Search

The task of identification of sound events in a particular surrounding is known as Sound Event Detection (SED) or Acoustic Event Detection (AED). The occurrence of sound events is unstructured and also displays wide variations in both temporal structure and frequency content. Sound events may be non-overlapped (monophonic) or overlapped (polyphonic) in nature. In real-time scenarios, polyphonic SED is most commonly seen as compared to monophonic SED. In this paper, a Mel-Pseudo Constant Q-Transform (MP-CQT) technique is introduced to perform polyphonic SED to effectively learn both monophonic and polyphonic sound events. A pseudo CQT technique is adapted to extract features from the audio files and their Mel spectrograms. The Mel-scale is believed to broadly simulate human perception system. The classifier used is a Convolutional Recurrent Neural Network (CRNN). Comparison of the performance of the proposed MP-CQT technique along with CRNN is presented and a considerable performance improvement is observed. The proposed method achieved an average error rate of 0.684 and average F1 score of 52.3%. The proposed approach is also analyzed for the robustness by adding an additional noise at different Signal to Noise Ratios (SNRs) to the audio files. The proposed method for SED task has displayed improved performance as compared to state-of-the-art SED systems. The introduction of new feature extraction technique has shown promising improvement in the performance of the polyphonic SED system. [ABSTRACT FROM AUTHOR]

Classification is one of the most well-known data mining branches used in diverse domains and fields. In the literature, many different classification techniques, such as statistical/intelligent, linear/nonlinear, fuzzy/crisp, shallow/deep, and single/hybrid, have been developed to cover data and systems with different characteristics. Intelligent classification approaches, especially deep learning classifiers, due to their unique features to provide accurate and efficient results, have recently attracted a lot of attention. However, in the learning process of the intelligent classifiers, a continuous distance-based cost function is used to estimate the connection weights, though the goal function in classification problems is discrete and using a continuous cost function in their learning process is unreasonable and inefficient. In this paper, a novel discrete learning–based methodology is proposed to estimate the connection weights of intelligent classifiers more accurately. In the proposed learning process, they are discretely adjusted and at once jumped to the target. This is in contrast to conventional continuous learning algorithms in which the connection weights are continuously adjusted and step by step near the target. In the present research, the proposed methodology is exemplarily applied to the deep neural network (DNN), which is one of the most recognized deep classification approaches, with a solid mathematical foundation and strong practical results in complex problems. Although the proposed methodology is just implemented on the DNN, it is a general methodology that can be similarly applied to other shallow and deep intelligent classification models. It can be generally demonstrated that the performance of the proposed discrete learning–based DNN (DIDNN) model, due to its consistency property, will not be worse than the conventional ones. The proposed DIDNN model is exemplarily evaluated on some well-known cancer classification benchmarks to illustrate the efficiency of the proposed model. The empirical results indicate that the proposed model outperforms the conventional versions of the selected deep approach in all data sets. Based on the performance analysis, the DIDNN model can improve the performance of the classic version by approximately 3.39%. Therefore, using this technique is an appropriate and effective alternative to conventional DNN-based models for classification purposes. [ABSTRACT FROM AUTHOR]

We study the stability of accuracy during the training of deep neural networks (DNNs). In this context, the training of a DNN is performed via the minimization of a cross-entropy loss function, and the performance metric is accuracy (the proportion of objects that are classified correctly). While training results in a decrease of loss, the accuracy does not necessarily increase during the process and may sometimes even decrease. The goal of achieving stability of accuracy is to ensure that if accuracy is high at some initial time, it remains high throughout training. A recent result by Berlyand, Jabin, and Safsten introduces a doubling condition on the training data, which ensures the stability of accuracy during training for DNNs using the absolute value activation function. For training data in R n , this doubling condition is formulated using slabs in R n and depends on the choice of the slabs. The goal of this paper is twofold. First, to make the doubling condition uniform, that is, independent of the choice of slabs. This leads to sufficient conditions for stability in terms of training data only. In other words, for a training set T that satisfies the uniform doubling condition, there exists a family of DNNs such that a DNN from this family with high accuracy on the training set at some training time t 0 will have high accuracy for all time t > t 0 . Moreover, establishing uniformity is necessary for the numerical implementation of the doubling condition. We demonstrate how to numerically implement a simplified version of this uniform doubling condition on a dataset and apply it to achieve stability of accuracy using a few model examples. The second goal is to extend the original stability results from the absolute value activation function to a broader class of piecewise linear activation functions with finitely many critical points, such as the popular Leaky ReLU. [ABSTRACT FROM AUTHOR]

Deep neural networks (DNNs) involving massive neural nodes grouped into different neural layers have been a promising innovation for function approximation and inference, which have been widely applied to various vertical applications such as image recognition. However, the computing burdens to train a DNN model with a limited latency may not be affordable for the user equipment (UE), which consequently motivates the concept of splitting the computations of DNN layers to not only the edge server but also the cloud platform. Despite the availability of more computing resources, computing tasks with such split computing also suffer packet transmission unreliability, latency, and significant energy consumption. A practical scheme to optimally split the computations of DNN layers to the UE, edge, and cloud is thus urgently desired. To solve this optimization, we propose a multi-tier deep reinforcement learning (DRL) scheme for the UE and edge to distributively determine the splitting points to minimize the overall training latency while meeting the constraints of overall energy consumption and image recognition accuracy. The performance evaluation results show the outstanding performance of the proposed design as compared with state-of-the-art schemes, to fully justify the practicability in the next-generation smart networks. [ABSTRACT FROM AUTHOR]

The advent of autonomous vehicles has heralded a transformative era in transportation, reshaping the landscape of mobility through cutting-edge technologies. Central to this evolution is the integration of artificial intelligence (AI), propelling vehicles into realms of unprecedented autonomy. Commencing with an overview of the current industry landscape with respect to Operational Design Domain (ODD), this paper delves into the fundamental role of AI in shaping the autonomous decision-making capabilities of vehicles. It elucidates the steps involved in the AI-powered development life cycle in vehicles, addressing various challenges such as safety, security, privacy, and ethical considerations in AI-driven software development for autonomous vehicles. The study presents statistical insights into the usage and types of AI algorithms over the years, showcasing the evolving research landscape within the automotive industry. Furthermore, the paper highlights the pivotal role of parameters in refining algorithms for both trucks and cars, facilitating vehicles to adapt, learn, and improve performance over time. It concludes by outlining different levels of autonomy, elucidating the nuanced usage of AI algorithms, and discussing the automation of key tasks and the software package size at each level. Overall, the paper provides a comprehensive analysis of the current industry landscape, focusing on several critical aspects. [ABSTRACT FROM AUTHOR]

