Emerging resource-constrained cellular Internet of Things (IoT) applications such as drone swarms, autonomous vehicles, and remote surgery via mixed reality demand millisecond-level time synchronization. Narrow-Band IoT (NB-IoT), the leading low-power wide-area technology, struggles to meet these requirements. The root cause lies in the non-deterministic delays inherent in the 5G protocol design. Uplink reliability and scheduling mechanisms introduce asymmetric latencies that disrupt conventional time synchronization algorithms such as the Network Time Protocol (NTP). Time-critical packets are further affected by deep-sleep wake-up latency, base station scheduling delays, uplink/downlink asymmetry, and unpredictable drift from inexpensive oscillators. Together, these factors can accumulate into timing errors on the order of hundreds of milliseconds. In this paper, we first quantify timing errors across five dimensions on a commercial NB-IoT network. We then present SynchroNB, an on-device framework that combines lightweight machine learning with a cross-layer control loop. SynchroNB forecasts 5G network volatility and crystal drift to adaptively wake the cellular modem, reserves uplink resources just in time, switches into resilience mode when the wireless link degrades, and prioritizes time synchronization packets in the MAC-layer queue. We deploy SynchroNB on commercial NB-IoT hardware and evaluate it over a live 5G network. Our experiments show that SynchroNB achieves single-millisecond-level synchronization accuracy under NB-IoT uplink/downlink asymmetry and, diverse wireless conditions, while requiring only \(36\%\) of the radio-on time and \(25\%\) of the bandwidth of the NTP baseline, transforming NB-IoT time synchronization from a reactive protocol into an intelligent, self-tuning control loop.

Collaborative Mixed Reality (CollabMR) enables multiple users to interact with both the physical and virtual worlds in real time. One of the major challenges CollabMR faces is the timely synchronization of shared virtual content, yet how network Quality of Service (QoS) maps to user Quality of Experience (QoE) remains poorly understood. To address this gap, we present an exploratory, multi-layer framework linking QoS inputs (latency, bandwidth) to system responsiveness and perceptual, cognitive, and behavioral QoE. We evaluate this framework through a controlled within-subject human study with 60 participants (30 pairs) using HoloLens 2 headsets and a collaborative 3D puzzle task under four network conditions. To validate, we analyze network-level measurements, interaction logs, task performance metrics, and post-task subjective questionnaires to examine how variations in system responsiveness manifest across our conceptual model.

Mixed Reality (MR) enables rich, embodied collaboration; however, it is uncertain whether sensor- and system-logged behavioral signals capture how users experience that collaboration. This disconnect stems from a fundamental gap. Behavioral signals are observable and continuous, while collaboration is interpreted subjectively and shaped by internal states like presence, cognitive availability, and social awareness. Our core insight is that sensor signals serve as observable manifestations of subjective experiences in MR collaboration, and they can be captured through sensor data such as shared gaze, speech, spatial movement, and other system-logged performance metrics. We propose the Sensor-to-Subjective (S2S ) Mapping Framework, a conceptual model that links observable interaction patterns to users’ subjective perceptions of collaboration and internal cognitive states through sensor-based indicators and task performance metrics. To evaluate this model, we conducted an exploratory study with 48 participants across 12 MR groups engaged in a collaborative image-sorting task. Our findings show a correlation between sensed behavior and perceived collaboration, particularly through shared attention and proximity.

COMPASS ’25, Toronto, ON, Canada

This paper investigates the security vulnerabilities of prompt-learning-based FL systems in a healthcare setting. Specifically, we use a backdoor attack that leverages learnable prompt vectors in vision-language medical foundation models to execute stealthy adversarial manipulations. We evaluate our attack across diverse healthcare datasets and FL configurations, showing that while FL is useful as a privacy-preserving mechanism, it is susceptible to targeted backdoor attacks that pose a threat to medical applications.

In-flight Wi-Fi provides high-speed Internet connectivity to travelers at 30,000 feet at premium fees. In this paper, we present the first systematic study of the architecture and security policies of in-flight Wi-Fi paywall systems using network tomography analysis. We discover that attackers can exploit the inherent architectural shortcomings of airborne networks to create covert channels and conceal data packets within certain ''always-allowed'' traffic for free Internet access. Moreover, broken device authentication policies in these systems allow unlimited complimentary Internet connectivity. Finally, insecure ARP policies allow attackers to steal paid users' bandwidth to access the free Internet even faster. We validate these issues in practice over two major in-flight Wi-Fi providers using common protocols, e.g., UDP, DNS, etc. We also find that the root causes of these issues stem from different design choices in the architectures of these systems and propose countermeasures to address these flaws and prevent similar attacks.

