AI/ML utilization in Open RAN plays a crucial role in enhancing the overall network performance, particularly for tasks such as resource and mobility management, quality of service, and security. Integrating AI and ML into the Open RAN framework offers significant improvements in network operations and functionality. However, this integration introduces several security threats, including data poisoning, data reconstruction, and adversarial attacks, which can compromise the integrity and reliability of ML models. To address these challenges and ensure that the final systems use robust, reliable, and private ML models, several techniques can be implemented. These include adversarial training, which helps models become more resilient to malicious inputs, and collaborative learning methodologies like federated learning and split learning, which enhance data privacy by decentralising the training process. By adopting these approaches, it is possible to mitigate ML security vulnerabilities, fostering a more resilient, secure, and trustworthy Open RAN ecosystem.

The parameter size of modern large language models (LLMs) can be scaled up to the trillion-level via the sparsely-activated Mixture-of-Experts (MoE) technique to avoid excessive increase of the computational costs. To further improve training efficiency, pipelining computation and communication has become a promising solution for distributed MoE training. However, existing work primarily focuses on scheduling tasks within the MoE layer, such as expert computing and all-to-all (A2A) communication, while neglecting other key operations including multi-head attention (MHA) computing, gating, and all-reduce communication. In this paper, we propose FlowMoE, a scalable framework for scheduling multi-type task pipelines. First, FlowMoE constructs a unified pipeline to consistently scheduling MHA computing, gating, expert computing, and A2A communication. Second, FlowMoE introduces a tensor chunk-based priority scheduling mechanism to overlap the all-reduce communication with all computing tasks. We implement FlowMoE as an adaptive and generic framework atop PyTorch. Extensive experiments with 675 typical MoE layers and four real-world MoE models across two GPU clusters demonstrate that our proposed FlowMoE framework outperforms state-of-the-art MoE training frameworks, reducing training time by14%-57%, energy consumption by 10%-39%, and memory usage by 7%-32%. FlowMoE’s code is anonymously available at https://anonymous.4open.science/r/FlowMoE.

Generative foundation models can revolutionize the design of semantic communication (SemCom) systems by enabling high fidelity exchange of semantic information at ultra-low rates. In this work, a generative SemCom framework utilizing pre-trained foundation models is proposed, where both uncoded forward-with-error and coded discard-with-error schemes are developed for the semantic decoder. Using the rate-distortion-perception theory, the relationship between regenerated signal quality and transmission reliability is characterized, which is proven to be non-decreasing. Based on this, semantic values are defined to quantify the semantic similarity between multimodal semantic features and the original source. We also investigate semantic-aware power allocation problems that minimize power consumption for ultra-low rate and high fidelity SemComs. Two semantic-aware power allocation methods are proposed by leveraging the non-decreasing property of the perception-error relationship. Based on the Kodak dataset, perception-error functions and semantic values are obtained for image tasks. Simulation results show that the proposed semantic-aware method significantly outperforms conventional approaches, particularly in the channel-coded case (up to 90% power saving).

To enable improved spectral efficiency and uniform service of cell-free massive multiple-input multiple-output (CF-mMIMO) systems, challenges such as excessive communication overheads for gathering global CSI at the central processor (CP) and high computational complexity for centralized interference management should be addressed. Deep learning (DL)-based decentralized optimization approaches have been proposed to tackle these challenges. However, conventional black-box deep neural network (DNN)-based fronthaul coordination methods require large DNNs and are less adaptable to changes in wireless channel statistics and environments. In this paper, we develop a model-based neural collaboration (NC) framework, which leverages DNN-assisted signal processing built upon our proposed low-complexity distributed weighted minimum mean squared error (L-D-WMMSE) algorithm, for downlink CF-mMIMO beamforming. DNN components for neural initialization (NI) and neural parameterization (NP) reduce the fronthaul overheads while improving the sum-rate of user equipments (UEs). Through our model-based approach, these DNN components are designed to be compact, thereby reducing computational complexity. We derive gradients for DNN updates that each AP can compute using its local CSI and the fronthaul messages exchanged for downlink beamforming. This enables a distributed online DNN adaptation that runs during normal operation of the CF-mMIMO system at zero additional fronthaul overheads.

