Abstract Recent efforts to improve the efficiency of neuromorphic and machine learning systems have centred on developing of specialised hardware for neural networks. These systems typically feature architectures that go beyond the von Neumann model employed in general-purpose hardware such as GPUs, offering potential efficiency and performance gains. However, neural networks developed for specialised hardware must consider its specific characteristics. This requires novel training algorithms and accurate hardware models, since they cannot be abstracted as a general-purpose computing platform. In this work, we present a bottom-up approach to training neural networks for hardware-based spiking neurons and synapses, built using ferroelectric capacitors (FeCAPs) and resistive random-access memories (RRAMs), respectively. Unlike the common approach of designing hardware to fit abstract neuron or synapse models, we start with compact models of the physical device to model the computational primitives. Based on these models, we have developed a training algorithm (BRUNO) that can reliably train the networks, even when applying hardware limitations, such as stochasticity or low bit precision. We analyse and compare BRUNO with Backpropagation Through Time. We test it on different spatio-temporal datasets. First on a music prediction dataset, where a network composed of ferroelectric leaky integrate-and-fire (FeLIF) neurons is used to predict at each time step the next musical note that should be played. The second dataset consists on the classification of the Braille letters using a network composed of quantised RRAM synapses and FeLIF neurons. The performance of this network is then compared with that of networks composed of LIF neurons. Experimental results show the potential advantages of using BRUNO by reducing the time and memory required to detect spatio-temporal patterns with quantised synapses.

Human Action Recognition (HAR) from video streams has many possible uses in areas like healthcare, surveillance, and human-computer interaction. The original purpose of video compression methods like SPIHT and others was to work with pixel-level quality measurements like PSNR and SSIM. These indicators have nothing to do with how well recognition works. In this paper, we present a Task-Aware Progressive SPIHT Framework that prioritises spatio-temporal data critical to actions during compression. By combining efficient pose estimation algorithms with lightweight motion and posture cues from opticalflow magnitude maps, you can make a significance mask that shows the areas that are most important for understanding action. We present a 3D Temporal-Priority SPIHT method that utilises motion-based dependencies among video frames, alongside spatial and temporal dependencies. Additional-ly, a Policy-Gradient-based Bit-Dropping method and Weighted Significance Testing are used to dynamically give bits to coefficients that are more important for the skeleton and motion while hiding background information that isn't important. Experimental tests show that the proposed framework works well for video analytics applications that need to work in real time and have limited resources. It greatly improves action detection accuracy at low bitrates while keeping compression efficiency competitive.

Towards efficient and accurate spiking neural networks via adaptive bit allocation.

In this article we propose a novel algorithm for lossless image compression based on 8 8 px block division and vector quantization.Each block is assigned to one of k classes and encoded using linear predictor assigned to a class (calculated using the Iterative Reweighted Least Squares (IRLS) algorithm).Several methods for initializing classes based on binary division of a set are proposed.Their advantages and disadvantages are shown, and the best performing ones are selected in order to develop an original dictionary initialization algorithm, which is used in the vector quantization method adapted to the lossless image compression.The hierarchical, binary tree-based initialization method is a combination of these algorithms, in which class initialization procedure follows the pattern of a complete binary tree with k leaves.The proposed initialization method significantly reduced time required for the main vector quantization process.Proposed codec belongs to the time-asymmetric compression methods with a short decoding time and is characterized by high compression efficiency, offering on average a 7.22% lower bit average compared to JPEG-LS.

Text-guided diffusion-based restoration of extremely compressed backgrounds for VCM

This paper proposes a new deep learning estimation algorithm of spectrum sensing and channel estimation in Cognitive Radio Sensor Networks (CRSNs) using Non-Orthogonal Multiple Access (NOMA). The given model will combine Long Short-Term Memory (LSTM) Networks with Bayesian Neural Networks (BNNs) to improve the work of the system in dynamic and unpredictable wireless conditions. The LSTM networks are used to predict with accuracy complex-valued Rayleigh Fading Channels and Bayesian is used to model the uncertainty with which such predictions are made. Also, a parallel Bayesian LSTM spectrum sensing model classifies activity of primary users (PU) to provide intelligent spectrum access and reduce interference. Prediction and spectrum sensing: Prediction and spectrum sensing is possible with the model in real-time and this is important in efficient spectrum management in CRSN where the Mean Absolute Error (MAE) of channel estimation is brought to under 0.02, implying high accuracy of channel condition prediction. Results of simulation demonstrate a significant enhancement when compared with traditional systems. The maximum accuracy of the spectrum sensing in the model is 98% at Signal to Noise Ratio of 10 dB and also a low Bit Error Rate (BER). LSTM combined with Bayesian inference structuring enables a combination of accurate channel estimation and trusted spectrum sensing, which are significant in terms of accuracy and the quantification of the uncertainties. These findings indicate the possibility of the proposed Bayesian-enhanced LSTM model to enhance the CRSN performance, especially in a low SNR and high-interference environment. This method is superior to the traditional models, which is a guarantee of the stable communication and spectrum utilization in the complicated wireless settings.