Detecting drones is a complex challenge, primarily due to their small feature size for extraction and variable lighting conditions. It is crucial to effectively capture and model features for drone detection. To accurately detect drones, we propose feature capture and modeling modules. The feature capture module has a minimal number of FLOPs and parameters. It consists of both local and global attention branches, which capture contextual information and global dependencies across the entire feature set. Complementing this, our feature modeling module innovatively calculates attention maps without any additional parameters. This module augments the capability of the feature capture mechanism to represent complex patterns more effectively. Finally, to ensure rapid deployment, we convert the proposed models to machine codes by introducing a compiler, accelerating drone detection. The compiler unifies inter- and intra-operator scheduling with task abstraction. It optimizes the codes for hardware. In compiling time, the effective schedule is performed. The compilation ensures that drone detection is real-time and accuracy remains unchanged. Through rigorous testing, our methods have demonstrated superiority over most current ones in several metrics, including accuracy, parameter quantities, FLOPs, average FPS, visual effects, and latency. Our method yields at least 23.5% and 12.47% higher $AR_{M}$ than existing methods on DUT-Anti-UAV and Online Drone datasets. Our inference speed is at least 6.49% higher than other methods on NVIDIA RTX 2080 Ti GPU.

The expansion of small satellite networks in earth's orbit has resulted in a plethora of earth optical imagery available to the civil, defense, and commercial sectors. Small satellites (less than 1000 kg in mass) and their constellations can be delivered rapidly and at low cost and are more difficult to target by adversaries—a key consideration in the defense industry. Yet, small satellite size constraints often result in reduced payload capacity, reduced power capacity, or loss of redundancy. Traditionally, the cost of an optical telescope on board a satellite scales at roughly the square of the aperture, meaning that it costs four times as much to double the resolution of the imaging hardware. However, deep learning has shown considerable success in the areas of super-resolution and enhancing the pixel resolution of optical imagery. These deep learning methods have the potential to provide optical resolution capabilities rivaling larger satellites and their telescopes, while maintaining the benefits of small satellites—smaller physical size (which lowers launch vehicle costs and provides a basis for large constellations), reduced manufacturing time, and lower manufacturing costs. By providing low-cost small satellite platforms with the same capabilities as larger satellites, the cost for high-resolution in-orbit optical imagery is reduced alongside time to orbit. In this work, we detail a deep-learning-based approach, which improves optical satellite imagery to five times the original pixel-based resolution without the need or expense of increasing the capabilities of the imager through larger telescope apertures. The approach—demonstrated on Terran Orbital's GEOStare SV2 mission imagery—is generally applicable to any optical satellite image and is agnostic to the mission, satellite manufacturer, optical payload specifications, or data source. This capability provides a basis for small satellite missions and constellations—and their optical payloads—to rival the native hardware-based resolutions available through larger satellites with wider telescope apertures at a significantly reduced cost.

Efficient Unmanned Aerial Vehicle (UAV) trajectory generation is crucial for successful area coverage missions, aiming to maximize coverage while minimizing resource consumption. In this research, we present a comprehensive study on optimizing UAV trajectory generation using Deep Neural Networks (DNNs) with the Adam optimization algorithm. The DNNs are trained on historical data to produce smooth and continuous trajectories, thereby reducing abrupt changes in direction and enhancing overall efficiency during the mission. To evaluate the performance of the proposed approach, we conducted experiments comparing different activation functions, namely tanh, sigmoid, and ReLU, with the Adam-optimized DNN model. The trajectories generated by each activation function were analyzed using key metrics, including Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-squared (R2) scores for both X and Y coordinates. The results of the comparative analysis revealed that the DNN model with the Adam optimizer exhibited superior performance over the other activation functions. It achieved lower MSE, MAE, and RMSE values, indicating better trajectory accuracy and smoother paths. Additionally, the R2 scores demonstrated a higher correlation between the generated trajectories and the actual trajectories, highlighting the model's ability to capture underlying patterns effectively. The findings underscore the significance of leveraging the Adam-optimized DNN approach for UAV trajectory planning, offering promising opportunities for resource optimization, increased mission success, and further advancements in autonomous aerial systems. This research contributes to the ongoing efforts in UAV path planning, optimization, and intelligent control strategies, paving the way for enhanced autonomous systems in various real-world applications. [ABSTRACT FROM AUTHOR]

Computer vision applications like traffic monitoring, security checks, self-driving cars, medical imaging, etc., rely heavily on machine learning models. It raises an essential growing concern regarding the dependability of machine learning algorithms, which cannot be entirely trusted due to their fragile nature. This leads us to a dire need for systematic analysis of adversarial settings in neural networks. Hence, this article presents a comprehensive study of vulnerabilities, possible attacks such as data poisoning and data access during training, evasion, and oracle attacks at the test time, and their defensive and preventive measures using novel taxonomies. The survey has covered the complete scenario where an adversary can make malicious manipulations and elaborated more on the most potent threat, i.e., test time evasion attack using an adversarial image (maliciously perturbed image). It expounds an intuition behind generating an adversarial image, covering all relevant adversarial attack algorithms and strategies for increasing robustness against adversarial images. The existence and effect of adversarial images, as well as their transferability, are also examined. The article guides the reader with an approach on building new models to enhance their reliability. Additionally, the survey presents the procedures that still demand further exploration with limitations in existing methods, enhancing future research directions. [ABSTRACT FROM AUTHOR]

