Deep Neural Network Classifier Research Articles

The impact of gut microbiome enterotypes on ulcerative colitis: identifying key bacterial species and revealing species co-occurrence networks using machine learning

ABSTRACT Ulcerative colitis (UC) is a chronic inflammatory intestinal disease affecting the colon and rectum, with its pathogenesis attributed to genetic background, environmental factors, and gut microbes. This study aimed to investigate the role of enterotypes in UC by conducting a hierarchical analysis, determining differential bacteria using machine learning, and performing Species Co-occurrence Network (SCN) analysis. Fecal bacterial data were collected from UC patients, and a 16S rRNA metagenomic analysis was performed using the QIIME2 bioinformatics pipeline. Enterotype clustering was conducted at the family level, and deep neural network (DNN) classification models were trained for UC and healthy controls (HC) in each enterotype. Results from eleven 16S rRNA gut microbiome datasets revealed three enterotypes: Bacteroidaceae (ET-B), Lachnospiraceae (ET-L), and Clostridiaceae (ET-C). Ruminococcus (R. gnavus) abundance was significantly higher in UC subjects with ET-B and ET-C than in those with ET-L. R. gnavus also showed a positive correlation with Clostridia in UC SCN for ET-B and ET-C subjects, with a higher correlation in ET-C subjects. Conversely, Odoribacter (O.) splanchnicus and Bacteroides (B.) uniformis exhibited a positive correlation with tryptophan metabolism and AMP-activated protein kinase (AMPK) signaling pathways, while R. gnavus showed a negative correlation. In vitro co-culture experiments with Clostridium (C.) difficile demonstrated that fecal microbiota from ET-B subjects had a higher abundance of C. difficile than ET-L subjects. In conclusion, the ET-B enterotype predisposes individuals to UC, with R. gnavus as a potential risk factor and O. splanchnicus and B. uniformis as protective bacteria, and those with UC may have ultimately become ET-C.

Description of tea quality using deep learning and multi-sensor feature fusion

To address the problem of the lack of rapid discrimination technology for comprehensive evaluation of product quality and grade in the field of tea quality control, this study uses near-infrared spectroscopy, electronic eye, electronic tongue, and electronic nose technology to sense the external and internal quality sensing characteristics of tea leaves, and realizes comprehensive evaluation of four quality factors such as "color, aroma, taste and shape" of tea leaves through cross-sensor multimodal quality characteristics data fusion. For the first time, a comprehensive method for assessing the grade quality of tea based on deep learning algorithms combining cross-sensor multimodal fusion features such as color, texture, shape, optical, electronic tongue, and electronic nose is established. The results showed that the constructed deep convolutional neural network (CNN) classification model had a misclassification rate of only 0.86% in the prediction set, which was better than the modeling results based on traditional classification methods. This study reveals that the CNN algorithm can effectively evaluate the quality grade of tea samples and realize the digital evaluation of multiple evaluation factors of tea sensory quality.

A Cryptographic based I2ADO-DNN Security Framework for Intrusion Detection in Cloud Systems

Cloud computing's popularity and success are directly related to improvements in the use of Information and Communication Technologies (ICT). The adoption of cloud implementation and services has become crucial due to security and privacy concerns raised by outsourcing data and business applications to the cloud or a third party. To protect the confidentiality and security of cloud networks, a variety of Intrusion Detection System (IDS) frameworks have been developed in the conventional works. However, the main issues with the current works are their lengthy nature, difficulty in intrusion detection, over-fitting, high error rate, and false alarm rates. As a result, the proposed study attempts to create a compact IDS architecture based on cryptography for cloud security. Here, the balanced and normalized dataset is produced using the z-score preprocessing procedure. The best attributes for enhancing intrusion detection accuracy are then selected using an Intelligent Adorn Dragonfly Optimization (IADO). In addition, the trained features are used to classify the normal and attacking data using an Intermittent Deep Neural Network (IDNN) classification model. Finally, the Searchable Encryption (SE) mechanism is applied to ensure the security of cloud data against intruders. In this study, a thorough analysis has been conducted utilizing various parameters to validate the intrusion detection performance of the proposed I2ADO-DNN model.

