Split Federated Learning is a system-efficient federated learning paradigm that leverages the rich computing resources at a central server to train model partitions. Data heterogeneity across silos, however, presents a major challenge undermining the convergence speed and accuracy of the global model. This paper introduces Step-wise Momentum Fusion (SMoFi), an effective and lightweight framework that counteracts gradient divergence arising from data heterogeneity by synchronizing the momentum buffers across server-side optimizers. To control gradient divergence over the training process, we design a staleness-aware alignment mechanism that imposes constraints on gradient updates of the server-side submodel at each optimization step. Extensive validations on multiple real-world datasets show that SMoFi consistently improves global model accuracy (up to 7.1%) and convergence speed (up to 10.25x). Furthermore, SMoFi has a greater impact with more clients involved and deeper learning models, making it particularly suitable for model training in resource-constrained contexts.

6G will require on-device antenna systems to operate at ultra-high frequency bands, achieve robust beamforming on the compact user devices, and be blockage-robust. Conventional edge-mounted antennas on devices have limited apertures, suffer from the 'death grip' caused by user-induced blockage, and have poor scalability at mmWave and sub-THz bands. To address these issues, motivated by the rapid evolution of transparent materials and antennas, we propose ScreenAnt in this work–which integrates a transparent antenna array onto the screens of future mobile devices. Specifically, we propose using a transparent on-screen uniform planar array and develop a framework to model its electromagnetic property, spatial configuration, and blockage robustness under realistic user-induced blockage. We also design a gradient-ascent-based algorithm to efficiently optimize power and phase control of on-screen antennas to maximize ScreenAnt's spectral efficiency. Our thorough simulations show that the proposed ScreenAnt can increase the uplink spectral efficiency by over 50

Deep learning (DL) models can now run on microcontrollers (MCUs). Through neural architecture search (NAS), we can search DL models that meet the constraints of MCUs. Among various constraints, energy and latency costs of the model inference are critical metrics. To predict them, existing research relies on coarse proxies such as multiply-accumulations (MACs) and model's input parameters, often resulting in inaccurate predictions or requiring extensive data collection. In this paper, we propose InstMeter, a predictor leveraging MCUs' clock cycles to accurately estimate the energy and latency of DL models. Clock cycles are fundamental metrics reflecting MCU operations, directly determining energy and latency costs. Furthermore, a unique property of our predictor is its strong linearity, allowing it to be simple and accurate. We thoroughly evaluate InstMeter under different scenarios, MCUs, and software settings. Compared with state-of-the-art studies, InstMeter can reduce the energy and latency prediction errors by $3\times$ and $6.5\times$, respectively, while requiring $100\times$ and $10\times$ less training data. In the NAS scenario, InstMeter can fully exploit the energy budget, identifying optimal DL models with higher inference accuracy. We also evaluate InstMeter's generalization performance through various experiments on three ARM MCUs (Cortex-M4, M7, M33) and one RISC-V-based MCU (ESP32-C3), different compilation options (-Os, -O2), GCC versions (v7.3, v10.3), application scenarios (keyword spotting, image recognition), dynamic voltage and frequency scaling, temperatures (21°C, 43°C), and software settings (TFLMv2.4, TFLMvCI). We will open our source codes and the MCU-specific benchmark datasets.

Running deep neural networks on microcontroller units (MCUs) is severely constrained by limited memory resources. While TinyML techniques reduce model size and computation, they often fail in practice due to excessive peak Random Access Memory (RAM) usage during inference, dominated by intermediate activations. As a result, many models remain infeasible on standalone MCUs. In this work, we present a fine-grained split inference system for networked MCUs that enables collaborative inference of Convolutional Neural Networks (CNN) models across multiple devices. Our key insight is that breaking the memory bottleneck requires splitting inference at sub-layer granularity rather than at layer boundaries. We reinterpret pre-trained models to enable kernel-wise and neuron-wise partitioning, and distribute both model parameters and intermediate activations across multiple MCUs. A lightweight, resource-aware coordinator orchestrates the inference across MCU devices with heterogeneous resources. We implement the proposed system on a real testbed and evaluate it on up to 8 MCUs using MobileNetV2, a representative CNN model. Our experimental results show that CNN models infeasible on a single MCU can be executed across networked MCUs, reducing the per-MCU peak RAM usage while maintaining the practical end-to-end inference latency. All the source code of this work can be found here: https://github.com/shashsuresh/split-inference-on-MCUs.