Construction materials undergo appearance and textural changes during the construction process. Accurate recognition of these changes is critical for effectively understanding the construction status; however, recognizing the various levels of detailed material conditions is not sufficiently explored. The primary challenge in the detailed recognition of the conditions of the material is the availability of labeled training data. To address this challenge, this study proposes a novel state-of-the-art deep learning model that leverages transfer learning, utilizing the pretrained Inception V3 to transfer knowledge to the limited labeled data set in the construction context. This enables the model to learn meaningful representations from the limited training data, enhancing its ability to accurately classify material conditions. In addition, gray-level co-occurrence matrix (GLCM)–based texture features are extracted from the images to capture the appearance and textural changes in construction materials, which are then concatenated with the transferred convolutional neural network (CNN) features to create a more comprehensive representation of the material conditions. The proposed model achieved an overall classification accuracy of 95% and 71% with limited (208 images) and very small (70 images) data sets, respectively. It outperformed different experimental architectures, including CNN models developed using limited data with and without augmentation, CNN model with data augmentation and transfer learning, separate models using local binary pattern (LBP) and GLCM texture features with super learners trained using augmented limited data. The findings suggest that the proposed model, which combines transfer learning with GLCM-based texture features, is effective in accurately recognizing the conditions of construction materials, even with limited labeled training data. This can contribute to improved construction management and monitoring. There are several practical applications of the proposed model combining CNN architecture with GLCM textures in construction industry. The primary applications are in activities such as quality inspection and progress monitoring. By analyzing images of different material conditions, such as concrete, plaster, or masonry walls, the model can be used to automatically detect defects or inconsistencies. This enables construction practitioners to ensure the quality of their structures, detect issues early on, and make informed decisions for maintenance and repair. Additionally, the model can be used to monitor the progress of construction projects by analyzing images to track the status and completion of different building components to estimate delays or cost overruns comparing with the expected material condition at a given time of the schedule. Moreover, because the model does not depend on large training data set, it enables construction managers to develop their project specific data sets and automate material condition detection and monitoring. [ABSTRACT FROM AUTHOR]

Human capability to perform routine tasks declines with age and age-related problems. Remote human activity recognition (HAR) is beneficial for regular monitoring of the elderly population. This paper addresses the problem of the continuous detection of daily human activities using a mm-wave Doppler radar. In this study, two strategies have been employed: the first method uses un-equalized series of activities, whereas the second method utilizes a gradient-based strategy for equalization of the series of activities. The dynamic time warping (DTW) algorithm and Long Short-term Memory (LSTM) techniques have been implemented for the classification of un-equalized and equalized series of activities, respectively. The input for DTW was provided using three strategies. The first approach uses the pixel-level data of frames (UnSup-PLevel). In the other two strategies, a convolutional variational autoencoder (CVAE) is used to extract Un-Supervised Encoded features (UnSup-EnLevel) and Supervised Encoded features (Sup-EnLevel) from the series of Doppler frames. The second approach for equalized data series involves the application of four distinct feature extraction methods: i.e., convolutional neural networks (CNN), supervised and unsupervised CVAE, and principal component Analysis (PCA). The extracted features were considered as an input to the LSTM. This paper presents a comparative analysis of a novel supervised feature extraction pipeline, employing Sup-ENLevel-DTW and Sup-EnLevel-LSTM, against several state-of-the-art unsupervised methods, including UnSUp-EnLevel-DTW, UnSup-EnLevel-LSTM, CNN-LSTM, and PCA-LSTM. The results demonstrate the superiority of the Sup-EnLevel-LSTM strategy. However, the UnSup-PLevel strategy worked surprisingly well without using annotations and frame equalization.

Human capability to perform routine tasks declines with age and age-related problems. Remote human activity recognition (HAR) is beneficial for regular monitoring of the elderly population. This paper addresses the problem of the continuous detection of daily human activities using a mm-wave Doppler radar. In this study, two strategies have been employed: the first method uses un-equalized series of activities, whereas the second method utilizes a gradient-based strategy for equalization of the series of activities. The dynamic time warping (DTW) algorithm and Long Short-term Memory (LSTM) techniques have been implemented for the classification of un-equalized and equalized series of activities, respectively. The input for DTW was provided using three strategies. The first approach uses the pixel-level data of frames (UnSup-PLevel). In the other two strategies, a convolutional variational autoencoder (CVAE) is used to extract Un-Supervised Encoded features (UnSup-EnLevel) and Supervised Encoded features (Sup-EnLevel) from the series of Doppler frames. The second approach for equalized data series involves the application of four distinct feature extraction methods: i.e., convolutional neural networks (CNN), supervised and unsupervised CVAE, and principal component Analysis (PCA). The extracted features were considered as an input to the LSTM. This paper presents a comparative analysis of a novel supervised feature extraction pipeline, employing Sup-ENLevel-DTW and Sup-EnLevel-LSTM, against several state-of-the-art unsupervised methods, including UnSUp-EnLevel-DTW, UnSup-EnLevel-LSTM, CNN-LSTM, and PCA-LSTM. The results demonstrate the superiority of the Sup-EnLevel-LSTM strategy. However, the UnSup-PLevel strategy worked surprisingly well without using annotations and frame equalization. [ABSTRACT FROM AUTHOR]

The prevalence of violence against women and children is concerning, and the initial step is to raise awareness of this issue. Certain forms of detection based techniques are not frequently regarded both socially and culturally permissible. Designing and implementing effective approaches in secondary and supplementary avoidance simultaneously depends on the characterization and assessment. Given the greater incidence of instances and mortalities resulting developing an early detection system is essential. Consequently, violence against women and children is a problem of human health of pandemic proportions. As a result, the focus of this survey is to analyze the existing methods used to identify violence in photos or films. Here, 50 research papers are reviewed and their techniques employed, dataset, evaluation metrics, and publication year are analyzed. The study reviews the potential future research areas by examining the difficulties in identifying violence against women and children in literary works for researchers to overcome in order to produce better results. [ABSTRACT FROM AUTHOR]

Power line inspection is one important task performed by electricity distribution network operators worldwide. It is part of the equipment maintenance for such companies and forms a crucial procedure since it can provide diagnostics and prognostics about the condition of the power line network. Furthermore, it helps with effective decision making in the case of fault detection. Nowadays, the inspection of power lines is performed either using human operators that scan the network on foot and search for obvious faults, or using unmanned aerial vehicles (UAVs) and/or helicopters equipped with camera sensors capable of recording videos of the power line network equipment, which are then inspected by human operators offline. In this study, we propose an autonomous, intelligent inspection system for power lines, which is equipped with camera sensors operating in the visual (Red–Green–Blue (RGB) imaging) and infrared (thermal imaging) spectrums, capable of providing real-time alerts about the condition of power lines. The very first step in power line monitoring is identifying and segmenting them from the background, which constitutes the principal goal of the presented study. The identification of power lines is accomplished through an innovative hybrid approach that combines RGB and thermal data-processing methods under a custom-made drone platform, providing an automated tool for in situ analyses not only in offline mode. In this direction, the human operator role is limited to the flight-planning and control operations of the UAV. The benefits of using such an intelligent UAV system are many, mostly related to the timely and accurate detection of possible faults, along with the side benefits of personnel safety and reduced operational costs. [ABSTRACT FROM AUTHOR]