As the exponential growth in advanced compute workloads drives intra-datacenter interconnects to ever increasing bitrates, optical networking equipment has risen to the challenge by shifting from NRZ signaling to bandwidth efficient modulation methods such as PAM4. As these modulation schemes introduce an inherent SNR penalty, maintaining low bit error rates (BER) forces optical links to operate at significantly higher optical powers. However, increasing the optical power leads to photodetectors reaching one of their fundamental bottlenecks caused by the space-charge effect, limiting their ability to provide a high-speed response under high-power illumination. This work presents the design, fabrication, and characterization of a waveguide-integrated photodiode with dual optical inputs (DIPD) designed to overcome this limitation. Specifically, we demonstrate that combining a dual-fed architecture with targeted cross-sectional geometric optimizations effectively distributes the photocurrent density to delay the onset of space-charge saturation. Experimental validation demonstrates a high responsivity of ≈0.91 [A/W] (for O-band wavelengths) and a large electro-optic bandwidth (EOBW) of ≈58 [GHz], all under high-power illumination and CMOS driving voltages.

We propose Distribution-Preserving Latent Steganography via Conditional Optimal Transport (DPL-COT), a coverless image steganography framework for latent diffusion models. Unlike classical cover-modifying schemes, DPL-COT embeds a bitstream directly into the initialization noise latent zT∼N(0,I) without model retraining. Our primary objective is high recoverability and a low bit error rate (BER) under deterministic inversion, which is inherently imperfect due to numerical discretization and VAE nonlinearity. To maximize decoding stability, we restrict embedding to the natural tails of the latent prior by selecting the largest-magnitude coordinates, thereby increasing the sign decision margin against inversion drift. To preserve distributional stealth, per-bit target values are analytically derived from truncated Gaussians matching the marginal distribution of the selected coordinates. Conditional 1D optimal transport is applied independently for each bit class, mapping every coordinate to its target value while preserving rank order. We generate 5000 stego images using a pretrained diffusion model and demonstrate a favorable capacity–reliability trade-off (e.g., 4916 bits/image with 0.473% mean BER) and strong robustness to JPEG compression (sub-1% mean BER at Q=60). Compared with LDStega, a recent LDM-based scheme reporting 99.28% clean-channel accuracy, DPL-COT achieves 99.53% at a comparable operating point and sustains above-99% accuracy under all tested JPEG quality factors. Latent-space tests further confirm negligible cover–stego distribution shift (mean KS2<0.003, mean W1<0.003), a property not formally addressed by prior methods.

Abstract In this paper, a fiber-wireless optical communication system is designed and investigated under foggy weather conditions via OptiSystem ver. 7. Wavelength division multiplexing (WDM) is used to enhance the capacity of the data transmission. The WDM technique was used to transmit data over the Fi-Wi link. Fog poses an obstacle to terrestrial optical communications using FSO technology because fog particles weaken the signal and scatter light during transmission. The simulation proceeded under foggy conditions. The simulation system is designed with eight channels capable of transmitting data at 5 Gb/s for each channel. The frequencies were selected (193.1, 193.2, 193.3, 193.4, 193.5, 193.6, 193.7, and 193.8 THz) as carrier waves. The system performance is tested in terms of Q-factor and BER. The simulation system is a mimic to compare the power source and data rates. The system is characterized by a low bit error rate and a high quality factor at maximum power. The system was implemented by sending data at rates of 1, 2.5, and 5 Gb/s. It was observed that the Q-factor decreases with an increase in the transmission rate. A comparison was made between RZ and NRZ pulse generating. It was noticed that when using RZ, the Q-factor increases and the bit error rate decreases.

We introduce LadderPrime, an exception-free scalar-point multiplication algorithm, which works on the Kummer line of an elliptic curve given by the equation B*y^2=x^3+A*x^2+ax+b. LadderPrime operates only on two coordinates and computes the correct output for all input points, all scalars, and all elliptic curves of characteristic > 2. This is achieved by an alternative set of differential-addition formulas which can handle not only generic points but also the problematic point (0 : Z) for Montgomery ladder. The main structure of LadderPrime is analogous to the scalar-point multiplication in Bernstein’s X25519 Diffie-Hellman key exchange (DH) specification. Unlike, X25519 which uses the non-prime order (h = 8) elliptic Curve25519, LadderPrime is able to work with prime order (h = 1) (and non-prime order) elliptic curves. When used with a prime order elliptic curve, LadderPrime does not need the initial raising of base point to a prime order subgroup. In other words, LadderPrime eliminates the need for masking lower bits of the scalar. LadderPrime also eliminates the need for Hamburg’s "Decaf" (CRYPTO 2015) and later refined "Ristretto" methods. Essential cryptographic protocols such as DH and qDSA can be instantiated over LadderPrime.