Wireless sensing has gained significant research interest in recent years. However, the focus has primarily been on improving sensing performance, such as enhancing accuracy, and little attention has been paid to the security aspects of wireless sensing. In this paper, we demonstrate that acoustic signal-based sensing, widely regarded as the safest wireless sensing modality due to its physical characteristics, can be stealthily attacked and compromised. The core of this stealthy injection attack lies in exploiting the “photoacoustic effect” to convert the energy of laser light into vibrations, inducing acoustic signals, which are then used to compromise acoustic sensing systems. It is important to note that the laser can be invisible to the human eye, and the acoustic signals it generates are inaudible to human ears. To make the attack even stealthier, we propose the concept of “Smartphone-Defined Laser” to use commodity smartphones to control low-cost laser hardware ($0.60), eliminating the need for bulky and expensive signal generators. Through hardware and software co-design, we successfully demonstrate the attack with smartphones and cheap laser hardware. Comprehensive experiments show that the proposed attack can compromise the state-of-the-art acoustic sensing techniques (both chirp-based and continuous wave-based), achieving a high average success rate of 96.1% across different tasks even when the target is 50 m away. We hope our findings raise awareness of the security risks associated with acoustic sensing and encourage further research into enhancing its security.

Due to the dual functionality of lighting and localization, received signal strength (RSS) based visible light positioning (VLP) has become a promising high-precision indoor positioning method in industrial environments by leveraging pervasively available lighting infrastructure. However, its widespread deployment is hindered by two key challenges: 1) the need for densely sampled fingerprint datasets, and 2) performance degradation due to LED aging in a long-lasting deployment. To address these limitations, we develop a data-efficient VLP method with an online calibration strategy that automatically compensates for gradual light intensity decay caused by LED aging, maintaining robustness without manual re-fingerprinting. Extensive simulations demonstrate that the proposed approach maintains centimeter-level precision across varying LED usage durations. A lightweight on-device model deployed on the Raspberry Pi Pico microcontroller achieves 1.62 cm mean positioning error with 8.39 ms inference latency, demonstrating the practicality of our approach for long-lasting indoor positioning in industrial and IoT scenarios. The source code is publicly available at https://github.com/joeywli/VLP.

Over the past decade, visible light positioning has become increasingly important for precise localization systems, yet its widespread adoption is limited due to the necessity of modifying existing lighting systems. This article presents HueLoc, a novel method that bypasses this issue by using inherent features of light, such as the dominant colors in white light-emitting diode (LED) lights, and employs affordable, energy-efficient hue sensors for location services. We propose that by extracting the power at dominant wavelengths of LEDs, these can be uniquely identified using a specifically designed signature. The unique signatures can be used by mobile objects for spatial awareness and further localization using the proposed regression-based learning approach. Our experiments demonstrate that HueLoc attains a location-mapping accuracy of 100% and achieves decimeter-level localization precision with a moving object in uncontrolled lighting conditions. Moreover, these unique signatures can be combined with other RF-based technologies to enhance their localization accuracy. As an example, this article details the integration of Bluetooth features with light signatures using a three-stage incremental learning approach. The experimental results show that this fusion method significantly improves Bluetooth localization by over 75%, overcoming challenges associated with severe indoor multipath and achieving highly precise location accuracy within decimeters.

The abundant bandwidth in the mmWave band supports high data rates and low latency communication, making it ideal for delivering realistic and seamless virtual reality experiences. However, a key challenge lies in adapting the mmWave beams to the highly dynamic user movements, which often cause beam misalignment, resulting in signal degradation and potential outages. Additionally, maintaining uninterrupted signal reception during beam re-alignment due to head rotation requires low-overhead and timely beam transitions to prevent signal drops caused by delayed switching. This paper addresses these challenges with a joint solution at both the access point (AP) and head-mounted display (HMD) ends. Specifically, the proposed solution integrates coordinated multi-point networks with dual-beam reception at the HMD to enhance diversity, improve channel gain, and mitigate outages caused by user movement. Evaluation using real HMD movement datasets demonstrates that dual-beam reception within a coordinated multi-AP setup achieves up to a 22.8% improvement in reliability by reducing outage rates compared to single-beam reception. Experimental validation further highlights the effectiveness of combining widely distributed APs with a locally distributed subarray configuration on the HMD, improving angular coverage during head rotations. Furthermore, our predictive beam transition approach anticipates the future beam during user movements, preventing received signal degradation from delayed transitions while reducing overhead by 43.8% compared to exhaustive periodic beam searches.

