MobiCom 2025 Papers

Efficient on-device neural network (NN) inference offers predictable latency, improved privacy and reliability, and lower operating costs for vendors than cloud-based inference. This has sparked recent development of microcontroller-scale NN accelerators, also known as neural processing units (μNPUs), designed specifically for ultra-low-power applications.

Ultrasonic microphone jammers (UMJs) use ultrasonic waves to interfere with concealed microphone recorders, offering significant values in confidential meetings and secret talks. Existing UMJs, however, performs in a short range of typically 2 m, which remains a huge gap to practical applications. The key challenge lies in the fact that the ultrasonic signal will produce audible sounds during its transmission, caused by nonlinear distortion of power amplifiers and loudspeakers in the transmission chain of the UMJ. In this paper, we propose Hedgehog, a room-scale ultrasonic jammer design, which enhances the jamming range through two key methods. First, it corrects nonlinear distortion by modeling the transmission chain and applying digital pre-distortion. Furthermore, it redesigns the jamming signal to improve its effectiveness by not only reducing the signal-to-noise ratio (SNR) but also by suppressing the semantic information in human speech. The experimental results show that Hedgehog achieved a word error rate (WER) of over 95% at 8 meters and over 80% at 10 meters, which is a 4 times increase in jamming distance compared to existing solutions.

We propose FocusX, the first mobile-deployable system achieving artifact-free all-in-focus synthesis in dynamic scenes. Our approach introduces three key innovations: 1) For focal stack acquisition, our depth prior-based dynamic focusing method that adaptively selects focus distances using real-time scene depth distribution analysis and depth-of-field constrained spatial clustering, reducing redundant captures while ensuring full depth coverage; 2) To reduce pixel misalignment caused by lens breathing, we adopt a one-time offline calibration to map the relationship between field-of-view and focus distance, aligning the images by cropping accordingly; 3) We design the Diff-MotionAIFNet, a conditional diffusion-based model that decouples moving-static components for artifact-free AIF reconstruction in dynamic scene while preserving scene fidelity. We further contribute DynaAIFSet, containing 5,500 dynamic scenes (120K images) for training and evaluation. Experiments show FocusX achieves state-of-the-art performance, outperforming baselines up by 59.6% in SSIM and 49.1% in PSNR, respectively. The deployment latency of FocusX is 4.8s on Honor Magic7 Pro. This work bridges computational photography theory with mobile implementation constraints, delivering practical AIF enhancement for user-generated content.

The mixture of experts (MoE) model is a sparse variant of large language models (LLMs), designed to hold a better balance between intelligent capability and computational overhead. Despite its benefits, MoE is still too expensive to deploy on resource-constrained edge devices, especially with the demands of on-device inference services. Recent research efforts often apply model compression techniques, such as quantization, pruning and merging, to restrict MoE complexity. Unfortunately, due to their predefined static model optimization strategies, they cannot always achieve the desired quality-overhead trade-off when handling multiple requests, finally degrading the on-device quality of service. These limitations motivate us to propose the D2MoE, an algorithm-system co-design framework that matches diverse task requirements by dynamically allocating the most proper bit-width to each expert. Specifically, inspired by the nested structure of matryoshka dolls, we propose the matryoshka weight quantization (MWQ) to progressively compress expert weights in a bit-nested manner and reduce the required runtime memory. On top of it, we further optimize the I/O-computation pipeline and design a heuristic scheduling algorithm following our hottest-expert-bit-first (HEBF) principle, which maximizes the expert parallelism between I/O and computation queue under constrained memory budgets, thus significantly reducing the idle temporal bubbles waiting for the experts to load. Evaluations on real edge devices show that D2MoE improves the overall inference throughput by up to 1.39× and reduces the peak memory footprint by up to 53% over the latest on-device inference frameworks, while still preserving comparable serving accuracy as its INT8 counterparts.

Large language models (LLMs) have emerged as a cornerstone for advancing AI technologies. It revolutionizes the way we interact with devices, websites, and information, and paves the way for the development of highly intuitive and capable virtual assistants. Training of today's LLMs happens in cloud data centers due to the requirement of enormous data and a significant amount of computing power. Despite extensive research in mobile edge computing, fine-tuning pre-trained LLMs using resource-constrained devices like commodity smartphones remains highly under-explored. In this paper, we propose Confidant, a practical collaborative training framework that allows modern LLMs to be fine-tuned across multiple off-the-shelf mobile devices. To this end, Confidant partitions an LLM into several sub-models, allowing each of them to fit in the memory of a mobile device. Multiple mobile devices then collaborate to train the LLM by employing a novel pipeline parallel training approach. In specific, Confidant encompasses a memory-aware dynamic model partitioning and intra-device multi-processor scheduler to minimize the training time across heterogeneous platforms. To ensure resilient distributed training, a hybrid fault tolerance mechanism is devised to proactively manage potential device and network failures. We fully implemented Confidant in C++/Python, and built a cross-framework adapter, enabling collaborative training on a variety of mobile platforms. Experimental results show that Confidant excels in achieving computation-, memory-efficient, and robust customization of LLMs - it manages to train state-of-the-art billion-sized LLMs including BERT, GPT-2, Phi2, and LLaMA3, and fine-tunes Phi2-2.7B on Alpaca in just 40.1 hours using three consumer-grade mobile devices.

