論文 - 大槻　知明

A Plug-and-Play Module for Enhancing Fault-Tolerant Distributed Inference Based on Gaussian Dropout

Hou Z., Ohtsuki T.

IEEE International Conference on Communications    4553 - 4558 2025年

ISSN  15503607

　概要を見る

Deploying deep neural network (DNN) models to the Internet of Things (IoT) for distributed inference (DI) is becoming an increasingly common requirement for AI applications. When IoT transmissions occur over lossy networks, data loss can significantly affect the accuracy of inference tasks. Therefore, we propose a plug-and-play module called Loss-Adapter, which can improve the accuracy of DI on lossy networks. To simulate network packet loss, we design a Gaussian distribution sampling dropout. This idea is utilized during the training of Loss-Adapter to acquire fault-tolerant inference capability on unreliable communication links. Moreover, we use the Tree-structured Parzen Estimator algorithm to optimize the parameters of the Gaussian dropout layer. To avoiding catastrophic forgetting, we implement freeze training. In addition, as a plug-and-play module, it is very easy to train and use. Our experiments show that by applying the Loss-Adapter, the inference accuracy of DNN models on lossy IoT is significantly improved.

• Access to Document (DOI)

A Novel Physical Spoofing Technique Using Radio Frequency Fingerprint Emulation and Model Fitting

Yao Z., Wang Y., Gui G., Otsuki T., Mao S., Wang X., Sari H.

IEEE International Conference on Communications    2629 - 2634 2025年

ISSN  15503607

　概要を見る

With the increasing demand for secure communication in 5G and beyond, authentication of wireless devices has become a crucial task for communication security. Radio frequency fingerprint identification (RFFI) leverages the hardware-specific features in radio frequency (RF) signals, known as radio frequency fingerprints (RFF), to achieve highprecision device identification. However, the dependence of RFFI on the physical characteristics of devices makes it vulnerable to physical spoofing attacks. This paper proposes an innovative physical spoofing attack framework that combines spoofed transmitter and legitimate transmitter models. It performs RFF modeling, RFF concealment (RFFC), and RFF spoofing (RFFS) sequentially to achieve precise spoofing of the original baseband signal. We validate the effectiveness of the proposed physical spoofing mechanism through simulations of seven types of transmitters using MATLAB Simulink. The performance is further evaluated on an RFFI model based on complexvalued convolutional neural networks (CVCNN). Experimental results show that neural networks (NN) significantly outperform the generalized memory polynomial (GMP) model in nonlinear data fitting and temporal relationship modeling. Consequently, NN-based physical spoofing methods exhibit superior attack effectiveness. Specifically, under the signal-to-noise ratio (SNR) of 15 dB, the NN-based physical spoofing method achieves a target attack success rate (TSR) as high as 98%, which is superior to adversarial attack methods. NN-based methods also enhanced performance in terms of stealthiness metrics.

• Access to Document (DOI)

A Novel MIMO FMCW Radar-Based Approach for Heart Rate Estimation Using Positional Feature Selection

Nakatani S., Bouazizi M., Ohtsuki T.

IEEE International Conference on Communications    916 - 921 2025年

ISSN  15503607

　概要を見る

This paper proposes a heart rate estimation method using Variational Mode Decomposition (VMD) employing Multiple-Input Multiple-Output (MIMO) Frequency Modulated Continuous Wave (FMCW) radar. The proposed method first estimates the human position within the radar's coverage area, then reduces noise by focusing on the signal from these positions. The signal is decomposed into Intrinsic Mode Function (IMF) signals using VMD, and only the IMF signal related to the heartbeat is retained. Heart rate signal is reconstructed by weighting IMF signals based on their energy within the specific spatial area in which the human is located. The reconstructed signals are utilized for heart rate estimation through peak detection. The estimation is done over consecutive time windows. Until the fourth time window, among the estimated heart rates from the different cells, the selection is based on energy and periodicity. From the fifth time window onwards, the heart rate closest to the average of the previous four estimates is chosen. This approach considers the gradual transition of heart rate estimation over time to mitigate extraneous variations. Validation experiments involved 4 subjects being seated and stationary within the radar coverage area, with heart rate observed via MIMO FMCW radar. Results showed an average Mean Absolute Error (MAE) of 2.54 BPM, with an exclusion rate of 2.12%.