In order to improve the performance of deep neural network (DNN) accelerators, it is necessary to optimize compute efficiency and operating frequency. However, the implementation of contemporary DNNs often requires excessive resources due to the heavy multiply-and-accumulate (MAC) computations. In this work proposes a MAC unit designed with a Co-ordinate Rotation DIgital Computer (CORDIC)-based architecture, which is both power and area-efficient for 8-bit and higher-bit precision. The CORDIC-based designs are typically associated with low throughput. To address this issue, a performance-centric pipelined architecture is investigated that increases throughput. The study conducts a detailed Pareto analysis of accuracy variation at different precision levels and required pipeline stages to achieve high performance. The proposed MAC unit's post-synthesis results at the 45nm technology node are provided, and performance is evaluated on a deep neural network using Vertex-7 FPGA board. The proposed fixed-point MAC architecture is scalable for all bit-precision and flexible for the decimal point implication. The study finds that the proposed Fixed Q 3.5 precision with five pipeline stage-based MAC shows better performance metrics compared to the recursive CORDIC-based MAC design. The proposed MAC design has a lower area-delay-product (ADP) which is 1.13 × , and higher throughput of 2.73 × compared to the recursive CORDIC-based MAC. The study evaluated the performance of the proposed MAC unit using the fully connected NN for the MNIST dataset and found that the throughput 1.89 × better compared to the conventional MAC-based design. [ABSTRACT FROM AUTHOR]

Copyright of Transactions of Nanjing University of Aeronautics & Astronautics is the property of Editorial Department of Journal of Nanjing University of Aeronautics & Astronautics and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

The advent of autonomous vehicles has heralded a transformative era in transportation, reshaping the landscape of mobility through cutting-edge technologies. Central to this evolution is the integration of artificial intelligence (AI), propelling vehicles into realms of unprecedented autonomy. Commencing with an overview of the current industry landscape with respect to Operational Design Domain (ODD), this paper delves into the fundamental role of AI in shaping the autonomous decision-making capabilities of vehicles. It elucidates the steps involved in the AI-powered development life cycle in vehicles, addressing various challenges such as safety, security, privacy, and ethical considerations in AI-driven software development for autonomous vehicles. The study presents statistical insights into the usage and types of AI algorithms over the years, showcasing the evolving research landscape within the automotive industry. Furthermore, the paper highlights the pivotal role of parameters in refining algorithms for both trucks and cars, facilitating vehicles to adapt, learn, and improve performance over time. It concludes by outlining different levels of autonomy, elucidating the nuanced usage of AI algorithms, and discussing the automation of key tasks and the software package size at each level. Overall, the paper provides a comprehensive analysis of the current industry landscape, focusing on several critical aspects.

Recent years have seen a significant amount of studies in the area of speech enhancement. This review looks at several speech improvement methods as well as Deep Neural Network (DNN) functions in speech enhancement. Speech transmissions are frequently distorted by ambient noise, background noise, and reverberations. There are processing methods, such as Short-time Fourier Transform, Short-time Autocorrelation, and Short-time Energy (STE), that can be used to enhance speech. To reduce speech noise, features such as the Mel-Frequency Cepstral Coefficients (MFCCs), Logarithmic Power Spectrum (LPS), and Gammatone Frequency Cepstral Coefficients (GFCCs) can be retrieved and input to a DNN. DNN is essential to speech improvement since it builds models using a lot of training data and evaluates the efficacy of the enhanced speech using certain performance metrics. Since the beginning of deep learning publications in 1993, a variety of speech enhancement methods have been examined in this study. This review provides a thorough examination of the several neural network topologies, training algorithms, activation functions, training targets, acoustic features, and databases that were employed for the job of speech enhancement and were gathered from various articles published between 1993 and 2022. [ABSTRACT FROM AUTHOR]

This study aims to develop a model for detecting heat stress in southern-type garlic using a multispectral snapshot camera. Raw snapshot images were obtained from garlic cloves during the garlic bulb enlargement period, capturing the visible (Vis) and near-infrared (NIR) regions. Image preprocessing was applied to obtain a 38-wavelength spectrum by combining a 16-wavelength image in the Vis region and a 22-wavelength image in the NIR region. These spectral data were then utilized to develop models, including PLS-DA, LS-SVM, DNN, and recurrence plots-based CNN (RP-CNN). On average, the LS-SVM model demonstrated the best performance in detecting heat stress during the garlic bulb enlargement period. This is attributed to the nonlinear nature of the spectral differences between groups caused by abiotic stress in garlic. The LS-SVM model is particularly effective at capturing such nonlinear relationships. Among the model images, LS-SVM yielded the best performance, followed by RP-CNN, DNN, and PLS-DA. Therefore, this study confirms the potential of snapshot-based multispectral imaging for measuring changes in garlic crops induced by high-temperature stress. [ABSTRACT FROM AUTHOR]

Backdoor attacks, which induce a trained model to behave as intended by an adversary for specific inputs, have recently emerged as a serious security threat in deep learning-based classification models. In particular, because a backdoor attack is executed solely by incorporating a small quantity of malicious data into a dataset, it poses a significant threat to authentication models, such as facial cognition systems. Depending on whether the label of the poisoned samples has been changed, backdoor attacks on deep learning-based face recognition methods are categorized into one of the two architectures: 1) corrupted label attack; and 2) clean label attack. Clean label attack methods have been actively studied because they can be performed without access to training datasets or training processes. However, the performance of previous clean label attack methods is limited in their application to deep learning-based face recognition methods because they only consider digital triggers with instance-specific characteristics. In this study, we propose a novel clean label backdoor attack, that solves the limitations of the scalability of previous clean label attack methods for deep learning-based face recognition models. To generate poisoned samples that are instance agnostic while including physical triggers, the proposed method applies three core techniques: 1) accessory injection; 2) optimization-based feature transfer; and 3) $N$ :1 mapping for generalization. From the experimental results under various conditions, we demonstrate that the proposed attack method is effective for deep learning-based face recognition models in terms of the attack success rate on unseen samples. We also show that the proposed method not only outperforms the recent clean label attack methods, but also maintains a comparable level of classification accuracy when applied to benign data.