Split Federated Learning (SFL) improves scalability of Split Learning (SL) by enabling parallel computing of the learning tasks on multiple clients. However, state-of-the-art SFL schemes neglect the effects of heterogeneity in the clients’ computation and communication performance as well as the computation time for the tasks offloaded to the cloud server. In this paper, we propose a fine-grained parallelization framework, called PipeSFL, to accelerate SFL on heterogeneous clients. PipeSFL is based on two key novel ideas. First, we design a server-side priority scheduling mechanism to minimize per-iteration time. Second, we propose a hybrid training mode to reduce per-round time, which employs asynchronous training within rounds and synchronous training between rounds. We theoretically prove the optimality of the proposed priority scheduling mechanism within one round and analyze the total time per round for PipeSFL, SFL and SL. We implement PipeSFL on PyTorch. Extensive experiments on seven 64-client clusters with different heterogeneity demonstrate that at training speed, PipeSFL achieves up to 1.65x and 1.93x speedup compared to EPSL and SFL, respectively. At energy consumption, PipeSFL saves up to 30.8% and 43.4% of the energy consumed within each training round compared to EPSL and SFL, respectively.

Over-the-air computation (AirComp) is a promising technology converging communication and computation over wireless networks, which can be particularly effective in model training, inference, and more emerging edge intelligence applications. AirComp relies on uncoded transmission of individual signals, which are added naturally over the multiple access channel thanks to the superposition property of the wireless medium. Despite significantly improved communication efficiency, how to accommodate AirComp in the existing and future digital communication networks, that are based on discrete modulation schemes, remains a challenge. This paper proposes a massive digital AirComp (MD-AirComp) scheme, that leverages an unsourced massive access protocol, to enhance compatibility with both current and next-generation wireless networks. MD-AirComp utilizes vector quantization to reduce the uplink communication overhead, and employs shared quantization and modulation codebooks. At the receiver, we propose a near-optimal approximate message passing-based algorithm to compute the model aggregation results from the superposed sequences, which relies on estimating the number of devices transmitting each code sequence, rather than trying to decode the messages of individual transmitters. We apply MD-AirComp to federated edge learning (FEEL), and show that it significantly accelerates FEEL convergence compared to state-of-the-art while using the same amount of communication resources.

The communication bottleneck severely constrains the scalability of distributed deep learning, and efficient communication scheduling accelerates distributed DNN training by overlapping computation and communication tasks. However, existing approaches based on tensor partitioning are not efficient and suffer from two challenges: 1) the fixed number of tensor blocks transferred in parallel can not necessarily minimize the communication overheads; 2) although the scheduling order that preferentially transmits tensor blocks close to the input layer can start forward propagation in the next iteration earlier, the shortest per-iteration time is not obtained. In this paper, we propose an efficient communication framework called US-Byte. It can schedule unequal-sized tensor blocks in a near-optimal order to minimize the training time. We build the mathematical model of US-Byte by two phases: 1) the overlap of gradient communication and backward propagation, and 2) the overlap of gradient communication and forward propagation. We theoretically derive the optimal solution for the second phase and efficiently solve the first phase with a low-complexity algorithm. We implement the US-Byte architecture on PyTorch framework. Extensive experiments on two different 8-node GPU clusters demonstrate that US-Byte can achieve up to 1.26x and 1.56x speedup compared to ByteScheduler and WFBP, respectively. We further exploit simulations of 128 GPUs to verify the potential scaling performance of US-Byte. Simulation results show that US-Byte can achieve up to 1.69x speedup compared to the state-of-the-art communication framework.