• Access to Document (DOI)

A Deep Transfer Learning-Based Low-Overhead Beam Prediction in Vehicle Communications

Xiao Z., Cao Y., Bouazizi M., Ohtsuki T., Mumtaz S.

IEEE Wireless Communications Letters 14 （ 12 ） 4147 - 4151 2025年

ISSN  21622337

　概要を見る

Existing transfer learning-based beam prediction approaches primarily rely on simple fine-tuning. When there is a significant difference in data distribution between the target domain and the source domain, simple fine-tuning limits the model’s performance in the target domain. To tackle this problem, we propose a transfer learning-based beam prediction method that combines fine-tuning with domain adaptation. We integrate a domain classifier into fine-tuning the pre-trained model. The model extracts domain-invariant features in adversarial training with domain classifier, which can enhance model performance in the target domain. Simulation results demonstrate that the proposed transfer learning-based beam prediction method achieves better achievable rate performance than the pure fine-tuning method in the target domain, and close to those when the training is done from scratch on the target domain.

• Access to Document (DOI)

Malware Traffic Classification via Expandable Class Incremental Learning With Architecture Search

Zhang X., Wang Y., Ohtsuki T., Gui G., Yuen C., Di Renzo M., Sari H.

IEEE Transactions on Information Forensics and Security 20   6074 - 6085 2025年

ISSN  15566013

　概要を見る

Malware traffic classification (MTC) is a crucial step in network intrusion detection, which is significant for network security and management. With the continuous evolution of malware traffic, traditional MTC methods are difficult to adapt efficiently to new traffic categories, and manually designed neural network structures suffer from performance bottlenecks and low design efficiency. Hence, we propose an enhanced MTC method based on expandable class incremental learning (CIL) with architecture search. The architecture search can automatically design the optimal neural network structure tailored to different network traffic characteristics, avoiding the limitations of manually designing network structures and improving classification performance. Meanwhile, expandable CIL allows the MTC model to gradually learn new traffic categories without forgetting previous knowledge, avoiding the computational overhead and efficiency loss caused by frequent retraining of the model. The experimental results demonstrate that the proposed CIL-MTC approach surpasses advanced incremental learning methods on both the Edge-IIoTset and ISCX VPN-nonVPN datasets, achieving superior classification performance while maintaining lower average trainable parameters and training costs. Especially, it achieves an average incremental accuracy of 98.55% and 99.09% on the Edge-IIoTset dataset with incremental tasks of 5 and 2, respectively.

• Access to Document (DOI)

Model-Based DRL for Dynamic Antenna Beamforming in HAPS Communication System

Yang S., Bouazizi M., Ohtsuki T.

2025 International Symposium on Antennas and Propagation Isap 2025 2025年

　概要を見る

High Altitude Platform Station (HAPS) systems require dynamic beamforming to handle platform mobility and uneven user distributions. However, traditional reinforcement learning (RL) and evolutionary algorithms run the risk of degrading throughput during exploration. We propose a model-based deep RL (DRL) framework that pre-trains an environment model to map antenna parameters and HAPS states to SINR distributions. This simulated feedback allows efficient parameter exploration without reducing average throughput below its initial level. Simulation results show that our method consistently maintains higher throughput than conventional RL approaches and swiftly adapts to network changes.

• Access to Document (DOI)

User Clustering and Trajectory Optimization for UAV-Assisted Wireless Sensor Network

Zhang X., Yin Y., Ohtsuki T.