Following the broad prospect of Radio Frequency Identification (RFID) technology is the security concern of unauthorized tag identification, which poses threats to the privacy of both objects and users. In this paper, we propose RF-Rock, the first RFID unauthorized identification attack that operates without tag activation, thereby evading almost all existing defenses. This attack exposes the vulnerabilities of current RFID networks in identification legitimacy and privacy. It is based on the intermodulation effect originating from intrinsic nonlinearity within tag circuits. To this end, we explore the distinctness and consistency of the intermodulation-based physical layer fingerprint of RFID tags with theoretical analysis and empirical validation, and optimize the attack accuracy and efficiency with delicate excitation plan. Real-world experiments show that RF-Rock achieves an attack success rate of 93.2% on average under various conditions. The entropy of our proposed fingerprint is 15.5 bits and implies sufficient capacity in practical attacks.

The recent proliferation of spacecraft in Earth's orbits has ushered in the rise of large-scale satellite constellations. However, this unprecedented growth of constellations has introduced a previously unforeseen challenge: ground track congestion. Specifically, the increasing density of orbital slots forces satellites to share similar orbit planes, causing their nadir-point projections on Earth's surface (i.e., ground tracks) to overlap or remain in close proximity within short time intervals. Such orbit-endowed ground track congestion can degrade constellation performance in remote sensing operations, specified by limited constellation coverage, redundant satellite count, and delayed data delivery.

As the adoption of LiDAR expands across various fields such as autonomous driving, robotics, and smart cities, the demand for adaptive scanning capabilities to better capture dynamic and complex scenes becomes paramount. Current LiDAR technologies, limited by fixed uniform scan patterns, struggle to prioritize critical areas, resulting in reduced perception accuracy and performance inefficiencies. This paper introduces SmartLiDAR, an advanced LiDAR system that enhances scanning efficiency and performance by adaptively optimizing scan focus through an intelligent, software-defined micro-mirror controller. Unlike traditional systems, SmartLiDAR dynamically adjusts its scan pattern based on environmental characteristics and application-specific requirements, concentrating sample points on key objects without increasing power consumption or scan time. SmartLiDAR achieves this by integrating a novel quadratic micro-mirror controller, an adaptive algorithm for generating fine-grained attention map with prioritized scan focus, and a carefully designed optimization algorithm that maps attention maps to practical scanning patterns. We prototype SmartLiDAR by building a software-defined LiDAR using commercially available optical components and FPGA. Our experimental results demonstrate that SmartLiDAR significantly enhances resolution in regions of interest by 3x and increases average object detection precision by up to 16.11%. Additionally, SmartLiDAR maintains negligible extra energy consumption and processing latency, making it suitable for real-time applications, such as autonomous vehicles.

Large Foundation Models (LFMs) have unlocked new possibilities in human-computer interaction, particularly with the rise of mobile Graphical User Interface (GUI) Agents capable of interacting with mobile GUIs. These agents allow users to automate complex mobile tasks through simple natural language instructions. However, the inherent probabilistic nature of LFMs, coupled with the ambiguity and context-dependence of mobile tasks, makes LFM-based automation unreliable and prone to errors. To address this critical challenge, we introduce VeriSafe Agent (VSA)1: a formal verification system that serves as a logically grounded safeguard for Mobile GUI Agents. VSA deterministically ensures that an agent's actions strictly align with user intent before executing the action. At its core, VSA introduces a novel autoformalization technique that translates natural language user instructions into a formally verifiable specification. This enables runtime, rule-based verification of agent's actions, detecting erroneous actions even before they take effect. To the best of our knowledge, VSA is the first attempt to bring the rigor of formal verification to GUI agents, bridging the gap between LFM-driven actions and formal software verification. We implement VSA using off-the-shelf LFM services (GPT-4o) and evaluate its performance on 300 user instructions across 18 widely used mobile apps. The results demonstrate that VSA achieves 94.33%–98.33% accuracy in verifying agent actions, outperforming existing LFM-based verification methods by 30.00%–16.33%, and increases the GUI agent's task completion rate by 90%–130%.