Federated Learning (FL) has emerged as a privacy-preserving framework for training models on data generated at the edge. However, the heterogeneity of data silos (e.g., label skew and domain shift) often leads to inconsistent learning objectives and suboptimal model performance. Inspired by the data-driven approach, we propose Flick, a novel data generation framework for heterogeneous **F**ederated **L**earning w**i**th **C**ommonsense **K**nowledge from Large Language Models (LLMs). In Flick, the client performs the local data summary to capture client-specific knowledge in textual form. The central server then distills task-relevant, high-quality knowledge from the out-of-the-box LLM -- guided by cross-client-specific insights -- to generate informative text prompts. These prompts direct a generative model in producing synthetic data, enabling global model fine-tuning and local data compensation. This process gradually aligns the label and feature distributions across clients. Extensive results on three datasets demonstrate that Flick improves the global model accuracy by up to 11.43\%, and accelerates convergence by up to 12.9$\times$, validating its effectiveness in addressing data heterogeneity.

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.

Machine learning models can now run on microcontrollers. Thanks to the advances in neural architectural search, we can automatically identify tiny machine learning (tinyML) models that satisfy stringent memory and energy requirements. However, existing methods often overlook the energy used during event detection and data gathering. This is critical for devices powered by renewable energy sources like solar power, where energy efficiency is paramount. To address it, we introduce SolarML, a solution designed specifically for solar-powered tinyML platforms, which optimizes the end-to-end system's inference accuracy and energy consumption, from data gathering and processing to model inference. Considering two applications of gesture recognition and keywords spotting, SolarML has the following contributions: 1) a hardware platform with an optimal event detection mechanism that reduces event detection costs by up to 10 x compared to state-of-the-art alternatives; 2) a joint optimization framework eNAS that reduces the energy consumption of the sensor and inference model by up to 2 x, compared to methods that only optimize the inference model. Jointly, they enable SolarML to run end-to-end gesture and audio inference on a battery-free tinyML platform by only harvesting solar energy for 30 and 57 seconds, respectively, in an office environment (500 lux). Source code is available at [1].

This paper presents a novel transparent amplifying intelligent surface (TAIS) architecture for uplink enhancement in indoor-to-outdoor mmWave communications. The TAIS is an amplifier-based transmissive intelligent surface that can refract and amplify the incident signal, instead of only refracting it with adjustable phase shift by most passive reconfigurable intelligent surfaces (RIS). With advanced indium tin oxide film and printing technology, TAIS can be fabricated on the windows without any visual effects. This paper primarily focuses on exploiting the TAIS-based architecture to boost the uplink spectral efficiency (SE) in indoor-to-outdoor mmWave communications. By jointly optimizing the TAIS’s phase shift matrix and transmit power of the user equipment, the uplink SE can be maximized by exploiting the nonlinearity in the TAIS’s amplification process. The key enabler is that we drive the optimal phase shift matrix that maximizes the SE and deduces its closed-form representation. The SE maximization is then proved to be transferred to the transmit power optimization problem. Another important enabler is that we design a low-complexity algorithm to solve the optimization problem using the difference of convex programming. Moreover, the asymptotic spectral efficiency under nonlinear amplification and power scaling law with infinitely large elements under both the sparse and rich scattering channel models are analyzed. Simulation results show that our proposed TAIS can increase the SE by up to 24.7% as compared to its alternative methods.

Federated Learning (FL) has emerged as a privacy-preserving paradigm for collaborative deep learning model training across distributed data silos. Despite its importance, FL faces challenges such as high latency and less effective global models. In this paper, we propose ShuffleFL, an innovative framework stemming from the hierarchical FL, which introduces a user layer between the FL devices and the FL server. ShuffleFL naturally groups devices based on their affiliations, e.g., belonging to the same user, to ease the strict privacy restriction-"data at the FL devices cannot be shared with others", thereby enabling the exchange of local samples among them. The user layer assumes a multi-faceted role, not just aggregating local updates but also coordinating data shuffling within affiliated devices. We formulate this data shuffling as an optimization problem, detailing our objectives to align local data closely with device computing capabilities and to ensure a more balanced data distribution at the intra-user devices. Through extensive experiments using realistic device profiles and five non-IID datasets, we demonstrate that ShuffleFL can improve inference accuracy by 2.81% to 7.85% and speed up the convergence by 4.11x to 36.56x when reaching the target accuracy.