Satellites supporting IoT connectivity may need to serve a large population of LoRa terminals, where collisions among packets using the same spreading factor (SF) can severely degrade uplink reliability. The ALOHA-based access used in LEO-IoT leads to frequent collisions under massive terminal access, which limits system capacity. Conventional signal separation methods that rely on the capture effect typically require a sufficiently large power difference between colliding signals. However, due to the channel characteristics of LEO links, this condition is often difficult to satisfy. We propose ConvLoRa, a collision demodulation method for co-SF LoRa uplink signals in LEO-IoT based on a fully convolutional neural network (FCN). To improve robustness to synchronization deviations, ConvLoRa uses an up-chirp in the preamble as a reference for feature matching, and employs data augmentation to emulate synchronization deviations during training. In addition, a multi-task design is adopted to estimate the payload length with minimal introduction of extra network parameters. Experiments show that ConvLoRa achieves lower demodulation bit error rate (BER) under collision conditions compared with baselines, including CoRa and SIC-based receivers. Under the condition of a two-signal collision with SNR = -9 dB and SF = 8, the BER of the proposed method is 21% that of CoRa and 28% that of the SIC-based method.

Directional modulation (DM) has emerged as a promising physical-layer security technique for next-generation wireless communication systems. This work presents a breakthrough photonic-assisted DM architecture that simultaneously resolves the fundamental constraints in conventional implementations, including limited operating frequency range, low symbol rate and sluggish phase shifter switching rates. By carefully controlling the biases of the dual-parallel Mach-Zehnder modulator (DPMZM), as well as the amplitude of the encoded signals, high-quality physical layer security communication in a specific direction is achieved. The verification experiment demonstrates a wideband directional binary phase shift keying (BPSK) modulation with a widely tunable operating frequency from 1 to 40 GHz and a high symbol rate of 2 GS/s. The experiment in 40 GHz is conducted via numerical simulations, whereas the actual hardware implementation and measurements are conducted up to 10 GHz to validate performance under realistic channel conditions. The desired direction receiver can successfully recover the symbol information with a low bit error rate (BER) approaching 10-7. In contrast, the undesired direction exhibits low power, serious phase distortion, and high BER, making it difficult to extract useful information. The innovative directional modulation technique holds substantial promise for advanced applications in various fields, including high-speed secure military communications and satellite communications.

Recent advancements in diffusion-based generative priors have enabled visually plausible image compression at extremely low bit rates. However, existing approaches suffer from slow sampling processes and suboptimal bit allocation due to fragmented training paradigms. In this work, we propose Accelerate Diffusion-based Image Compression via Consistency Prior Refinement (DiffCR), a novel compression framework for efficient and high-fidelity image reconstruction. At the heart of DiffCR is a Frequency-aware Skip Estimation (FaSE) module that refines the epsilon-prediction prior from a pre-trained latent diffusion model and aligns it with compressed latents at different timesteps via Frequency Decoupling Attention (FDA). Furthermore, a lightweight consistency estimator enables fast two-step decoding by preserving the semantic trajectory of diffusion sampling. Without updating the backbone diffusion model, DiffCR achieves substantial bitrate savings (27.2% BD-rate(LPIPS) and 65.1% BD-rate(PSNR)) and over 10 times speed-up compared to SOTA diffusion-based compression baselines.

A lightweight model for perceptual image compression via implicit priors.

Both laser and write current influences write-ability and areal density capability (ADC) in heat-assisted magnetic recording (HAMR). Sufficiently high laser and write current is required to achieve a high proportion of magnetization switching of media grains, as measured by THMap, and obtain low bit error rate. However, increasing these currents also lead to a wider write width, resulting in a loss of ADC. This study compares these conditions necessary to achieve saturation of magnetization switching and those for maximizing ADC. ADC increases with write current up to the level required for saturation of magnetization switching. In contrast, higher laser current than that necessary for saturation of magnetization switching is required to maximize ADC. This is because increasing the laser current enhances the down-track thermal gradient and reduces bit-to-bit transition jitter. This finding suggests that achieving saturation of magnetization switching is a necessary condition for maximizing ADC. Furthermore, optimizing the laser current to reduce jitter plays a significant role in improving ADC.

