Subject: Cutting-Edge Advances in Wireless Communications, Sensing, and Signal Processing

Hi Elman,

This newsletter dives into the latest preprints exploring the forefront of wireless communications, sensing, and signal processing. We'll cover a range of topics, from integrated sensing and communication to the application of machine learning in biomedical signal analysis.

General Overview

This collection of preprints explores diverse challenges and advancements in wireless communications, sensing, and signal processing. Several papers focus on improving the performance and security of emerging wireless technologies. Galappaththige et al. (2025) Galappaththige et al. (2025) provide a comprehensive survey of cell-free integrated sensing and communication (CF-ISAC), highlighting its potential for enhanced spectral and energy efficiency, distributed multi-static sensing, and robust multi-user communication. They categorize state-of-the-art developments, discuss key challenges like synchronization and interference management, and outline future directions, including next-generation antenna technologies and machine learning integration. Similarly, Celebi et al. (2025) Celebi et al. (2025) investigate OTFS waveform application in LEO-satellite systems with HAPS relaying, deriving closed-form outage probability expressions and demonstrating improved performance with increased OTFS order and transmit antennas. Addressing security concerns in RIS-aided positioning, Li et al. (2025) Li et al. (2025) analyze the impact of unauthorized RIS interference on positioning accuracy, evaluating different codebook design strategies and system arrangements. Tomic et al. (2025) Tomic et al. (2025) propose a novel voting scheme based on clustering and weighted central mass for secure target localization and attacker detection in wireless sensor networks, demonstrating improved accuracy and near-perfect attacker detection rates under specific conditions.

Another prominent theme is the application of advanced signal processing and machine learning techniques. Buciulea et al. (2025) Buciulea et al. (2025) introduce a greedy approach for learning the topology of a simplicial complex using simplicial signals, leveraging signal smoothness and sparsity assumptions. Yue et al. (2025) Yue et al. (2025) present NeRFCom, a novel system for 3D scene semantic transmission using feature transform coding and neural radiance fields, demonstrating robust performance under adverse channel conditions. Park et al. (2025) Park et al. (2025) propose a transformer-based nonlinear transform coding approach for multi-rate CSI compression in MIMO-OFDM systems, achieving state-of-the-art rate-distortion performance. In the biomedical domain, Cai et al. (2025) Cai et al. (2025) introduce SuPreME, a supervised pre-training framework for multimodal ECG representation learning using LLMs to extract clinical entities from ECG reports, enabling zero-shot classification of unseen diseases. Furthermore, Moulaeifard et al. (2025) (Moulaeifard et al., 2025) benchmark various machine learning approaches for PPG analysis, finding that deep neural networks operating on raw time series achieve the best results for blood pressure and atrial fibrillation prediction. A separate study by Moulaeifard et al. (2025) (Moulaeifard et al., 2025) investigates the generalizability of deep learning models for PPG-based blood pressure estimation, highlighting the importance of out-of-distribution generalization and proposing domain adaptation techniques.

Several contributions focus on specific applications and system designs. Shahid et al. (2025) Shahid et al. (2025) propose ReVeal, a physics-informed neural network for high-fidelity radio environment mapping, demonstrating significant improvements in accuracy compared to existing models. Wang et al. (2025) Wang et al. (2025) present an electromagnetically reconfigurable fluid antenna system (ER-FAS) for wireless communications, demonstrating enhanced spectral efficiency through joint optimization of analog phase shift and radiation state. Rostami Ghadi et al. (2025) Rostami Ghadi et al. (2025) analyze the performance of FAS-aided covert communications, highlighting the trade-off between covertness and transmission success. Kulkarni et al. (2025) Kulkarni et al. (2025) propose a KAN-powered large-target detection method for automotive radar, demonstrating improved performance over traditional OS-CFAR. Heo and Choi (2025) Heo and Choi (2025) introduce a DFT-based near-field beam alignment scheme for XL-MIMO systems, combining model-based and data-driven approaches for enhanced accuracy and reduced complexity. Bazzi and Chafii (2025) Bazzi and Chafii (2025) explore ISAC RIS-enabled passive radar target localization, proposing a normalized least mean squares method for joint target detection and angle-of-arrival estimation.