Deep learning-based Hand Gesture Recognition (HGR) via surface Electromyogram (sEMG) signals have recently shown considerable potential for development of advanced myoelectric-controlled prosthesis. Although deep learning techniques can improve HGR accuracy compared to their classical counterparts, classifying hand movements based on sparse multichannel sEMG signals is still a challenging task. Furthermore, existing deep learning approaches, typically, include only one model as such can hardly extract representative features. In this paper, we aim to address this challenge by capitalizing on the recent advances in hybrid models and transformers. In other words, we propose a hybrid framework based on the transformer architecture, which is a relatively new and revolutionizing deep learning model. The proposed hybrid architecture, referred to as the Transformer for Hand Gesture Recognition (TraHGR), consists of two parallel paths followed by a linear layer that acts as a fusion center to integrate the advantage of each module. We evaluated the proposed architecture TraHGR based on the commonly used second Ninapro dataset, referred to as the DB2. The sEMG signals in the DB2 dataset are measured in real-life conditions from 40 healthy users, each performing 49 gestures. We have conducted an extensive set of experiments to test and validate the proposed TraHGR architecture, and compare its achievable accuracy with several recently proposed HGR classification algorithms over the same dataset. We have also compared the results of the proposed TraHGR architecture with each individual path and demonstrated the distinguishing power of the proposed hybrid architecture. The recognition accuracies of the proposed TraHGR architecture for the window of size 200ms and step size of 100ms are 86.00%, 88.72%, 81.27%, and 93.74%, which are 2.30%, 4.93%, 8.65%, and 4.20% higher than the state-of-the-art performance for DB2 (49 gestures), DB2-B (17 gestures), DB2-C (23 gestures), and DB2-D (9 gestures), respectively.

Deep learning has transformed data generation, particularly in creating synthetic sensor data. This capability is invaluable in fields like autonomous driving, robotics, and computer science. To achieve this, we train models using real data, enabling them to replicate sensor data closely. These models introduce variations and noise, enhancing diversity and realism. Prominent techniques, including generative adversarial networks (GANs), variational autoencoders (VAEs), and recurrent neural networks (RNNs), excel in generating synthetic sensor data. Our paper focuses on Autoregressive Convolutional Recurrent Neural Networks (CRNN) for Multivariate Time Series Prediction. We incorporate Denoising Autoencoders (DAE) to mimic real-world noise characteristics. Our model is trained and validated using Ultra Wide Band (UWB) and Ultra High-Frequency Radio Frequency Identification (UHF-RFID) sensor data. It integrates sensor measurements and diverse information sources to produce synthetic data complementing real-world data. While demonstrated with UHF-RFID and UWB sensors, these techniques extend to industrial automation, healthcare and environmental monitoring. While our methodology exhibits broad potential, we present practical demonstrations with UHF-RFID and UWB sensors. Our deep neural network model allows researchers to construct datasets for algorithm validation, eliminating the need for costly and time-consuming data collection.

As we advance towards 6G communication systems, the number of network devices continues to increase resulting in spectrum scarcity. With the help of Spectrum Sensing (SS), Cognitive Radio (CR) exploits the frequency spectrum dynamically by detecting and transmitting in underutilized bands. The performance of 6G networks can be enhanced by utilizing Deep Neural Networks (DNNs) to perform SS. This paper provides a detailed survey of several Deep Learning (DL) algorithms used for SS by classifying them as Multilayer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, combined CNN-LSTM architectures and Autoencoders (AEs). The works are discussed in terms of the input provided to the DL algorithm, data acquisition technique used, data pre-processing technique used, architecture of each algorithm, evaluation metrics used, results obtained, and comparison with standard SS detectors. This survey further provides an overview of traditional Machine Learning (ML) algorithms and simple Artificial Neural Networks (ANNs) while highlighting the drawbacks of conventional SS approaches for completeness. A description of some publicly available Radio Frequency (RF) datasets is included and the need for comprehensive RF datasets and Transfer Learning (TL) is discussed. Furthermore, the research challenges related to the use of DL for SS are highlighted along with potential solutions.

This paper proposes a memory-efficient deep neural network (DNN) framework-based symbol level precoding (SLP). We focus on a DNN with realistic finite precision weights and adopt an unsupervised deep learning (DL) based SLP model (SLP-DNet). We apply a stochastic quantization (SQ) technique to obtain its corresponding quantized version called SLP-SQDNet. The proposed scheme offers a scalable performance vs memory trade-off, by quantizing a scalable percentage of the DNN weights, and we explore binary and ternary quantizations. Our results show that while SLP-DNet provides near-optimal performance, its quantized versions through SQ yield $\sim 3.46\times $ and $\sim 2.64\times $ model compression for binary-based and ternary-based SLP-SQDNets, respectively. We also find that our proposals offer $\sim 20\times $ and $\sim 10\times $ computational complexity reductions compared to SLP optimization-based and SLP-DNet, respectively.

The microwave backscattering model is one of the most effective tools for surface soil moisture (SSM) inversion, which has strong theoretical support, but the inverse problem is difficult to solve. Advance in artificial intelligence offers possibilities to learn complex nonlinear relationships in a data-driven way, but it lacks physical mechanism. To combine the advantages of model-driven and data-driven methods, an SSM inversion approach that couples the AIEM-Oh model with deep neural networks (DNNs) was proposed in this study. DNNs with different inputs were trained with a large number of simulation data generated from the AIEM-Oh model, thus embedding physical mechanisms in the data-driven scheme. Two field experiments at different scales were carried out to evaluate the performances of the proposed approach over bare surfaces. The effects of polarization modes and prior knowledge of surface roughness on SSM inversion were explored, and the accuracy of the approach was compared with the existing methods. The results suggest that satisfactory accuracy was obtained by the proposed approach, the RMSE between the measured and estimated values of SSM was 0.03–0.04 cm3 · cm−3 with prior knowledge of soil roughness, and the RMSE was 0.08–0.10 cm3 · cm−3 without the prior soil roughness information. VV polarization was more sensitive to SSM over bare surfaces than VH polarization. Moreover, the approach showed stable performance in different experimental regions. The results demonstrate the capability and reliability of the coupled approach for SSM inversion over bare surfaces.