2025 30th Asia Pacific Conference on Communications Apcc 2025 2025年

　概要を見る

With the advances in flexibility, wide coverage, and cost-effectiveness, integrating unmanned aerial vehicles (UAVs) with wireless sensor networks (WSNs) has been regarded as a promising technology in data collection and transmission within complex environments. Due to the limited battery capacity of sensor nodes (SNs) and UAVs, UAV-assisted WSNs face challenges in achieving energy efficiency and optimizing resource allocation. To reduce the total energy consumption of UAV-assisted WSNs, we design a multifactor fitness function which includes the distance between SNs, energy levels of SNs, and UAV trajectory length. Based on that, we propose a balanced clustering and cluster head (CH) selection strategy. As for UAV trajectory design, we employ the branch and bound method according to the selection of CHs. Simulation results show that the proposed algorithm reduces the average energy consumption of all SNs and achieves short UAV trajectory length compared to Low-Energy Adaptive Clustering Hierarchy (LEACH), Hybrid Energy-Efficient Distributed clustering (HEED), Particle Swarm Optimization (PSO), and Stable Election Protocol (SEP).

• Access to Document (DOI)

Towards Robust and Real-Time State Estimation: A Neural Kalman Filter Framework with Mamba Architecture

Sun J., Shen J., Tong J., Tang T., Yin Y., Ohtsuki T., Gui G.

IEEE Transactions on Vehicular Technology 2025年

ISSN  00189545

　概要を見る

The rapid advancement of intelligent perception systems and autonomous navigation technologies imposes increasingly stringent requirements on state estimation algorithms regarding modeling capacity, robustness, and real-time performance. While traditional Kalman Filter (KF) demonstrates computational efficiency and mathematical interpretability for linear Gaussian systems, their effectiveness diminishes in nonlinear, non-Gaussian, and non-stationary dynamic scenarios. Recent developments in neural state space modeling offer new perspectives for recursive estimation, yet existing approaches suffer from structural complexity, training difficulties, and inadequate real-time capability. This paper presents Mamba-KF, a novel state estimation framework integrating deep modeling with recursive filtering. Leveraging the cutting-edge Mamba state space neural network, our method dynamically generates state transition parameters while establishing a differentiable covariance propagation pathway, enabling end-to-end optimization. We further design a lightweight multi-modal fusion architecture tailored for high-frequency sensory data processing in real-time applications. Evaluated on the EuRoC Micro Aerial Vehicle (MAV) dataset, Mamba-KF attains the lowest errors across states, reducing position Root Mean Square Error (RMSE) by up to 33.3% relative to Extended Kalman Filter (EKF) and 24.1% relative to Unscented Kalman Filter (UKF) at 10 dB Signal-to-Noise Ratio (SNR), with an average velocity RMSE of 0.097 m/s and an orientation error of 2.3°. It also achieves superior consistency, with a mean Normalized Estimation Error Squared (NEES) of 6.02 and 93.7% of estimates within the χ²-based 95% confidence interval. Under robustness evaluations, Mamba-KF yields the lowest RMSEs and suppresses divergence observed in EKF. Computationally, it maintains near real-time performance with 0.65 ms per step, offering a favorable accuracy-efficiency trade-off for visual-inertial navigation and autonomous systems, while providing significant theoretical insights and promising engineering applications.

• Access to Document (DOI)

Stress Detection Using Eye Activity via Recurrent Neural Networks

Bouazizi M., Feghoul R., Feghoul K., Ohtsuki T.

Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society EMBS 2025年

ISSN  1557170X

　概要を見る

Chronic stress has become a significant issue in modern society, negatively affecting individuals' overall well-being and potentially leading to severe physiological and psychological disorders. Most existing methods for automatic stress detection rely on physiological data, which requires direct skin contact with sensors. Not only does this make them intrusive, but it also limits their practicality for everyday use due to accuracy constraints. In this paper, we propose a novel approach using eye data captured by a standard video camera -specifically, eye gaze direction and eye landmarks- for stress detection. We use two types of Recurrent Neural Networks (RNNs), namely a Long Short-Term Memory (LSTM) network and a Gated Recurrent Unit (GRU) network. Through experiments on the UBFC-Phys dataset, we show that our method achieves a stress detection accuracy of up to 83.69% using gaze data, while eye landmarks further improve accuracy to 90.12%. Combined, these two modalities reach accuracy of 90.83%. These results highlight the effectiveness of our approach in enabling accurate and non-intrusive stress detection.