Distractions in mixed reality (MR) environments can significantly influence user experience, affecting key factors such as presence, reaction time, cognitive load, and Break in Presence (BIP). Presence measures immersion, reaction time captures user responsiveness, cognitive load reflects mental effort, and BIP represents moments when attention shifts from the virtual to the real world, breaking immersion. While prior work has established that distractions impact these factors individually, the relationship between these constructs remains underexplored, particularly in MR environments where users engage with both real and virtual stimuli. To address this gap, we have presented a theoretical model to understand how congruent and incongruent distractions affect all these constructs. We conducted a within-subject study (N = 54) where participants performed image-sorting tasks under different distraction conditions. Our findings show that incongruent distractions significantly increase cognitive load, slow reaction times, and elevate BIP frequency, with presence mediating these effects.

Autonomous robots, autonomous vehicles, and humans wearing mixed-reality headsets require accurate and reliable tracking services for safety-critical applications in dynamically changing real-world environments. However, the existing tracking approaches, such as Simultaneous Localization and Mapping (SLAM), do not adapt well to environmental changes and boundary conditions despite extensive manual tuning. On the other hand, while deep learning-based approaches can better adapt to environmental changes, they typically demand substantial data for training and often lack flexibility in adapting to new domains. To solve this problem, we propose leveraging the neurosymbolic program synthesis approach to construct adaptable SLAM pipelines that integrate the domain knowledge from traditional SLAM approaches while leveraging data to learn complex relationships. While the approach can synthesize end-to-end SLAM pipelines, we focus on synthesizing the feature extraction module. We first devise a domain-specific language (DSL) that can encapsulate domain knowledge on the essential attributes for feature extraction and the real-world performance of various feature extractors. Our neurosymbolic architecture then undertakes adaptive feature extraction, optimizing parameters via learning while employing symbolic reasoning to select the most suitable feature extractor. Our evaluations demonstrate that our approach, neurosymbolic Feature EXtraction (nFEX), yields higher-quality features. It also reduces the pose error observed for the state-of-the-art baseline feature extractors ORB and SIFT by up to 90% and up to 66%, respectively, thereby enhancing the system’s efficiency and adaptability to novel environments.

Data heterogeneity among Federated Learning (FL) users poses a significant challenge, resulting in reduced global model performance. The community has designed various techniques to tackle this issue, among which Knowledge Distillation (KD)-based techniques are common. While these techniques effectively improve performance under high heterogeneity, they inadvertently cause higher accuracy degradation under model poisoning attacks (known as \emph{attack amplification}). This paper presents a case study to reveal this critical vulnerability in KD-based FL systems. We show why KD causes this issue through empirical evidence and use it as motivation to design a hybrid distillation technique. We introduce a novel algorithm, Hybrid Knowledge Distillation for Robust and Accurate FL (HYDRA-FL), which reduces the impact of attacks in attack scenarios by offloading some of the KD loss to a shallow layer via an auxiliary classifier. We model HYDRA-FL as a generic framework and adapt it to two KD-based FL algorithms, FedNTD and MOON. Using these two as case studies, we demonstrate that our technique outperforms baselines in attack settings while maintaining comparable performance in benign settings.

Measuring presence is critical to improving user involvement and performance in Mixed Reality (MR). Presence, a crucial aspect of MR, is traditionally gauged using subjective questionnaires, leading to a lack of time-varying responses and susceptibility to user bias. Inspired by the existing literature on the relationship between presence and human performance, the proposed methodology systematically measures a user's reaction time to a visual stimulus as they interact within a manipulated MR environment. We explore the user reaction time as a quantity that can be easily measured using the systemic tools available in modern MR devices. We conducted an exploratory study (N = 40) with two experiments designed to alter the users' sense of presence by manipulating place illusion and plausibility illusion. We found a significant correlation between presence scores and reaction times with a correlation coefficient -0.65, suggesting that users with a higher sense of presence responded more swiftly to stimuli. We develop a model that estimates a user's presence level using the reaction time values with high accuracy of up to 80%. While our study suggests that reaction time can be used as a measure of presence, further investigation is needed to improve the accuracy of the model.