The outstanding performance of deep neural networks (DNNs) in multiple computer vision in recent years has promoted its widespread use in aerial image semantic segmentation. Nonetheless, prior research has demonstrated the high susceptibility of DNNs to adversarial attacks. This poses significant security risks when applying DNNs to safety-critical earth observation missions. As an essential means of attacking DNNs, data poisoning attacks destroy model performance by contaminating model training data, allowing attackers to control prediction results by carefully crafting poisoning samples. Toward building a more robust DNNs-based aerial image semantic segmentation model, in this study, we proposed a robust invariant feature enhancement network (RIFENet) that can resist data poisoning attacks and has superior semantic segmentation performance. The constructed RIFENet improves the resistance to poisoning attacks by extracting and enhancing robust invariant features. Specifically, RIFENet uses a texture feature enhancement module (T-FEM), structural feature enhancement module (S-FEM), global feature enhancement module (G-FEM), and multi-resolution feature fusion module (MR-FFM) to enhance the representation of different robust features in the feature extraction process to suppress the interference of poisoning samples. Experiments on several benchmark aerial image datasets demonstrate that the proposed method is more robust and exhibits better generalization than other state-of-the-art methods. [ABSTRACT FROM AUTHOR]

With the advent of deep neural networks (DNNs) in the last two decades, tremendous developments have been made in many fields, such as image classification/recognition, voice recognition, and action recognition. These advanced DNNs require large amounts of labeled data, whose collection is costly and requires great effort. In this paper, we provide a convergence method for DNNs to solve some of these difficulties. First, we consider how to create labeled data using a generative adversarial network (GAN), one DNN method, and add additional networks to improve the quality of generated data. Then, we propose a convergence method for the DNNs and use a three-step evaluation to confirm this approach and show how to use the automatically generated data for training. With the method proposed in this paper, we hope that the manual work of labeling data can be reduced for many DNN applications. [ABSTRACT FROM AUTHOR]

Recent years have seen deep learning (DL) architectures being leveraged for learning the nonlinear relationships across the parameters in seismic inversion problems in order to better analyse the subsurface, such as improved velocity model building (VMB). In this study, we focus on deep-learning-based inversion (DLI) for velocity model building, leveraging on a conditional generative adversarial network (PIX2PIX) with ResNet-9 as generator, as well as a comprehensive mathematical methodology for generating samples of multi-stratified heterogeneous velocity models for training the DLI architecture. We demonstrate that the proposed architecture can achieve state-of-the-art performance in reconstructing velocity models using only one seismic shot, thus reducing cost and computational complexity. We also demonstrate that the proposed solution is generalisable across linear multi-layer models, curved or folded structures, structures with salt bodies as well as higher-resolution structures built from geological images through quantitative and qualitative evaluation. [ABSTRACT FROM AUTHOR]

Profiting from the powerful feature extraction and representation capabilities of deep learning (DL), aerial image semantic segmentation based on deep neural networks (DNNs) has achieved remarkable success in recent years. Nevertheless, the security and robustness of DNNs deserve attention when dealing with safety-critical earth observation tasks. As a typical attack pattern in adversarial machine learning (AML), backdoor attacks intend to embed hidden triggers in DNNs by poisoning training data. The attacked DNNs behave normally on benign samples, but when the hidden trigger is activated, its prediction is modified to a specified target label. In this article, we systematically assess the threat of backdoor attacks to aerial image semantic segmentation tasks. To defend against backdoor attacks and maintain better semantic segmentation accuracy, we construct a novel robust generative adversarial network (RFGAN). Motivated by the sensitivity of human visual systems to global and edge information in images, RFGAN designs the robust global feature extractor (RobGF) and the robust edge feature extractor (RobEF) that force DNNs to learn global and edge features. Then, RFGAN uses robust global and edge features as guidance to obtain benign samples by the constructed generator, and the discriminator to obtain semantic segmentation results. Our method is the first attempt to address the backdoor threat to aerial image semantic segmentation by constructing the robust DNNs model architecture. Extensive experiments on real-world scenes aerial image benchmark datasets demonstrate that the constructed RFGAN can effectively defend against backdoor attacks and achieve better semantic segmentation results compared with the existing state-of-the-art methods. [ABSTRACT FROM AUTHOR]

In this work, we build a general piece-wise model to analyze data-parallel (DP) training costs of convolutional neural networks (CNNs) on clusters of GPUs. This general model is based on i) multi-layer perceptrons (MLPs) in charge of modeling the NVIDIA cuDNN/cuBLAS library kernels involved in the training of some of the state-of-the-art CNNs; and ii) an analytical model in charge of modeling the NVIDIA NCCL Allreduce collective primitive using the Ring algorithm. The CNN training scalability study performed using this model in combination with the Roofline technique on varying batch sizes, node (floating-point) arithmetic performance, node memory bandwidth, network link bandwidth, and cluster dimension unveil some crucial bottlenecks at both GPU and cluster level. To provide evidence of this analysis, we validate the accuracy of the proposed model against a Python library for distributed deep learning training. [ABSTRACT FROM AUTHOR]

One of the major focuses in the remote sensing community is the rapid processing of deep neural networks (DNNs) on very high resolution (VHR) aerial images. Few studies have investigated the acceleration of training and prediction by optimizing the architecture of the DNN system rather than designing a lightweight DNN. Parallel processing using multiple graphics processing units (GPUs) increases VHR image processing performance. It drives extremely large and frequent data transfers (input/output(I/O)) from random access memory (RAM) to GPU memory. As a result, the system bus congestion causes the system to hang, resulting in long latency in training/predicting. In this paper, we evaluate the causes of long latency and propose a space-to-speed (S2S) DNN system to overcome the aforementioned challenges. A three-level memory system aiming to reduce data transfer during system operation is presented. Distribution optimization with parallel processing was used to accelerate the training. Training optimizations on VHR images (such as hot-zone searching and image/ground truth queues for data saving) were used to train the VHR images efficiently. Inference optimization was performed to speed up prediction in the release mode. To verify the efficiency of the proposed system, we used aerial image labeling from the Institut National de Recherche en Informatique et en Automatique (INRIA) and benchmarks from the Massachusetts Institute of Technology Aerial Imagery for Roof Segmentation (MITAIRS) to test the system performance and accuracy. Without the loss of accuracy, the S2S system improved prediction speed on the testing dataset by eight GPUs in a normal setting in both the INRIA dataset (from 534 to 72 s) and the MITAIRS dataset (818 to 120 s). With the prediction in half-float (using float-16 data), an 8-GPU parallel processing increased the speed to 38 s in the INRIA dataset and 83 s in the MITAIRS dataset. In a pressure test, our proposed system operated on 18,000 images with a size of 5000 × 5000 from 18.2 to 1.8 h with the prediction in full-float (using float-32 data) and 43 min with the prediction in half-float, increasing the speed by a factor of 9.78 and 25.3, respectively, when compared to system runs without optimization. [ABSTRACT FROM AUTHOR]