• Access to Document (DOI)

Stress Detection Through Eye Tracking Incorporating Custom Lasso Stress Index

Meng X., Bouazizi M., Ohtsuki T.

IEEE International Conference on Communications    3174 - 3179 2025年

ISSN  15503607

　概要を見る

Stress has become an increasingly common issue, significantly impacting individuals' health and productivity. Traditional methods for assessing stress levels often rely on physiological signals, such as heart rate and skin conductance. However, these methods can be impractical for continuous, non-invasive monitoring. To address these limitations, we propose a machine learning approach for stress assessment using eye-tracking data. We collected eye-tracking and Electrocardiogram (ECG) data under different stress conditions. From this dataset, we extracted relevant features and conducted preliminary analysis. We further proposed a novel stress index, the Least Absolute Shrinkage and Selection Operator Stress Index (Lasso Stress Index), specifically tailored for stress detection in our dataset. In previous studies, researchers commonly used predefined subjective stress index or Baevsky's Stress Index (BSI) for stress classification. We compared these methods and found that our proposed Lasso Stress Index yielded the highest accuracy, achieving up to 90.6% in stress classification.

• Access to Document (DOI)

Source Channel Model Selection for Downlink CSI Feedback Estimation using Transfer Learning in Massive MIMO System

Jingyu L., Bouazizi M., Ohtsuki T.

2025 International Conference on Computing Networking and Communications Icnc 2025    276 - 281 2025年

　概要を見る

Transfer learning-based methods for Channel State Information (CSI) feedback can achieve a good CSI reconstruction performance with small amounts of computation time and cost. However, for higher efficacy, such methods need to fine-tune a properly selected source channel model each time. This is because fine-tuning a model trained on drastically different data will suffer from slow convergence and non-optimal performance, which defies the purpose of transfer-learning. In this paper, we use a one-stage algorithm utilizing k-nearest neighbors (k-NN) and Deep Neural Network (DNN) methods to classify the downlink Clustered Delay Line (CDL) channels into one of 5 channel models with the help of a selected set of features. Based on the classification results, the appropriate source channel model for fine-tuning is selected, and the model is fine-tuned on the received data. To demonstrate the usefulness of our method, we compare its results to those of existing methods. Through simulation, we show that the overall classification accuracy for 5 CDL channel models using the one-step algorithm is 80.73%. Upon classification, and by selecting the proper source model, our method can reach a better normalized mean square error (NMSE) of CSI feedback performance in a shorter time and with fewer samples.

• Access to Document (DOI)

Semantic Communication in Vehicular Networks: A Multi-Modal Approach for Faithful Image Transmission

Bouazizi M., Nagase R., Yin Y., Yang S., Ohtsuki T.

IEEE Vehicular Technology Conference 2025年

ISSN  15502252

　概要を見る

Semantic communication (SC) has emerged as a promising paradigm to address the bandwidth limitations of traditional wireless communication systems by transmitting only the essential meaning of data. This paper investigates an advanced SC framework for vehicular communication, employing diverse feature extraction techniques to encode multi-modal information, such as textual descriptions, object poses, semantic segmentation, and sketches, into compact semantic representations. These semantic encoders are evaluated based on their output size, faithfulness of reconstruction, and resilience to data loss. The proposed system model considers vehicular communication scenarios where vehicles transmit important information extracted from camera-collected data to other vehicles and road users, in bandwidth-constrained environments. Simulation results show the effectiveness of the proposed SC framework, with a reconstruction performance reaching 17.24 in Fréchet Inception distance (FID) and a an RMSE equal to 0.029 between the transmitted image and the reconstructed one. This performance is achieved while the data saving between the size of the original image and the transmitted semantics is equal to 92.25%.

