2024). Identifying users by their hand tracking data in augmented and virtual reality. Int. J. Human–Computer Interact.40, 409–424. 10.1080/10447318.2022.2120845

  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  McNeillD. (1992). Hand and mind: what gestures reveal about thought univ.

  • Google Scholar
  • View reference in article

• 11

  MerzC.GöttfertC.WienrichC.LatoschikM. E. (2024). “Universal access for social xr across devices: the impact of immersion on the experience in asymmetric virtual collaboration,” in 2024 IEEE Conference Virtual Reality and 3D User Interfaces (VR), Orlando, FL, USA, 16-21 March 2024 (IEEE), 859–869. 10.1109/VR58804.2024.00105738

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  MillerM. R.HerreraF.JunH.LandayJ. A.BailensonJ. N. (2020). Personal identifiability of user tracking data during observation of 360-degree VR video. Sci. Rep.10, 17404. 10.1038/s41598-020-74486-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  MillerR.BanerjeeN. K.BanerjeeS. (2020). “Within-system and cross-system behavior-based biometric authentication in virtual reality,” in 2020 IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW), Atlanta, GA, USA, 22-26 March 2020 (IEEE), 311–316. 10.1109/VRW50115.2020.00070

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  MillerR.BanerjeeN. K.BanerjeeS. (2021). “Using Siamese neural networks to perform cross-system behavioral authentication in virtual reality,” in 2021 IEEE Virtual Reality and 3D User Interfaces (VR), Lisboa, Portugal, 27 March 2021 - 01 April 2021 (IEEE), 140–149. 10.1109/VR50410.2021.00035

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  MooreA. G.DoT. D.RuozziN.McMahanR. P. (2023). “Identifying virtual reality users across domain-specific tasks: a systematic investigation of tracked features for assembly,” in 2023 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), Sydney, Australia, 16-20 October 2023 (IEEE), 396–404. 10.1109/ISMAR59233.2023.00054

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  MooreA. G.DoT. D.ChawlaN. N.IriarteA. J.McMahanR. P. (2024). The full-scale assembly simulation testbed (fast) dataset. 10.48550/arXiv.2403.08969.

  • Google Scholar
  • View reference in article

• 17

  MusgraveK. (2025). PyTorch metric learning original-date: 2019-10-23T17:20:35Z.

  • Google Scholar
  • View reference in article

• 18

  NairV.GuoW.MatternJ.WangR.O’BrienJ. F.RosenbergL.et al (2023a). Unique identification of 50,000+ virtual reality users from head & hand motion data, 895–910.

  • Google Scholar
  • View reference in article

• 19

  NairV.GuoW.WangR.O’BrienJ. F.RosenbergL.SongD. (2023b). Berkeley open extended reality recordings 2023 (BOXRR-23): 4.7 million motion capture recordings from 105,852 extended reality device users. ArXiv:2310.00430. 10.1109/TVCG.2024.3372087.

  • Google Scholar
  • View reference in article

• 20

  NairV.MillerM. R.WangR.HuangB.RackC.LatoschikM. E.et al (2024a). “Effect of data degradation on motion Re-Identification,” in 2024 IEEE 25th International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM), Perth, Australia, 04-07 June 2024 (IEEE), 85–90. 10.1109/WoWMoM60985.2024.00026764

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  NairV.RosenbergL.O’BrienJ. F.SongD. (2024b). Truth in motion: the unprecedented risks and opportunities of extended reality motion data. IEEE Secur. Priv.22, 24–32. 10.1109/MSEC.2023.3330392

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  OBS Project (2024). OBS studio. OBS project. Available online at: https://obsproject.com/.

  • Google Scholar
  • View reference in article

• 23

  RackC.HothoA.LatoschikM. E. (2022). “Comparison of data encodings and machine learning architectures for user identification on arbitrary motion sequences,” in 2022 IEEE International Conference on Artificial Intelligence and Virtual Reality (AIVR), CA, USA, 12-14 December 2022 (IEEE), 11–19. 10.1109/AIVR56993.2022.00010

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  RackC.FernandoT.YalcinM.HothoA.LatoschikM. E. (2023). Who is alyx? A new behavioral biometric dataset for user identification in XR. Front. Virtual Real. 44, 1272234. 10.3389/frvir.2023.1272234

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  RackC.KobsK.FernandoT.HothoA.LatoschikM. E. (2024a). Versatile user identification in extended reality using pretrained similarity-learning ArXiv:2302.

  • Google Scholar
  • View reference in article

• 26

  RackC.NairV.SchachL.FoschumF.RothM.LatoschikM. E. (2024b). “Navigating the kinematic maze: analyzing, standardizing and unifying XR motion datasets,” in 2024 IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW), Orlando, FL, USA, 16-21 March 2024 (IEEE), 507–514. 10.1109/VRW62533.2024.00098

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  RackC.SchachL.AchterF.ShehadaY.LinJ.LatoschikM. E. (2024c). Motion passwords. 10.1145/3641825.3687711

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  RogersC. E.WittA. W.SolomonA. D.VenkatasubramanianK. K. (2015). “An approach for user identification for head-mounted displays,” in Proceedings of the 2015 ACM International Symposium on Wearable Computers - ISWC ’15, Osaka, Japan (New York, NY: Association for Computing Machinery), 143–146. 10.1145/2802083.2808391

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  ThielF. J.SteedA. (2022). Developing an accessibility metric for VR games based on motion data captured under game conditions. Front. Virtual Real.3, 909357. 10.3389/frvir.2022.909357

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  WenE.GuptaC.SasikumarP.BillinghurstM.WilmottJ.SkowE.et al (2024). VR.net: a real-world dataset for virtual reality motion sickness research. IEEE Trans. Vis. Comput. Graph.30, 2330–2336. 10.1109/TVCG.2024.3372044

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article