Most existing works on cardiac echocardiography segmentation require a large number of ground-truth labels to appropriately train a neural network; this, however, is time consuming and laborious for physicians. Self-supervision learning is one of the potential solutions to address this challenge by deeply exploiting the raw data. However, existing works mainly exploit single type/level of pretext task. In this work, we propose fusion of the multi-level and multi-type self-generated knowledge. We obtain multi-level information of sub-anatomical structures in ultrasound images via a superpixel method. Subsequently, we fuse various types of information generated through multi-types of pretext tasks. In the end, we transfer the learned knowledge to our downstream task. In the experimental studies, we have demonstrated the prove the effectiveness of this method through the cardiac ultrasound segmentation task. The results show that the performance of our proposed method for echocardiography segmentation matches the performance of fully supervised methods without requiring a high amount of labeled data. • Fusing multi-level multi-type self-generated knowledge. • Extracting the multi-level knowledge via a superpixel-based algorithm. • Extracting the multi-type knowledge via three types of tasks. • Accurate cardiac segmentation without requiring a large number of labels. [ABSTRACT FROM AUTHOR]

In this study, neurocomputational models are proposed for the acquisition of radar-based microwave images of breast tumors using deep neural networks (DNNs) and convolutional neural networks (CNNs). The circular synthetic aperture radar (CSAR) technique for radar-based microwave imaging (MWI) was utilized to generate 1000 numerical simulations for randomly generated scenarios. The scenarios contain information such as the number, size, and location of tumors for each simulation. Then, a dataset of 1000 distinct simulations with complex values based on the scenarios was built. Consequently, a real-valued DNN (RV-DNN) with five hidden layers, a real-valued CNN (RV-CNN) with seven convolutional layers, and a real-valued combined model (RV-MWINet) consisting of CNN and U-Net sub-models were built and trained to generate the radar-based microwave images. While the proposed RV-DNN, RV-CNN, and RV-MWINet models are real-valued, the MWINet model is restructured with complex-valued layers (CV-MWINet), resulting in a total of four models. For the RV-DNN model, the training and test errors in terms of mean squared error (MSE) are found to be 103.400 and 96.395, respectively, whereas for the RV-CNN model, the training and test errors are obtained to be 45.283 and 153.818. Due to the fact that the RV-MWINet model is a combined U-Net model, the accuracy metric is analyzed. The proposed RV-MWINet model has training and testing accuracy of 0.9135 and 0.8635, whereas the CV-MWINet model has training and testing accuracy of 0.991 and 1.000, respectively. The peak signal-to-noise ratio (PSNR), universal quality index (UQI), and structural similarity index (SSIM) metrics were also evaluated for the images generated by the proposed neurocomputational models. The generated images demonstrate that the proposed neurocomputational models can be successfully utilized for radar-based microwave imaging, especially for breast imaging. [ABSTRACT FROM AUTHOR]

Deep Neural Networks (DNNs) have developed rapidly in data privacy protection applications such as medical treatment and finance. However, DNNs require high-speed and high-memory computers in terms of computation, otherwise training can be very lengthy. Furthermore, DNNs are often not available in resource-constrained mobile devices. Therefore, training and executing DNNs are increasingly using cloud computing. In the paper, the Power Normalized Cepstrum-based Robust Feature Detector (PNC-RFD), with deep learning in the cloud computing, is proposed for data privacy protection. The proposed PNC-RFD extracts a specified number of signal segments of high robustness used to embed and extract various data. For the sake of embedding and extracting the data, a method of information hiding employing Dual-Tree Complex Wavelet Packet Transform (DT CWPT) is therefore presented. The presented scheme simultaneously embeds multiple data into coefficients of the DT CWPT of signal segments. By embedding the data into the orthogonal spaces, the proposed method ensures the independent extraction of the multiple data. In line with the performance analysis, the superiority of the presented scheme is elaborated through making the comparison with the current state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Cochlear implant technology successfully restores hearing function to patients with sensory impairment. Although cochlear implant users generally hear well in quiet, they still find noisy conditions very challenging, hence the need to employ noise reduction algorithms in these systems to enhance the user experience. This paper reviews noise reduction algorithms in cochlear implants. Traditionally, such algorithms have been classified as either single- or multiple-channel, depending on the number of microphones they use. This review retains this general classification in looking at recent papers and extends it to reflect recent interest in machine learning techniques. The review concludes with consideration of promising future areas of research. [ABSTRACT FROM AUTHOR]

A sketch is a black-and-white, 2-D graphical representation of an object and contains fewer visual details as compared to a colored image. Despite fewer details, humans can recognize a sketch and its context very efficiently and consistently across languages, cultures, and age groups, but it is a difficult task for computers to recognize such low-detail sketches and get context out of them. With the tremendous increase in popularity of IoT devices such as smartphones and smart cameras, etc., it has become more critical to recognize free hand-drawn sketches in computer vision and human-computer interaction in order to build a successful artificial intelligence of things (AIoT) system that can first recognize the sketches and then understand the context of multiple drawings. Earlier models which addressed this problem are scale-invariant feature transform (SIFT) and bag-of-words (BoW). Both SIFT and BoW used hand-crafted features and scale-invariant algorithms to address this issue. But these models are complex and time-consuming due to the manual process of features setup. The deep neural networks (DNNs) performed well with object recognition on many large-scale datasets such as ImageNet and CIFAR-10. However, the DDN approach cannot be carried out for hand-drawn sketches problems. The reason is that the data source is images, and all sketches in the images are, for example, ‘birds’ instead of their specific category (e.g., ‘sparrow’). Some deep learning approaches for sketch recognition problems exist in the literature, but the results are not promising because there is still room for improvement. This article proposed a convolutional neural network (CNN) architecture called Sketch-DeepNet for the sketch recognition task. The proposed Sketch-DeepNet architecture used the TU-Berlin dataset for classification. The experimental results show that the proposed method beats the performance of the state-of-the-art sketch classification methods. The proposed model achieved 95.05% accuracy as compared to existing models DeformNet (62.6%), Sketch-DNN (72.2%), Sketch-a-Net (77.95%), SketchNet (80.42%), Thinning-DNN (74.3%), CNN-PCA-SVM (72.5%), Hybrid-CNN (84.42%), and human recognition accuracy of 73% on the TU-Berlin dataset.