Differentiable matrix product states for simulating variational quantum computational chemistry

Quantum Computing is believed to be the ultimate solution for quantum chemistry problems. Before the advent of large-scale, fully fault-tolerant quantum computers, the variational quantum eigensolver (VQE) is a promising heuristic quantum algorithm to solve real world quantum chemistry problems on near-term noisy quantum computers. Here we propose a highly parallelizable classical simulator for VQE based on the matrix product state representation of quantum state, which significantly extend the simulation range of the existing simulators. Our simulator seamlessly integrates the quantum circuit evolution into the classical auto-differentiation framework, thus the gradients could be computed efficiently similar to the classical deep neural network, with a scaling that is independent of the number of variational parameters. As applications, we use our simulator to study commonly used small molecules such as HF, HCl, LiH and H2O, as well as larger molecules CO2, BeH2 and H4 with up to 40 qubits. The favorable scaling of our simulator against the number of qubits and the number of parameters could make it an ideal testing ground for near-term quantum algorithms and a perfect benchmarking baseline for oncoming large scale VQE experiments on noisy quantum computers.

Differentiation of frontotemporal dementia subtypes using neuroimaging‐based multi‐type parallel feature embedding

AbstractBackgroundFrontotemporal dementia (FTD) represents a collection of neurocognitive syndromes with frontotemporal lobar degeneration (FTLD) neuropathology, and is associated with significant clinical, pathological, and genetic heterogeneity. We trained a deep neural network (DNN) classifier to differentiate behavioral‐variant FTD (bvFTD), semantic variant primary progressive aphasia (svPPA), and non‐fluent variant (nfvPPA) patients using MRI scans drawn from two multi‐site neuroimaging consortiums.MethodBvFTD (N = 173), nfvPPA (N = 63), and svPPA (N = 41) patients with T1‐MRI were extracted from the FTLD Neuroimaging Initiative (FTLDNI) and the ARTFL‐LEFFTDS Longitudinal Frontotemporal Lobar Degeneration (ALLFTD) databases. MRI data were preprocessed with FreeSurfer, with cortical thickness, cortical volume, and subcortical volumes extracted. Cortical measures were parcellated into 360 patches (i.e., ROIs) according to the HCP‐MMP1 atlas. Both the patch‐based cortical thickness and volume features were harmonized to control confounding effects of sex, age, total intracranial volume (TIV), cohort, and scanner. Multi‐type features were parallelly fed into a multi‐layer‐perceptron (MLP)‐based classifier. Weighted cross‐entropy loss function was used to account for unbalanced sample sizes across FTD subtypes. 10‐fold nested cross‐validation was used to evaluate the robustness of the classification model, with data split into 80/10/10 of training/validation/test data sets.ResultVisual evaluation z‐scores of cortical features among FTLDNI and ALLFTD groups revealed site‐specific differences significantly reduced by feature harmonization, especially for volume (Figure 1). The balanced accuracy of the ensembled DNN classifier of each FTD subtype on test sets across all ten folds reached 0.76 ± 0.09 for bvFTD, 0.79 ± 0.07 for nfvPPA, and 0.88 ± 0.07 for svPPA (Figure 2). Figure 3 shows confusion matrices for classification performance in the test set across ten cross‐validation folds.ConclusionWe developed a deep‐learning‐based framework to classify three FTD subtypes: bvFTD, nfvPPA, and svPPA. The combination of feature harmonization and parallel multi‐type feature embedding framework showed promising differentiation power. This work can be used to recognize at‐risk populations for early and precise diagnosis, to aid intervention planning. Future studies will compare classification based on different input features and use visualization methods to identify the most discriminative regions and explore their clinical relevance.

A Deep Stacked Ensemble Model for Microarray Data Classification with Boosted Meta Classifier

Classification of microarray data is one of the major research interests in the biomedical field. It allows physicians for early detection of cancer through analysis of the Deoxyribonucleic acid (DNA). Classification of these sensitive data is still challenging due to the small sample and more feature size. In this paper, the authors have used an ensemble model for classifying two types of leukemia as Acute lymphocytic leukemia (ALL) and Acute myelocytic leukemia (AML). The work is carried out with other types of genetic data such as Leukemia, lung tumor, liver cancer, and liver Cirrhosis. The biomedical data are imbalanced. The ensemble classifier is based on a stacked approach where deep neural network (DNN) classifiers are used as the base classifier. The structure of each DNN is chosen as the homogenous type for the same training process for all classifiers. Because of the adaptive nature and random weight initialization, it provides different results for each classifier. The outputs of the base classifiers are again fed to a gradient boosting ensemble model termed a meta classifier. The meta classifier provides the final classification output. For comparison purposes, two types of meta-classifiers such as support vector machine (SVM) and ensemble gradient boosting are used in the proposed work. The performance of the model is verified well and the results are provided in the result section. From the experimental result, it is observed that the classification accuracy is 96% with SVM and 98% with boosted meta classifier for leukemia data, whereas 99.04% for lung tumor and 99.03% for liver Cirrhosis.