• Access to Document (DOI)

Remote RRI Estimation: A Signal Reconstruction Approach with Temporal Modeling

Wang S., Bouazizi M., Ohtsuki T.

2025 30th Asia Pacific Conference on Communications Apcc 2025 2025年

　概要を見る

We propose a neural network-based framework for radar-based heart rate monitoring and inter-beat interval (IBI) estimation that reconstructs the underlying cardiac waveform, rather than relying on Conventional peak detection methods. Conventional approaches typically estimate IBI by detecting peaks in radar signals after extensive signal processing, which can be sensitive to noise and transient signal degradation. Building on our previous work, which utilized a U-Net architecture with cross-channel attention, we introduce a enhanced model by incorporating Bidirectional Long Short-Term Memory (BiLSTM) layers and temporal self-attention into the bottleneck of the U-Net. This modification allows the network to explicitly model temporal dependencies and heartbeat periodicity, improving resilience to transient input degradation and physiological variability. The results are evaluated on a public radar-Electrocardiogram(ECG) dataset under Valsalva, Tilt Up, and Tilt Down conditions, our model achieves an average IBI estimation error of 22.54 ms Root Mean Square Error(RMSE), which outperforming both conventional peak-based methods and our prior U-Net-based design. These results demonstrate improved temporal generalization and signal stability, advancing the feasibility of radar-based cardiac monitoring in clinical and home settings.

• Access to Document (DOI)

Reinforcement Learning With Clustering Optimization for Antenna Parameter Adjustment in HAPS Networks

Bouazizi M., Yang S., Li Z., Yin Y., Ohtsuki T.

IEEE Access 13   172062 - 172079 2025年

　概要を見る

High-Altitude Platform Stations (HAPS) promise to extend broadband coverage over wide areas by acting as quasi-stationary aerial base stations. However, wind-induced movements cause antenna misalignment and unbalanced user distributions, leading to throughput degradation. This work presents a low-complexity, equal clustering (EC) -aided deep Q-network (DQN) algorithm for dynamic antenna control in HAPS networks. The proposed EC algorithm partitions users into balanced clusters by identifying low-density sectors in a single pass, avoiding the iterative refinements of other clustering algorithms such as K-Means. A DQN then fine-tunes the antenna tilt angles and beam widths based on the EC results to maximize the throughput of low-throughput users. The proposed method’s performance is evaluated through simulations under rotation and shift scenarios and benchmarked against three Reinforcement Learning (RL)-based approaches and a K-Means clustering baseline. Compared with these methods, the EC-aided DQN delivers higher throughput, especially for users in the low-throughput tail, while reducing computational complexity. For instance, the percentage of low-throughput users was reduced by over 10% in the case of 5 km shift in the regions of Osaka and Sendai, compared to K-means. Similarly, in the case of 30° rotation, the percentage is reduced by 4.1% and 6.5% for the same regions. Statistical significance tests confirm that the improvements are not due to random simulation noise.

• Access to Document (DOI)

Nullforming Strategy Based on User Distribution for Spectrum Sharing Between High-Altitude Platforms and Terrestrial Networks

Fontaine K., Kalør A.E., Ohtsuki T., Ishikawa T.

IEEE Vehicular Technology Conference 2025年

ISSN  10903038

　概要を見る

High-Altitude Platform Stations (HAPSs) enable wide-area coverage in 6G networks but introduce interference when sharing spectrum with terrestrial networks (TNs). Null-forming is a technique to mitigate this interference by directing low-power beams (nulls) toward terrestrial users. Traditional nullforming methods, such as Null-Sweeping, rely on changing null directions across the resource blocks (RB) to improve the impact of nullforming. Yet these null directions are predefined for uniform user distributions and may not fully account for nonuniform deployments. We propose a user-aware nullforming approach that leverages K-means clustering to adapt null positions to dense user regions in the terrestrial cells, while time-frequency resources are allocated proportionally to user density. Simulations show that our method reduces HAPS interference for terrestrial users and improves fairness in interference distribution.