The accurate localization and tracking of dynamic targets, such as equipment, people, vehicles, drones, robots, and the assets that they interact with in GPS-denied indoor environments is critical to enabling safe and efficient operations in the next generation of spatially-aware industrial facilities. This paper presents DragonFly, a 3D localization system of highly dynamic backscatter tags using a single MIMO mmWave radar. The system delivers the first demonstration of a mmWave backscatter system capable of exploiting the capabilities of MIMO radars for the 3D localization of mmID tags moving at high speeds and accelerations at long ranges by introducing a critical Doppler disambiguation algorithm and a fully-integrated cross-polarized dielectric-lens-based mmID tag consuming a mere 68 μW. DragonFly was extensively evaluated in static and dynamic configurations, including on a flying quadcopter, and benchmarked against multiple baselines, demonstrating its ability to track the positions of multiple tags with a median 3D accuracy of 12 cm at speeds and acceleration on the order of 10 m s–1 and 4 m s–2 and at ranges of up to 50 m.

The Fifth Generation (5G) core network is designed as a set of Virtual Network Functions (VNFs) hosted on Commercial-Off-the-Shelf (COTS) hardware. This creates a growing demand for general-purpose computing resources. Given their elastic infrastructure, cloud services like Amazon Web Services (AWS) are attractive platforms to address this need. Therefore, it is crucial to understand the Quality of Service (QoS) requirements associated with deploying the 5G core in the cloud. We developed the 5G-MAP (5G Measurement and Assessment Platform) to understand the trade-offs between different deployment strategies. Our framework facilitates detailed control and user plane performance assessments in varied deployment scenarios. We integrated 5G-MAP with the OpenAirInterface (OAI) 5G core and utilized it in a series of deployments across seven countries, leveraging eight AWS regions and eighteen edge zones. Our evaluations cover from HTTP transactions to user plane throughput and packet loss. We identify topologies that can considerably lower the 5G core service chain latencies due to a significant reduction in the number of inter-site hops. Such actionable performance improvements illustrate how operators can leverage 5G-MAP to optimize their cloud-based 5G deployments.

Quantum Annealing (QA)-accelerated MIMO detection is an emerging research approach in the context of NextG wireless networks. The opportunity is to enable large MIMO systems and thus improve wireless performance. The approach aims to leverage QA to expedite the computation required for theoretically optimal but computationally-demanding Maximum Likelihood detection to overcome the limitations of the currently deployed linear detectors. This paper presents X-ResQ, a QA-based MIMO detector system featuring flexible parallelism that is uniquely enabled by quantum Reverse Annealing (RA). Unlike prior designs, X-ResQ has many desirable parallel QA system properties and has effectively improved detection performance as more qubits are assigned. In our evaluations on a state-of-the-art quantum annealer, fully parallel X-ResQ achieves near-optimal throughput for 4 × 6 MIMO with 16-QAM using approx. 240 qubits achieving 2.5–5× gains compared against other classical and quantum detectors. We also implement and evaluate X-ResQ in the non-quantum digital setting for more comprehensive evaluations. This classical X-ResQ showcases the potential to realize ultra-large 1024 × 1024 MIMO, significantly outperforming other MIMO detectors, including the state-of-the-art RA detector classically implemented in the same way.

In this paper, we design and fabricate cost-effective, electronics-free millimeter-wave (mmWave) fiducial markers that support clutter-free detection on radars. Fiducial markers are widely used in camera-based systems to provide spatial information in robotic navigation applications. QR-code-like fiducial markers can be readily printed on paper - low-cost, easy to produce, and electronics-free. Yet, an analogous mmWave solution is yet to appear, which stems from the unique challenge in the mmWave context: its vulnerability to clutter. Existing solutions either trade hardware simplicity for clutter resilience or stay simple but remain vulnerable to environmental multipath. This paper seeks to solve this dilemma.

Hidden spy cameras are a growing global threat to personal privacy. With the emergence of translucent screen technology, a new security risk has arisen: cameras can now hide behind devices' screens like TVs and monitors that are common in private places, e.g., hotel rooms. The screen's covering over the hidden camera not only makes the cameras behind it unnoticeable to human eyes but also makes existing camera detection methods less effective. Inspired by recent advances in representing real-world scenes accurately using neural networks, we propose Neural Infrared Reflectance Field (NIRF) to learn the intricate optical properties of the screen and the cameras hidden behind it. Through NIRF, we design a new camera detection system by leveraging the unique reflective properties of behind-screen cameras and screens. We evaluate NIRF with thorough experiments on five smartphones. Our NIRF archives over 90% detection rate and is robust to different conditions, including varied backgrounds, ambient light levels, screen protectors, and screen contents. Besides, we conduct a field study by deploying 18 common spy cameras behind a 65-inch translucent TV and recruiting 27 people to compare NIRF with commercial hidden camera detectors. NIRF achieves an 89.5% detection rate, significantly outperforming the best commercial hidden camera detector that only has a 14.4% detection rate of behind-screen cameras.