Further contributions explore diverse topics in wireless systems and signal processing. Jiang et al. (2025) Jiang et al. (2025) introduce MTCA, a multi-task channel analysis framework for wireless communication, demonstrating improved performance in channel prediction, extrapolation, identification, and scenario classification. Asadi Ahmadabadi et al. (2025) Asadi Ahmadabadi et al. (2025) investigate DRL-based secure spectrum-reuse D2D communications with RIS assistance, demonstrating enhanced secrecy capacity. Chu et al. (2025) Chu et al. (2025) propose a transfer learning approach for fast design migration of mm-Wave passive networks across technology nodes, demonstrating significant reductions in required dataset size. Dureppagari et al. (2025) Dureppagari et al. (2025) introduce a two-stage weighted projection method for TDOA-based localization, demonstrating improved accuracy and robustness under challenging conditions. Hanif et al. (2025) Hanif et al. (2025) present a novel passive attack method for user localization based on modulation classification, highlighting security vulnerabilities in wireless standards. Finally, Tamo Amougou et al. (2025) Tamo Amougou et al. (2025) propose a self-supervised conformal prediction method for uncertainty quantification in Poisson imaging problems, demonstrating comparable performance to supervised methods without requiring ground truth data.

These preprints collectively demonstrate significant progress in various aspects of wireless communications and signal processing, ranging from novel system architectures and signal processing techniques to security considerations and machine learning applications. The emphasis on integrated sensing and communication, reconfigurable intelligent surfaces, and data-driven approaches reflects the current trends in the field, paving the way for more efficient, robust, and intelligent wireless systems. The exploration of both theoretical analysis and practical implementations, including real-world experiments and simulations, strengthens the validity and applicability of the presented research.

Paper Highlights

Cell-Free Integrated Sensing and Communication: A Deep Dive into the Future of Wireless Networks

Cell-Free Integrated Sensing and Communication: Principles, Advances, and Future Directions by Diluka Galappaththige, Mohammadali Mohammadi, Gayan Aruma Baduge, Chintha Tellambura https://arxiv.org/abs/2502.20345

Caption: This image illustrates a cell-free integrated sensing and communication (CF-ISAC) network. Multiple access points (APs), represented by the towers, are connected to a central server and simultaneously serve communication users (people with phones) and perform sensing tasks, highlighting the distributed nature of CF-ISAC and its ability to enhance both communication and sensing performance. The dashed lines represent the connections between the APs and the central server.

Cell-free (CF) integrated sensing and communication (ISAC) is rapidly emerging as a key technology for next-generation wireless networks. This paper offers a comprehensive overview of CF-ISAC, exploring its fundamental principles, recent advancements, and future research directions. Unlike traditional co-located ISAC, CF-ISAC utilizes a distributed network of access points (APs) to simultaneously serve communication users and perform multi-static sensing. This distributed approach offers substantial advantages in spectral efficiency, sensing accuracy, and overall system robustness.

The paper begins with a foundational review of cell-free massive MIMO (CFMM) and radar sensing, providing the necessary background for understanding CF-ISAC. It then delves into the principles of conventional ISAC, covering integration levels, design philosophies, key sensing metrics, and potential applications. The core of the paper focuses on CF-ISAC systems, emphasizing their unique characteristics and the inherent benefits of multi-static sensing. This includes a detailed discussion of distributed antenna systems, seamless handovers, interference management techniques, AP cooperation strategies, and user/target-centric operations.

The authors categorize state-of-the-art CF-ISAC research into four key areas: performance analysis, resource allocation, security considerations, and user/target-centric designs. They provide a thorough review of existing literature and present insightful case studies. One notable example is the definition of achievable sensing spectral efficiency (SE) as SSen = maxTr(XXH)<pmax 1/L log2(det(XHRsCX + IL)), where pmax represents the maximum allowable transmit power, X denotes the transmitted signal, Rs is the spatial correlation matrix of the target response, and L represents the number of symbols.