Cardiac patterns are being used to provide hard-to-forge biometric signatures in identification applications. However, this performance is obtained under controlled scenarios where cardiac signals maintain a relatively uniform pattern, facilitating the identification process. In this work, we analyze cardiac signals collected in more realistic (uncontrolled) scenarios and show that their high signal variability makes them harder to obtain stable and distinct features. When faced with these irregular signals, the state-of-the-art (SOTA) reduces its performance significantly. To solve these problems, we propose the CardioID framework 1 with two novel properties. First, we design an adaptive method that achieves stable and distinct features by tailoring the filtering process according to each user’s heart rate. Second, we show that users can have multiple cardiac morphologies, offering us a bigger pool of cardiac signals compared to the SOTA. Considering three uncontrolled datasets, our evaluation shows two main insights. First, while using a PPG sensor with healthy individuals, the SOTA’s balanced accuracy (BAC) reduces from 90–95% to 75–80%, while our method maintains a BAC above 90%. Second, under more challenging conditions (using smartphone cameras or monitoring unhealthy individuals), the SOTA’s BAC reduces to values between 65–75%, and our method increases the BAC to values between 75–85%.

While out-vehicle sensing has achieved great success with the development of vehicle radar and Lidar systems, in-vehicle sensing attracts a lot of attention recently. However, the popular camera-based solutions raise privacy concerns and pose requirement on lighting conditions. Researchers recently utilize wireless signals for sensing. However, besides requiring dedicated hardware, the rich multipath in a small cabin space causes severe interference, degrading the sensing reliability. In this paper, we propose a new sensing modality for in-vehicle sensing, leveraging the leaked EM signals from electric vehicles. The key observation is that the human body can capture the leaked signals, and body motions affect the signal variation patterns. Our solution involves designing conductive cloth tags on the seat to effectively collect body-captured signals and adopting a reference tag to deal with interference. Through extensive experiments conducted over 100 hours, covering a driving distance of 4000 kilometers on various real roads, our system, EVLeSen, can achieve over 90% accuracy in recognizing body motions utilizing just the leaked ambient signals.

Visible light positioning (VLP) based on the received signal strength (RSS) can leverage a dense deployment of LEDs in future lighting infrastructure to provide accurate and energy-efficient indoor positioning. However, its positioning accuracy heavily depends on the density of collected fingerprints, which is labor-intensive. In this work, we propose a data pre-processing method, including data cleaning and data augmentation, to construct reliable and dense fingerprint samples, thereby alleviating the impact of noisy samples as well as reducing labor intensity. Extensive experiments demonstrate that our proposed method achieves an average positioning error of 1.7 cm, utilizing a sparse dataset that reduces the fingerprint collection effort by 98 percent. Running a tinyML-based model for VLP on the Arduino Nano microcontroller, we also show the possibilities for deploying RSS fingerprint-based VLP systems on resource-constrained embedded devices for real-world applications.

Achieving accurate and low-latency spectrum sensing on resource-constrained devices is essential but very difficult. Traditional In-phase and Quadrature (I/Q)-based and the Short-Time Fourier Transform (STFT)-based methods fail to balance the computational overhead and classification accuracy. In this paper, we propose a novel framework -Spectrum Painting (SP)-which enables on-device signal classification with low latency and high accuracy. We design new signal processing methods to compress spectrograms while keeping global signal features and augmenting the salient features of small objects. SP achieves high-accuracy signal classification, assisted further by our proposed Dual-channel Convolutional Neural Network (DualCNN). We collect diverse datasets to evaluate the proposed SP, including synthesized data, and testbed data (from up to 18 commodity devices) obtained from real-world environments in the wild and office settings. Experimental results of SP running on Raspberry Pi 4B show a great reduction in latency up to 20x while maintaining a 95% accuracy. Furthermore, SP demonstrates superior performance within both the centralized learning architecture and the Federated Learning (FL) architecture. For example, the challenging cross-environment evaluation of the SP in the iid-FL scenario yields a substantial accuracy improvement, on average from 24.6% to 83.8%.