Various wireless technologies have been utilized for sensing. Although promising, these wireless sensing technologies have inherent limitations. Prior research mainly focuses on overcoming the limitations of an individual wireless sensing technology, and little attention has been paid to the potential benefits of sensing with more than one wireless technology. In this paper, we introduce the concept of cross-technology sensing for the first time, and propose LoFiSen to enable LoRa-to-WiFi sensing. LoFiSen leverages the strengths of both LoRa and WiFi—combining LoRa's long-range capability with WiFi's pervasiveness. The chirp characteristic of LoRa signal significantly improves the sensing range of WiFi, and the widespread availability of WiFi devices makes LoRa sensing more pervasive. LoFiSen is fully compatible with LoRa and WiFi protocols, and can work on commodity LoRa and WiFi hardware. The key component of our design is enabling the WiFi receiver to capture fine-grained LoRa signal variations for sensing. Real-world experiments demonstrate that LoFiSen improves the WiFi sensing range for respiration monitoring from 8 m to 41 m, and pushes the walking sensing range from 16 m to 73.5 m. Through-wall passive respiration monitoring, previously infeasible with state-of-the-art WiFi sensing, is now possible with LoFiSen.

Localization is a critical task for underwater robots, yet today's underwater localization systems are limited by their accuracy, scalability, and/or energy consumption (i.e., longevity). We present the design, implementation, and evaluation of EchoBLUE- an accurate, scalable, and low-power localization system for underwater robots.

Radio frequency (RF) techniques have shown promise for continuous contactless healthcare applications. However, real-world indoor environments pose challenges for existing systems, which may struggle to detect subtle physiological signals. This paper proposes Medusa, a novel wireless vital-sign sensing system designed for multi-view setups. It enables users to deploy distributed Multiple Input Multiple Output (MIMO) arrays into their daily living environments, facilitating vital-sign sensing in real-world settings. Unlike most existing single Commercial Off-The-Shelf (COTS) radar-based systems that operate under controlled settings Medusa's primary novelty lies in the design of a first-of-its-kind flexible multi-view vital sign sensing system that is view-agnostic, pose-agnostic, contactless, and can sense basic human vitals with good accuracy. Through our well-engineered hardware and software co-design, Medusa enables real-time processing of large distributed MIMO arrays, while balancing the tradeoff between Signal-to-Noise Ratio (SNR) and spatial diversity gain across each of its four distributed 4 × 4 sub-arrays for increased robustness. This is achieved using our novel unsupervised learning model which effectively recovers vital sign waveforms by decomposing the received signals. Extensive evaluations with 21 participants demonstrate Medusa's spatial diversity gain for real-world vital-sign monitoring, enabling free movement and orientation of subjects in both familiar and unfamiliar indoor environments.

Human activity monitoring in the water is essential for pool management and drowning prevention. Existing camera-based solutions pose significant concerns about privacy and extra installation costs. Although sonars have been widely used for underwater sensing in open aquatic environments such as oceans and lakes, monitoring human activities with sonars in a pool setup is challenging. In this work, we propose AquaScan, the first scanning sonar-based underwater sensing system for human activity monitoring. To overcome the low frame rate, we propose a novel scanning strategy and apply an image reconstruction method to accelerate the scanning speed without compromising the performance of motion detection. We develop a novel signal processing pipeline based on a physical model to remove noises and localize human subjects. We extract features like motion, time, and spatial information from sonar images and develop a state-transfer-based activity recognition system to recognize five common water activities. We deployed AquaScan on three public swimming pools for a total period of 94 hours. The evaluation results show that AquaScan can successfully recognize the five activities in the water at about 91.5%.

Bluetooth Low Energy (BLE) and LoRa are two dominant wireless protocols for the Internet of Things (IoT), each built with specific design goals, rendering them non-interoperable. Cross-Technology Communication (CTC) enables heterogeneous communication between BLE and LoRa. Despite these efforts, existing works suffer from low throughput and short communication range. In this paper, we present a novel system named BoRa, which enables bi-directional communication between the COTS BLE and COTS LoRa chips. The key idea of BoRa to emulate the linear frequency changes of the LoRa chirp is continuously tuning its Modulation Offset (MO), which is originally designed to compensate for the Carrier Frequency Offset (CFO) in BLE chips. BoRa can be readily run on COTS chips without any hardware modifications.