Hybrid Sneaky algorithm-based deep neural networks for Heart sound classification using phonocardiogram

ABSTRACT In the diagnosis of cardiac disorders Heart sound has a major role, and early detection is crucial to safeguard the patients. Computerized strategies of heart sound classification advocate intensive and more exact results in a quick and better manner. Using a hybrid optimization-controlled deep learning strategy this paper proposed an automatic heart sound classification module. The parameter tuning of the Deep Neural Network (DNN) classifier in a satisfactory manner is the importance of this research which depends on the Hybrid Sneaky optimization algorithm. The developed sneaky optimization algorithm inherits the traits of questing and societal search agents. Moreover, input data from the Phonocardiogram (PCG) database undergoes the process of feature extraction which extract the important features, like statistical, Heart Rate Variability (HRV), and to enhance the performance of this model, the features of Mel frequency Cepstral coefficients (MFCC) are assisted. The developed Sneaky optimization-based DNN classifier’s performance is determined in respect of the metrics, namely precision, accuracy, specificity, and sensitivity, which are around 97%, 96.98%, 97%, and 96.9%, respectively.

300-Gbit/s/λ 8-Level Pulse-Amplitude-Modulation (PAM8) with a silicon microring modulator utilizing long short term memory regression and deep neural network classification

We present the first experimental demonstration up the authors' knowledge a high-capacity short-reach optical communication link which employs a long-short-term-memory (LSTM) and deep-neural-network (DNN) for enabling 300-Gbit/s (100 Gbaud) 8-level pulse amplitude modulation (PAM8) generated by a single 55-GHz bandwidth silicon microring modulator (SiMRM) with a driving voltage of 1.8-Vpp. As PAM8 modulation has a lower average level-to-level transition energy when compared with that in PAM4, it could be possible to achieve higher energy efficiency when utilizing PAM8 at the same driving voltage. To recover the PAM8 modulation, LSTM and DNN are utilized for regression and classification respectively. The LSTM contains memory cells for dealing with signal time domain dependencies, and has the ability to store, read, and reject data passing through the neural network. The subsequent DNN performs the classification of the 8 levels. Experimental results show that 300 Gbit/s PAM8 modulation is achieved at back-to-back (B2B) and 270 Gbit/s PAM8 is achieved after 1 km standard single-mode-fiber (SSMF) transmission satisfying the soft-decision forward error correction (SD-FEC) bit-error-rate requirement (BER) of 2.4 × 10-2.

Integrated bulk and single-cell transcriptomes reveal pyroptotic signature in prognosis and therapeutic options of hepatocellular carcinoma by combining deep learning.

Although some pyroptosis-related (PR) prognostic models for cancers have been reported, pyroptosis-based features have not been fully discovered at the single-cell level in hepatocellular carcinoma (HCC). In this study, by deeply integrating single-cell and bulk transcriptome data, we systematically investigated significance of the shared pyroptotic signature at both single-cell and bulk levels in HCC prognosis. Based on the pyroptotic signature, a robust PR risk system was constructed to quantify the prognostic risk of individual patient. To further verify capacity of the pyroptotic signature on predicting patients' prognosis, an attention mechanism-based deep neural network classification model was constructed. The mechanisms of prognostic difference in the patients with distinct PR risk were dissected on tumor stemness, cancer pathways, transcriptional regulation, immune infiltration and cell communications. A nomogram model combining PR risk with clinicopathologic data was constructed to evaluate the prognosis of individual patients in clinic. The PR risk could also evaluate therapeutic response to neoadjuvant therapies in HCC patients. In conclusion, the constructed PR risk system enables a comprehensive assessment of tumor microenvironment characteristics, accurate prognosis prediction and rational therapeutic options in HCC.