Deep neural networks (DNNs) in the wireless communication domain have been shown to be hardly generalizable to scenarios where the train and test datasets follow a different distribution. This lack of generalization poses a significant hurdle to the practical utilization of DNNs in wireless communication. In this paper, we propose a generalizable deep learning approach for millimeter wave (mmWave) beam selection using sub-6 GHz channel state information (CSI) measurements, referred to as PARAMOUNT. First, we provide a detailed discussion on physical aspects of the electromagnetic wave scattering in the mmWave and sub-6 GHz bands. Based on this discussion, we develop the augmented discrete angle delay profile (ADADP) which is a novel linear transformation for the sub-6 GHz CSI that extracts the angle-delay attributes and provides a semantic visual representation of the multi-path clusters. Next, we introduce a convolutional neural network (CNN) structure that can learn the signatures of the path clusters in the sub-6 GHz ADADP representation and transform it to mmWave band beam indices. We demonstrate by extensive simulations on several different datasets that PARAMOUNT can generalize beyond the training dataset which is mainly due to transfer learning principles that allow transferring information from previously learned tasks to the learning of new unseen tasks.

With the huge number of broadband users, automated network management becomes of huge interest to service providers. A major challenge is automated monitoring of user Quality of Experience (QoE), where Artificial Intelligence (AI) and Machine Learning (ML) models provide powerful tools to predict user QoE from basic protocol indicators such as Round Trip Time (RTT), retransmission rate, etc. In this paper, we introduce an effective feature selection method along with the corresponding classification algorithms to address this challenge. The simulation results show a prediction accuracy of 78% on the benchmark ITU ML5G-PS-012 dataset, improving 11% over the state-of-the-art result whilst reducing the model complexity at the same time. Moreover, we show that the local area network round trip time (LAN RTT) value during daytime and midweek plays the most prominent factor affecting the user QoE.

With the large number of antennas and subcarriers the overhead due to pilot transmission for channel estimation can be prohibitive in wideband massive multiple-input multiple-output (MIMO) systems. This can degrade the overall spectral efficiency significantly, and as a result, curtail the potential benefits of massive MIMO. In this paper, we propose a neural network (NN)-based joint pilot design and downlink channel estimation scheme for frequency division duplex (FDD) MIMO orthogonal frequency division multiplex (OFDM) systems. The proposed NN architecture uses fully connected layers for frequency-aware pilot design, and outperforms linear minimum mean square error (LMMSE) estimation by exploiting inherent correlations in MIMO channel matrices utilizing convolutional NN layers. Our proposed NN architecture uses a non-local attention module to learn longer range correlations in the channel matrix to further improve the channel estimation performance.We also propose an effective pilot reduction technique by gradually pruning less significant neurons from the dense NN layers during training. This constitutes a novel application of NN pruning to reduce the pilot transmission overhead. Our pruning-based pilot reduction technique reduces the overhead by allocating pilots across subcarriers non-uniformly and exploiting the inter-frequency and inter-antenna correlations in the channel matrix efficiently through convolutional layers and attention module.

Massive multiple-input multiple-output (MIMO) systems require downlink channel state information (CSI) at the base station (BS) to achieve spatial diversity and multiplexing gains. In a frequency division duplex (FDD) multiuser massive MIMO network, each user needs to compress and feedback its downlink CSI to the BS. The CSI overhead scales with the number of antennas, users and subcarriers, and becomes a major bottleneck for the overall spectral efficiency. In this paper, we propose a deep learning (DL)-based CSI compression scheme, calledDeepCMC, composed of convolutional layers followed by quantization and entropy coding blocks. In comparison with previous DL-based CSI reduction structures, DeepCMC proposes a novel fully-convolutional neural network (NN) architecture, with residual layers at the decoder, and incorporates quantization and entropy coding blocks into its design. DeepCMC is trained to minimize a weighted rate-distortion cost, which enables a trade-off between the CSI quality and its feedback overhead. Simulation results demonstrate that DeepCMC outperforms the state of the art CSI compression schemes in terms of the reconstruction quality of CSI for the same compression rate. We also propose a distributed version of DeepCMC for a multi-user MIMO scenario to encode and reconstruct the CSI from multiple users in a distributed manner. Distributed DeepCMC not only utilizes the inherent CSI structures of a single MIMO user for compression, but also benefits from the correlations among the channel matrices of nearby users to further improve the performance in comparison with DeepCMC. We also propose a reduced-complexity training method for distributed DeepCMC, allowing to scale it to multiple users, and suggest a cluster-based distributed DeepCMC approach for practical implementation.