A case study evaluates a generalized CF-ISAC system with multiple targets and users. Simulation results, utilizing a 3GPP Urban micro (UMi) channel model, demonstrate that increasing the number of antennas at both uplink (UL) and downlink (DL) APs significantly improves both communication and sensing SE. This improvement is attributed to enhanced spatial diversity and reduced interference. Specifically, increasing the number of antennas from 10 to 20 with 9 UL/DL APs resulted in a 31.3% gain in communication SE and a 51.7% gain in sensing SE. Further analysis reveals a trade-off between communication and sensing performance, as allocating more resources to sensing tasks with a higher number of targets necessarily reduces the resources available for communication.

Finally, the paper identifies key challenges in CF-ISAC. These include synchronization across distributed APs, multi-target detection in cluttered environments, effective interference management between sensing and communication tasks, and limitations in fronthaul capacity and latency. The authors also explore emerging trends and open research directions, such as network-assisted CF-ISAC, the integration of new antenna technologies like holographic MIMO and fluid antennas, near-field CF-ISAC for enhanced localization, the consolidation of complementary technologies like RIS and NOMA, and the application of machine learning techniques for optimized resource allocation and interference management. These insights provide a valuable roadmap for future research and development in this promising field, paving the way for the next generation of wireless networks.

Revolutionizing ECG Analysis: Supervised Pre-training with SuPreME

SuPreME: A Supervised Pre-training Framework for Multimodal ECG Representation Learning by Mingsheng Cai, Jiuming Jiang, Wenhao Huang, Che Liu, Rossella Arcucci https://arxiv.org/abs/2502.19668

Caption: SuPreME's Cardiac Fusion Network (CFN) aligns ECG feature maps with text-based cardiac query embeddings using a multi-head cross-attention mechanism within a Transformer decoder architecture. This allows the model to learn relationships between ECG signals and medical terminology, enabling zero-shot classification of cardiac conditions. The input ECG signal is processed to generate an ECG feature map, while the medical terminology from the ECG report is embedded using a pre-trained medical BERT model and positional encoding.

SuPreME (Supervised Pre-training framework for Multimodal ECG representation learning) is a novel framework poised to revolutionize electrocardiogram (ECG) analysis. This innovative approach addresses the limitations of traditional methods that rely on extensive labeled datasets, which are resource-intensive and difficult to obtain. SuPreME leverages the power of Large Language Models (LLMs) to extract structured clinical entities from free-text ECG reports, effectively filtering out noise and irrelevant information. This automated pipeline generates a high-quality, fine-grained labeled dataset with 295 standardized medical terminologies, eliminating the need for manual labeling and ensuring scalability and consistency. Furthermore, instead of using traditional categorical labels, SuPreME utilizes text-based cardiac queries, enabling zero-shot classification of unseen diseases without the need for additional fine-tuning – a significant advancement in the field.

The core of SuPreME's innovation lies in its ability to align ECG signals with structured entity labels. This alignment is achieved by embedding both ECG signals and text-based cardiac queries into a shared latent space. A Vision Transformer (ViT) is used for ECGs and a pre-trained medical BERT model (MedCPT) is used for text. A specialized Cardiac Fusion Network (CFN), composed of multi-layer Transformer decoders, then aligns these embeddings, capturing cross-modal dependencies and grounding the interpretation of ECG signals in medical queries. This sophisticated architecture allows SuPreME to learn rich, clinically meaningful representations without resorting to complex data augmentations or pretext tasks that can distort the semantic integrity of the signal, a common issue in self-supervised learning methods.

Trained on a massive dataset of 771,500 ECG signals paired with extracted entities from the MIMIC-IV-ECG dataset, SuPreME demonstrates remarkable performance. Evaluation on six downstream datasets covering 127 cardiac conditions showcased its superiority. In a zero-shot scenario, SuPreME achieved an overall AUC of 77.20%, outperforming state-of-the-art multimodal methods and most non-multimodal self-supervised learning models, even those utilizing linear probing with labeled data. Impressively, SuPreME achieved comparable performance to multimodal contrastive learning frameworks with significantly fewer training epochs (16 vs. 50). Further analysis revealed that SuPreME maintains a significant performance edge even with limited pre-training data, highlighting its efficiency and robust generalization capabilities.