• Access to Document (DOI)

Non-contact Heart Rate Estimation Using VMD with Spectrogram-Based Preprocessing via CW Radar

Wang X., Bouazizi M., Ohtsuki T.

Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society EMBS 2025年

ISSN  1557170X

　概要を見る

Vital signs such as heart rate are important indicators for assessing human health. In particular, noninvasive heart rate measurement using non-contact devices such as continuous wave (CW) radar has attracted attention as a promising technique for estimating heart rate by detecting small cardiovascular movements. However, existing estimation methods have difficulty in accurately estimating heart rate when strongly affected by noise such as respiration and body movement. In this research, we propose a heart rate estimation method with high robustness against noise by combining a pre-processing method using spectrogram and signal decomposition based on Variational Mode Decomposition (VMD). In the proposed method, preprocessing based on the spectrogram is first performed before applying VMD. The obtained integrated spectrum containing the heartbeat-related components is then processed using VMD, which decomposes it into multiple intrinsic mode function (IMF) signals. To select the IMF signals containing heartbeat components, the autocorrelation function and the frequency variation characteristic of heartbeat signals are used as criteria for selection. To evaluate the characteristics of the proposed method, measurements were performed on 10 subjects in a sitting position while electrocardiogram (ECG) was simultaneously recorded as a reference signal. Experimental results showed that the proposed method achieved a low mean absolute error (MAE) of 5.46 bpm compared to the conventional methods, effectively suppressing the effect of noise and improving the accuracy of heart rate estimation.

• Access to Document (DOI)

NoisyNet-DDQN-Based Sequential Handoff Algorithm for UAV Networks With End-to-End Network Slicing

Yang F., Li M., Ohtsuki T., Wu W., Si P., Sun T.

IEEE Transactions on Vehicular Technology 74 （ 12 ） 19496 - 19512 2025年

ISSN  00189545

　概要を見る

End-to-end network slicing in uncrewed aerial vehicle (UAV) networks has attracted significant attention for delivering differentiated services tailored to the quality of service (QoS) requirements of user terminals (UTs). In a sliced UAV network, the dual mobility of both UTs and UAVs, combined with the three-layer association of UT-UAV-network slice (NS), makes handoff a critical and challenging problem for maintaining the communication performance of UTs. This paper proposes a sequential handoff process where a central controller makes handoff decisions for each UT sequentially according to the UT queue. This process is modeled as a Markov decision process (MDP). In the MDP, the state space includes NS available resources, UTs' QoS requirements, and UT queue information. Additionally, the reward function accounts for UTs' QoS and handoff overhead. The double deep Q-network (DDQN) technique is then employed to design a handoff algorithm. This algorithm achieves the optimization goal of ensuring UTs' QoS requirements while avoiding frequent handoffs by maximizing the total reward of the MDP. Additionally, a state-dependent exploration technique with adaptive intensity adjustment, called noisy network (NoisyNet), is utilized to drive the DDQN for effective exploration, thereby enhancing the performance of the handoff algorithm. Finally, the runtime complexity of the proposed NoisyNet-DDQN-based sequential handoff (NoisyNet-DDQN-SH) algorithm is analyzed. The simulation results show that the NoisyNet-DDQN-SH algorithm converges well. Moreover, it exhibits strong robustness. Additionally, the NoisyNet technique effectively enhances the algorithm's performance, making the NoisyNet-DDQN-SH algorithm superior to the vanilla DDQN (with an \varepsilon-greedy strategy), greedy, maximum signal-to-interference-plus-noise ratio (Max-SINR), and random algorithms.

• Access to Document (DOI)

Multi-Dimensional Spectrum Prediction Method Based on Efficient Adaptive Broad Learning

Ji N., Wang S., Zhang Y., Peng Y., Otsuki T., Niyato D., Gui G.