Massive multiple-input multiple-output (MIMO) systems require downlink channel state information (CSI) at the base station (BS) to better utilize the available spatial diversity and multiplexing gains. However, in a frequency division duplex (FDD) massive MIMO system, CSI feedback overhead degrades the overall spectral efficiency. Deep Learning (DL)-based CSI feedback compression schemes have received a lot of attention recently as they provide significant improvements in compression efficiency; however, they still require reliable feedback links to convey the compressed CSI information to the BS. Instead, we propose here a Convolutional neural network (CNN)-based analog feedback scheme, called AnalogDeepCMC, which directly maps the downlink CSI to uplink channel input. Corresponding noisy channel outputs are used by another CNN to reconstruct the downlink channel estimate. The proposed analog scheme not only outperforms existing digital CSI feedback schemes in terms of the achievable downlink rate, but also simplifies the feedback transmission as it does not require explicit quantization, coding, and modulation, and provides a low-latency alternative particularly in rapidly changing MIMO channels, where the CSI needs to be estimated and fed back periodically.

This paper presents a low complexity while accurate Heart Rate (HR) estimation technique from signals captured by Photoplethysmographic (PPG) sensors worn on the wrist during intensive physical exercise. Wrist-type PPG signals experience severe Motion Artifacts (MA) that hinder efficient HR estimation especially during intensive physical exercises. To suppress the motion artifacts efficiently, simultaneous 3 dimensional acceleration signals are used as reference MAs. The proposed method achieves an Average Absolute Error (AAE) of 1.19 Beats Per Minute (BPM) on the 12 benchmark PPG recordings in which subjects run at speeds of up to 15 km/h. This method also achieves an AAE of 2.17 BPM on the whole benchmark database of 23 recordings that include both running and arm movement activities. This performance is comparable with state-of-the-art algorithms while at a significantly reduced computational cost which makes its standalone implementation on wearable devices feasible. The proposed algorithm achieves an average processing time of 32 milliseconds per input frames of length 8 seconds (2 channel PPG and 3D ACC signals) on a 3.2 GHz processor.

In this letter, we propose a sparsity promoting feedback acquisition and reconstruction scheme for sensing, encoding and subsequent reconstruction of spectrally sparse signals. In the proposed scheme, the spectral components are estimated utilizing a sparsity-promoting, sliding-window algorithm in a feedback loop. Utilizing the estimated spectral components, a level signal is predicted and sign measurements of the prediction error are acquired. The sparsity promoting algorithm can then estimate the spectral components iteratively from the sign measurements. Unlike many batch-based compressive sensing algorithms, our proposed algorithm gradually estimates and follows slow changes in the sparse components utilizing a sliding-window technique. We also consider the scenario in which possible flipping errors in the sign bits propagate along iterations (due to the feedback loop) during reconstruction. We propose an iterative error correction algorithm to cope with this error propagation phenomenon considering a binary-sparse occurrence model on the error sequence. Simulation results show effective performance of the proposed scheme in comparison with the literature.