Power line inspection is one important task performed by electricity distribution network operators worldwide. It is part of the equipment maintenance for such companies and forms a crucial procedure since it can provide diagnostics and prognostics about the condition of the power line network. Furthermore, it helps with effective decision making in the case of fault detection. Nowadays, the inspection of power lines is performed either using human operators that scan the network on foot and search for obvious faults, or using unmanned aerial vehicles (UAVs) and/or helicopters equipped with camera sensors capable of recording videos of the power line network equipment, which are then inspected by human operators offline. In this study, we propose an autonomous, intelligent inspection system for power lines, which is equipped with camera sensors operating in the visual (Red–Green–Blue (RGB) imaging) and infrared (thermal imaging) spectrums, capable of providing real-time alerts about the condition of power lines. The very first step in power line monitoring is identifying and segmenting them from the background, which constitutes the principal goal of the presented study. The identification of power lines is accomplished through an innovative hybrid approach that combines RGB and thermal data-processing methods under a custom-made drone platform, providing an automated tool for in situ analyses not only in offline mode. In this direction, the human operator role is limited to the flight-planning and control operations of the UAV. The benefits of using such an intelligent UAV system are many, mostly related to the timely and accurate detection of possible faults, along with the side benefits of personnel safety and reduced operational costs.

Rheumatoid Arthritis (RA) falls under the group of chronic autoimmune diseases, which affects the joints and muscles, and can lead to considerable damage to the joint structure and their functionality. RA diagnosis much in early stages is quite critical in stopping the progression of the disease. In this technical work, Dynamic Neutrosophic Cognitive Map with Improved Cuckoo Search Algorithm (DNCM-ICSA) with ensemble classifier is introduced for obtaining the gene expression profile, which distinguishes between the persons affected with RA and probable subjects of control. There are four important steps involved in this work, which include data preprocessing, feature selection, prediction and classification. The initial phase of the work comprises of data preprocessing, and second phase of the work comprises of gene selection process with T-test, chi-squared test, relief-F and Minimum Redundancy Maximum Relevance (mRMR). Next, the disease prediction is carried out using the ensemble mechanism, which increases the prediction accuracy. The ensemble mechanism integrates the process of Adaptive Neuro Fuzzy Inference System (ANFIS) and Deep Neural Networks (DNNs). The ensemble mechanism of classifiers is a group of classifiers whose decisions individually are integrated generally by weighted means for the classification of new RA examples. The disease of the patients may be avoided from getting to the severe stages. At last, DNCM-ICSA algorithm is utilized for gene classification. The results of the newly introduced classifier are analysed in terms of the metrics including precision, recall, F-measure and accuracy.

The size of deep neural network (DNN) models has been exploding rapidly, demanding a colossal amount of memory capacity. For example, Google has recently scaled its Switch Transformer to have a parameter size of up to 6.4 TB. However, today's HBM DRAM-based memory system for GPUs and DNN accelerators is suboptimal for these extreme-scale DNNs as it fails to provide enough capacity while its massive bandwidth is poorly utilized. Thus, we propose Leviathan, a DNN inference accelerator, which integrates a cost-effective flash-based storage system, instead. We carefully architect the storage system to provide enough memory bandwidth while preventing performance drop caused by read disturbance errors. Our evaluation of Leviathan demonstrates an 8.3× throughput gain compared to the iso-FLOPS DNN accelerator with conventional SSDs and up to 19.5× higher memory cost-efficiency than the HBM-based DNN accelerator. [ABSTRACT FROM AUTHOR]

Processing-in-memory (PIM) architecture has been proposed to accelerate state-of-the-art neuro-inspired algorithms, such as deep neural networks. In this article, we present PIMulator-NN, an event-driven, cross-level simulation framework for PIM-based neural network accelerators. By employing an event-driven simulation mechanism, PIMulator-NN is able to model architecture details and capture design details of the architecture. Moreover, we integrate the main-stream circuit-level simulation framework with PIMulator-NN to accurately simulate the area, latency, and energy consumption of analog computation units. To demonstrate the usage of PIMulator-NN, we implement several PIM designs with PIMulator-NN and perform detailed simulation. The simulation results show that memory access and interconnects make considerable impacts on system-level performance and energy. Note that such results are hard to be captured by conventional performance model-based estimations. We found some anti common-sense results while modeling the architecture details with PIMulator-NN. With several architecture templates, PIMulator-NN provides the users with a platform to build up their PIM architecture quickly. PIMulator-NN is able to capture the impacts of different design choices (e.g., dataflow, interconnect, data parallelism, etc.), and this could enable users to explore their design space efficiently. [ABSTRACT FROM AUTHOR]

Although deep neural networks (DNNs) have been increasingly applied in industrial cyber physical systems (ICPSs), they are vulnerable to security attacks due to the tight interaction between cyber elements and physical elements. In this article, we aim to protect the core IP of DNNs, i.e., the model weights, against security attacks. Different from conventional approaches, a layerwise protection framework is proposed to ensure the confidentiality of DNN model weights during the inference procedure such that the security quality is maximized, while satisfying the latency constraint of the DNN task. Based on the layerwise execution characteristics of DNN tasks, the encrypted layer-related weights are decrypted and fed to the next layer of DNN in plaintext. CPU-field programmable gate array (FPGA) coscheduling is considered to accelerate the execution of confidentiality protection, where CPU is utilized to conduct the decryption of weights and FPGA is used to perform the layer execution of DNN. Considering to provide optimal confidential protection for each layer, the problem is transformed into a quality of security maximization problem subject to layerwise execution constraint and deadline constraint of the DNN application. Due to the problem being NP-hard, a fast approximation algorithm is proposed to obtain the near-optimal solution under given real-time and security constraints. Extensive experiments and a real-life ICPS application evaluate the efficiency of the proposed techniques. [ABSTRACT FROM AUTHOR]