IEEE Wireless Communications and Networking Conference Wcnc 2025年

ISSN  15253511

　概要を見る

Multi-dimensional spectrum prediction is essential for spectrum sharing and dynamic spectrum access (DSA), tack-ling spectrum scarcity and improving wireless communication. Traditional methods often use machine learning (ML), which requires manual feature extraction, or deep learning (DL), which demands high computational resources. This paper proposes a lightweight multi-dimensional spectrum prediction model using an adaptive broad learning network (ABLN). The model employs a sliding window to preprocess data and establishes input layers using randomly generated feature and enhancement nodes. The weights of broad learning are determined by solving the pseudo-inverse, and the structure is incrementally extended without retraining, reducing computational complexity. An adaptive node increment module optimizes hyperparameters efficiently. Experimental results demonstrate that ABLN reduces computational overhead while maintaining robust prediction performance across various scenarios.

• Access to Document (DOI)

More is Better: Channel-Robust Radio Frequency Fingerprinting with Random Overlay Augmentation

Wang Y., Meneghello F., Wang S., Otsuki T., Yuen C., Gui G., Wang X.

IEEE Wireless Communications and Networking Conference Wcnc 2025年

ISSN  15253511

　概要を見る

Radio Frequency Fingerprinting (RFF) is a critical technology for enhancing physical-layer security by leveraging the unique RF characteristics of hardware, enabling authentication and anti-counterfeiting for wireless communication devices. In recent years, Deep Learning (DL) has been extensively applied in RFF, significantly improving identification accuracy and efficiency. However, DL- based RFF methods still encounter challenges regarding robustness, particularly in cross-channel scenarios. To address these challenges, we propose a channel-robust RFF method based on a Multi-Scale Convolutional Attention Network (MSCAN) with Random Overlay Augmentation (ROA). Specifically, MSCAN extracts and fuses features at different scales, allowing it to capture more comprehensive signal characteristics. Additionally, ROA is a combinatorial data augmentation (DA) strategy designed to simulate diverse characteristics of wireless propagation environments, thereby enhancing the adaptability and robustness of RFF in complex channel conditions. Experiments conducted on the ORACLE dataset demonstrate that our proposed method achieves over 92 % accuracy in cross-channel scenarios, outperforming the previously proposed DA strategy. The codes will be published in GitHub¹¹https://github.com/BeechburgPieStar/SDG-for-Robust-SEI

• Access to Document (DOI)

MobiLLM: Enabling On-Device Fine-Tuning of Billion-Sized LLMs via Server-Assisted Side-Tuning

Li L., Yang X., Wu W., Wang H., Ohtsuki T., Fu X., Pan M., Shen X.

IEEE Journal on Selected Topics in Signal Processing 19 （ 7 ） 1251 - 1265 2025年

ISSN  19324553

　概要を見る

On-device fine-tuning of large language models (LLMs) has attracted a lot of attention because of its tailoring personalized models while retaining user data locally on the mobile device. However, it faces significant challenges due to prohibitive memory requirements and slow training speeds. In this paper, we propose MobiLLM, a novel scheme enabling memory-efficient LLM fine-tuning on a single mobile device via server-assisted side-tuning. Particularly, MobiLLM strategically offloads backpropagation computations to an edge server while allowing the resource-constrained mobile device to retain merely a pretrained backbone model with frozen parameters during finetuning. It constructs a backpropagation bypass via parallel adapters decoupled from the backbone. During forward propagation, the device employs low bitwidth quantization for transmitting intermediate activations to the server to reduce communication overhead. The advantage of MobiLLM lies in: 1) confining training data strictly to the mobile device, and 2) eliminating on-device backpropagation while overlapping local computations with server execution. Collectively, MobiLLM ensures the data never leaves the local mobile device while significantly reducing mobile memory and computational burdens. We implement MobiLLM on several popular mobile devices, including NVIDIA Jetson Xavier NX and CPU-only laptops. Extensive experimental results demonstrate that MobiLLM can enable a resource-constrained mobile device to fine-tune billion-sized LLMs, achieving up to 4x memory reduction and 2.3x faster convergence as compared to state-of-the-art baselines.

• Access to Document (DOI)