Abstract In this paper, we present a novel underwater acoustic (UWA) communication system based on a multiple chirp rate shift keying with index modulation (MCrSK-IM) scheme and a convolutional neural network (CNN) based demodulator. The proposed system is designed to mitigate the problems of low data transmission rate and high bit error rate (BER) caused by severe multiple path propagation and Doppler effects in UWA channels. At the transmitter, the chirp-indexed modulation approach maps input bits to specific combinations of up and down chirp rates, improving spectral efficiency without increasing bandwidth. At the receiver, a trained CNN model performs adaptive feature extraction and accurate symbol classification, effectively replacing the traditional Fractional Fourier Transform, Matched Filter and Multilayer Perceptron detection scheme. Simulation results obtained from the static BELLHOP channel and the dynamic NOF1 channel from the Watermark UWA channel set demonstrate that the proposed MCrSK-IM-CNN system achieves reliable underwater data transmission at 2000 bit/s with a BER lower than 10 −3 under UWA channel conditions.

Abstract To address the issue of quality degradation during high-definition video online transmission using traditional coding and decoding technologies under bandwidth constraints or high-resolution conditions, we propose the VidSemFlow model—based on Stable Diffusion for video semantic transmission—and constructed an online course video semantic transmission platform. The model converts video frames into semantic prompt vectors for transmission. The model incorporates three key innovations: gradient descent prompt fitting to ensure pixel-level reconstruction alignment, low-rank decomposition for dynamic bitrate control, and temporal prompt interpolation to minimize inter-frame redundancy. Experiments demonstrate that, at the same level of perceptual quality, this transmission model reduces bandwidth requirements by over four times compared to traditional H.265 and Variational Auto-Encoders(VAE) approaches. At extremely low bitrates, it achieves superior performance in the Learning Perceptual Image Patch Similarity(LPIPS) metric, with a substantial decrease in the proportion of severely distorted frames, and realizes real-time generation at over 150 Frames Per Second(FPS)—thus establishing a new paradigm for efficient video transmission.

Quantum key distribution (QKD) can be used to establish a secret key between trusted parties. Many practical use-cases in communication networks, however, involve parties who do not trust each other. A fundamental cryptographic building block for such distrustful scenarios is quantum coin flipping, which has been investigated only in few experimental studies to date, all of which used probabilistic quantum light sources imposing fundamental limitations. Here, we experimentally implement a quantum strong coin flipping protocol using single-photon states and demonstrate a quantum advantage compared to both classical realizations and implementations using faint laser pulses. We achieve this by employing a state-of-the-art deterministic quantum dot light source in combination with fast, random polarization-state encoding enabling sufficiently low quantum bit error ratio. By demonstrating a single-photon quantum advantage in a cryptographic primitive beyond QKD, our work represents a major advance towards the implementation of complex cryptographic tasks in a future quantum internet.

With the rapid development of electronic technology and increasing requirements on signal utilization efficiency and low bit error rates, diverse signal modulation schemes have emerged to accommodate diverse scenarios, including electronic countermeasures and military competition. The growing complexity of modulation schemes has made modulation recognition increasingly challenging, positioning deep learning-based automatic modulation recognition as a major research focus. This study uses the RadioML 2018.01A dataset from Kaggle to evaluate modulation recognition performance under signal-to-noise ratios ranging from 0 dB to 12 dB. A conventional convolutional neural network (CNN) and a residual neural network (ResNet), both commonly employed in image classification, are adopted and compared. The objective is to identify the modulation scheme used by the signals in the dataset and classify them into the corresponding categories. For certain modulation schemes, such as 64QAM and 128QAM, the residual neural network achieves classification accuracies that are up to approximately 20% higher than those of the convolutional neural network. Although the recognition accuracy of ResNet for some modulation schemes is slightly lower than that of the CNN, it remains within an acceptable range. Experimental results demonstrate that under medium signal-to-noise ratio conditions, the residual neural network outperforms the convolutional neural network.

Conventional chaotic communication relies on high-quality chaos synchronization, but this is difficult to achieve in practice due to the extreme parameter sensitivity of communication systems based on chaos synchronization. Unlike previous works, the proposed scheme utilizes a source separation approach to decrypt the message at the receiver end by separating it directly from chaos at the receiver side. In this paper, we take Conv-TasNet as an example to illustrate the feasibility of the separation scheme and achieve decryption at a low bit error rate(BER). The robustness of the separation performance to noise (demonstrated as signal-to-noise ratio (SNR)), masking coefficient, bit rate of the message, and parameters of the chaotic transmitter are studied. Simulation results show that the proposed system has low sensitivity to time delay signature (TDS). Simultaneously, the bit rate of the message, the overall gain of the electro-optical feedback loop, the high-pass cutoff frequency, and the masking coefficient could significantly affect the separation performance.