SuPreME’s success underscores the power of integrating LLMs with supervised pre-training for multimodal ECG representation learning. By leveraging structured, clinically relevant knowledge, SuPreME overcomes the limitations of traditional methods and paves the way for zero-shot diagnosis of cardiac conditions, potentially transforming clinical practice. However, the framework’s reliance on LLMs for entity extraction presents a potential challenge in capturing highly specialized medical knowledge. Future research will focus on enhancing the robustness of clinical entity extraction, developing adaptive strategies for diverse ECG data, and further optimizing the pre-training process to address these limitations and fully unlock the potential of this groundbreaking approach.

Robust Over-the-Air Computation: TBMA Fights Back Against Byzantine Attacks

Robust Over-the-Air Computation with Type-Based Multiple Access by Marc Martinez-Gost, Ana Pérez-Neira, Miguel Ángel Lagunas https://arxiv.org/abs/2502.19014

Caption: This graph compares the Normalized Mean Squared Error (NMSE) of Direct Aggregation (DA) and a robust Type-Based Multiple Access (TBMA) algorithm (Algorithm 1) under varying ratios of Byzantine attackers. It demonstrates the vulnerability of DA to attacks, with NMSE increasing rapidly, while Algorithm 1 maintains significantly lower NMSE, showcasing its robustness.

Over-the-air computation (AirComp) presents a promising approach for aggregating data directly from distributed devices, leveraging the superposition property of wireless channels. However, its vulnerability to Byzantine attacks, where malicious nodes inject false data to corrupt the aggregated result, poses a significant security challenge. Traditional direct aggregation (DA) AirComp, which encodes information in signal amplitude, is particularly susceptible to these attacks. This paper introduces a novel approach utilizing type-based multiple access (TBMA) to enhance the robustness of AirComp against Byzantine adversaries.

Unlike DA, TBMA allocates orthogonal radio resources based on data values rather than individual devices. This allows the receiver to construct a histogram, or type, of the transmitted data: $\textbf{r} = [\frac{K_1}{K},...,\frac{K_L}{K}] + [\tilde{w}_1,...,\tilde{w}_L] = \textbf{p} + \textbf{w}$, where $K_l$ is the number of devices transmitting data $l$, $K$ is the total number of devices, and $\textbf{w}$ represents the noise. This structure enables the integration of robust estimation techniques to detect and mitigate the impact of malicious transmissions. The proposed method incorporates noise thresholding, percentile truncation, and local outlier compensation to refine the received type before applying the desired aggregation function $\Psi(\textbf{r})$. Critically, TBMA decouples attack detection and compensation from the function computation, allowing for greater flexibility compared to DA.

The paper evaluates the performance of the proposed robust TBMA approach against DA and a median-based TBMA method under various attack scenarios and for different aggregation functions, including the arithmetic and geometric mean. The results demonstrate that DA is highly vulnerable to Byzantine attacks, with performance degrading rapidly as the number of attackers increases. While median-based TBMA offers some improvement, it remains susceptible to a high ratio of attackers and is limited to the arithmetic mean. In contrast, the proposed robust TBMA scheme exhibits significant resilience, maintaining high accuracy even with a substantial number of attackers. For example, in an experiment with 10,000 devices and 256 radio resources, the robust TBMA approach maintained accuracy even with up to 50% of devices acting as attackers, while DA completely failed under the same conditions.

The practical applicability of robust TBMA is further validated in a federated learning (FEEL) use case for image classification using the MNIST dataset. With 6% of the devices acting as Byzantine attackers, DA achieved only 10% accuracy, equivalent to random guessing. Robust TBMA, however, maintained accuracy close to the ideal scenario without attacks, highlighting its potential for real-world deployments. These findings establish TBMA as a scalable and robust solution for AirComp, paving the way for secure and efficient aggregation in future wireless networks.

Conclusion

This newsletter highlights the exciting advancements being made in wireless communications and signal processing. From the innovative CF-ISAC architecture that promises to revolutionize network efficiency and sensing capabilities, to the groundbreaking SuPreME framework that leverages LLMs for enhanced ECG analysis, and the robust TBMA approach that safeguards against Byzantine attacks in AirComp, these preprints showcase the dynamic nature of the field. The common thread weaving through these diverse contributions is the increasing reliance on data-driven approaches and the integration of intelligent algorithms to address complex challenges and unlock the full potential of next-generation wireless systems. These advancements not only push the boundaries of theoretical understanding but also offer practical solutions with tangible real-world implications.