A prerequisite to improving the presence of a user in mixed reality (MR) is the ability to measure and quantify presence. Traditionally, subjective questionnaires have been used to assess the level of presence. However, recent studies have shown that presence is correlated with objective and systemic human performance measures such as reaction time. These studies analyze the correlation between presence and reaction time when technical factors such as object realism and plausibility of the object's behavior change. However, additional psychological and physiological human factors can also impact presence. It is unclear if presence can be mapped to and correlated with reaction time when human factors such as conditioning are involved. To answer this question, we conducted an exploratory study ($N=60$) where the relationship between presence and reaction time was assessed under three different conditioning scenarios: control, positive, and negative. We demonstrated that human factors impact presence. We found that presence scores and reaction times are significantly correlated (correlation coefficient of -0.64), suggesting that the impact of human factors on reaction time correlates with its effect on presence. In demonstrating that, our study takes another important step toward using objective and systemic measures like reaction time as a presence measure.

There is a lack of datasets for visual-inertial odometry applications in Extended Reality (XR). To the best of our knowledge, there is no dataset available that is captured from an XR headset with a human as a carrier. To bridge this gap, we present a novel pose estimation dataset --- called HoloSet --- collected using Microsoft Hololens 2, which is a state-of-the-art head mounted device for XR. Potential applications for HoloSet include visual-inertial odometry, simultaneous localization and mapping (SLAM), and additional applications in XR that leverage visual-inertial data.

Critical machine learning applications (medical image guidance, task prediction, anomaly detection) require large amounts of data that could not be sufficiently supplied from a single entity, so multiple edge devices collaboratively train their collected data. But this raises privacy and overhead concerns. Federated learning (FL) can be a promising solution to enable these applications while preserving data privacy and mitigating communication overhead. However, an FL model originating from edge deployments with heterogeneous resources may be biased towards a set of devices. We observe that existing bias mitigation techniques in FL focus mainly on the bias that originates from label heterogeneity (due to the skewed distribution of data). We argue that sample feature heterogeneity due to different feature representations at devices is a major contributor to bias in FL. In this paper, we present an analysis of the bias that arises from sampling feature heterogeneity, and analyze the potential of existing performance enhancing techniques (normalization) to overcome bias. Our results demonstrate that normalization techniques do not eliminate bias and motivate the need for dedicated bias mitigation techniques in FL.

This paper exhaustively explores the threat landscape of coordinated spatiotemporal attacks in mixed reality systems. Novel devicelevel and cross-device time translation and spatial shift attacks are launched, and their impact on deep learning based sensor fusion is evaluated. A major focus of this work is to establish stealthiness in the presence of sophisticated security mechanisms with an added constraint that mixed reality systems allow minimal time durations for covert operation. The efficacy of proposed attacks is evaluated through a preliminary study on inertial and visual data streams.

Emerging countries predominantly rely on room-level air conditioning units (window ACs, space heaters, ceiling fans) for thermal comfort. These distributed units have manual, decentralized control leading to suboptimal energy usage for two reasons: excessive setpoints by individuals and inability to interleave different conditioning units for energy savings. We propose a novel inverted HVAC approach: cheaply retrofitting these distributed units with “on-off” control and providing centralized control augmented with room and environmental sensors. Our binary control approach exploits an understanding of device consumption characteristics and factors this into the control algorithms to reduce consumption. We implement this approach as H awadaar in a prototype 180ft 2 room to evaluate its efficacy over a 7-month period experiencing both hot and cold climates. Through a post analysis, we show that our on-off algorithms are not far from a theoretically optimal approach based on a priori information that precisely knows the optimal control points to minimize consumption. We collect enough evidence to plausibly scale our empirical evaluation, demonstrating countrywide benefits: with just 20% market penetration, H awadaar can save up to 6% of electricity per capita in residential and commercial sectors—resulting in a substantial countrywide impact.

Wireless energy transfer has recently emerged as a promising alternative to realize the vision of perpetual embedded sensing. However, this technology transforms the notion of energy from merely a node’s local commodity to, similarly to data, a deployment-wide shareable resource. The challenges of managing a shareable energy resource are much more complicated and radically different from the research of the past decade: Besides energy-efficient operation of individual devices, we also need to optimize networkwide energy distribution. To counteract these challenges, we propose an energy stack , a layered software model for energy management in future transiently powered embedded networks. An initial specification of the energy stack, which is based on the historically successful layered approach for data networking, consists of three layers: (i) the transfer layer, which deals with the physical transfer of energy; (ii) the scheduling layer, which optimizes energy distribution over a single hop; and (iii) the network layer, creates a global view of the energy in the network for optimizing its networkwide distribution. As a contribution, we define the interfacing APIs between these layers, delineate their responsibilities, identify corresponding challenges, and provide a first implementation of the energy stack. Our evaluation, using both experimental deployments and high-level simulations, establishes the feasibility of a layered solution to energy management under transient power.