The authors propose asynchronous level crossing (LC) A/D converters for low redundancy voice sampling. They propose to utilise the family of iterative methods with adaptive thresholding (IMAT) for reconstructing voice from non‐uniform LC and adaptive LC (ALC) samples thereby promoting sparsity. The authors modify the basic IMAT algorithm and propose the iterative method with adaptive thresholding for level crossing (IMATLC) algorithm for improved reconstruction performance. To this end, the authors analytically derive the basic IMAT algorithm by applying the gradient descent and gradient projection optimisation techniques to the problem of square error minimisation subjected to sparsity. The simulation results indicate that the proposed IMATLC reconstruction method outperforms the conventional reconstruction method based on low‐pass signal assumption by 6.56 dBs in terms of reconstruction signal‐to‐noise ratio (SNR) for LC sampling. In this scenario, IMATLC outperforms orthogonal matching pursuit, least absolute shrinkage and selection operator and smoothed L0 sparsity promoting algorithms by average amounts of 12.13, 10.31, and 10.28 dBs, respectively. Finally, the authors compare the performance of the proposed LC/ALC‐based A/Ds with the conventional uniform sampling‐based A/Ds and their random sampling‐based counterparts both in terms of perceptual evaluation of speech quality and reconstruction SNR.

The common voice channels existing in cellular communication networks provide reliable, ubiquitously available and top priority communication mediums. These properties make voice dedicated channels an ideal choice for high priority, real time communication. However, such channels include voice codecs that hamper the data flow by compressing the waveforms prior to transmission. This study designs codebooks of speech‐like symbols for reliable data transfer through the voice channel of cellular networks. An efficient algorithm is proposed to select proper codebook symbols from a database of natural speech to optimise a desired objective. Two variants of this codebook optimisation algorithm are presented: One variant minimises the symbol error rate and the other maximises the capacity achievable by the codebook. It is shown both analytically and by the simulation results that under certain circumstances, these two objective functions reach the same performance. Simulation results also show that the proposed codebook optimisation algorithm achieves higher data rates and lower symbol error rates compared with previously reported results while requiring lower computational complexity for codebook optimisation. The Gilbert–Elliot channel model is utilised to study the effects of adaptive compression rate adjustment of the vocoder on overall voice channel capacity. Finally, practical implementation issues are addressed.

This letter considers the problem of casual heart rate tracking during intensive physical exercise using simultaneous 2 channel photoplethysmographic (PPG) and 3 dimensional (3D) acceleration signals recorded from wrist. This is a challenging problem because the PPG signals recorded from wrist during exercise are contaminated by strong Motion Artifacts (MAs). In this work, a novel algorithm is proposed which consists of two main steps of MA Cancellation and Spectral Analysis. The MA cancellation step cleanses the MA-contaminated PPG signals utilizing the acceleration data and the spectral analysis step estimates a higher resolution spectrum of the signal and selects the spectral peaks corresponding to HR. Experimental results on datasets recorded from 12 subjects during fast running at the peak speed of 15 km/hour showed that the proposed algorithm achieves an average absolute error of 1.25 beat per minute (BPM). These experimental results also confirm that the proposed algorithm keeps high estimation accuracies even in strong MA conditions.

The increased penetration of cellular networks has made voice channels widely available ubiquitously. On the other hand, mobile voice channels possess properties that make them an ideal choice for high priority, low‐rate real‐time communications. Mobile voice channel with the mentioned properties, could be utilised in emergency applications in vehicular communications area such as the standardised emergency call system planned to be launched in 2015. This study aims to investigate the challenges of data transmission through these channels and proposes an efficient data transfer structure. To this end, a proper statistical model for the channel distortion is proposed and an optimum detector is derived considering the proposed channel model. Optimum symbols are also designed according to the derived rule and analytical bounds on error probability are obtained for the orthogonal signaling and sphere packing techniques. Moreover, analytical evaluation is performed and appropriate simulation results are presented. Finally, it is observed that the proposed structure based on the sphere packing technique achieves superior performance compared with prior works in this field. Although the ideas offered in this study are utilised to cope with voice channel non‐idealities, the steps taken in this study could also be applied to channels with similar conditions.