A Hybrid Dimensionality Reduction for Network Intrusion Detection

Due to the wide variety of network services, many different types of protocols exist, producing various packet features. Some features contain irrelevant and redundant information. The presence of such features increases computational complexity and decreases accuracy. Therefore, this research is designed to reduce the data dimensionality and improve the classification accuracy in the UNSW-NB15 dataset. It proposes a hybrid dimensionality reduction system that does feature selection (FS) and feature extraction (FE). FS was performed using the Recursive Feature Elimination (RFE) technique, while FE was accomplished by transforming the features into principal components. This combined scheme reduced a total of 41 input features into 15 components. The proposed systems’ classification performance was determined using an ensemble of Support Vector Classifier (SVC), K-nearest Neighbor classifier (KNC), and Deep Neural Network classifier (DNN). The system was evaluated using accuracy, detection rate, false positive rate, f1-score, and area under the curve metrics. Comparing the voting ensemble results of the full feature set against the 15 principal components confirms that reduced and transformed features did not significantly decrease the classifier’s performance. We achieved 94.34% accuracy, a 93.92% detection rate, a 5.23% false positive rate, a 94.32% f1-score, and a 94.34% area under the curve when 15 components were input to the voting ensemble classifier.

Conditional generative adversarial network based data augmentation for fault diagnosis of diesel engines applied with infrared thermography and deep convolutional neural network

This paper tries to introduce a new intelligent method for the early fault diagnosis of diesel engines. Firstly, infrared thermography (IRT) is introduced into diesel engine condition monitoring, then infrared images of diesel engines in four health states, such as normal condition, single-cylinder misfire, multi-cylinder misfire and air filter blockage, are collected and the region of interest (ROI) of infrared images are extracted. Next, conditional generative adversarial network (CGAN) is deployed to perform data augmentation on infrared image datasets. Then, deep convolutional neural network (DCNN) and Softmax regression (SR) classifier are used for automatically extracting infrared image fault features and pattern recognition, respectively. Finally, a comparison with three deep learning (DL) models is performed. The validation results show that the data augmentation method proposed in the paper can significantly improve the early fault diagnosis accuracy, and DCNN has the best fault diagnosis effect and resistance to temperature fluctuation interference among the four DL models.

Improvement of emotion classification performance using multi-resolution variational mode decomposition method

Automated speech emotion recognition (SER) has been gaining popularity among researchers for three decades because of its vast number of applications in the real world. It is helpful to improve the relationship between humans and machines, online marketing and education, customer relations, medical treatment, safe driving, online search, etc. Researchers adopt various methods to improve emotion recognition performance from speech signals, like using different combinations of features (acoustic, non-acoustic, or both), classifiers (machine learning, deep learning, or both). Our study tried to improve emotion classification performance using features based on the multi-resolution variational mode decomposition (MRVMD) method. We first decompose each signal frame into several sub-signals known as modes or intrinsic mode functions (IMFs) using the MRVMD method. Then the proposed features, multi-resolution variational mode mel-frequency cepstral coefficient (MRVMMFCC), multi-resolution variational mode approximate entropy (MRVMAE), and multi-resolution variational mode permutation entropy (MRVMPE), were extracted using the MRVMD-decomposed IMF signals. Finally, different combinations of the proposed features are used to classify the emotion using a deep neural network (DNN) classifier. From the experimental results, we found that combination of the proposed feature (MRVMMFCC + MRVMAE + MRVMPE) performs better than the other combination in recognizing emotion using speech signals. The proposed feature combination with a DNN classifier achieved an emotion classification accuracy of 83.4%, 85.01%, and 90.51% for the SAVEE, EMOVO, and EMO-DB datasets, respectively. We found that the proposed MRVMD method performed better than the state-of-the-art methods.

Lungs X-Ray Image Segmentation and Classification of Lung Disease using Convolutional Neural Network Architectures

Lung disease is one of the biggest causes of death in the world. The SARS-CoV-2 virus causes diseases like COVID-19, and the bacteria Streptococcus sp., which causes pneumonia, are two sample causes of lung disease. X-ray images are used to detect the lung disease. This study aimed to combine the stages of segmentation and classification of lung disease. This study in segmentation aims to separate the features contained in the lung images. The classification aimed to provide holistic information on lung disease. This research method used the Deep Residual U-Net (DrU-Net) segmentation architecture and the Deep Residual Neural Network (DResNet) classification architecture. DrU-Net is a modified U-Net architecture with dropout added in its convolutional layers. DResNet is a modified Residual Network (ResNet) architecture with dropout added in its convolutional block layers. The result of this study was segmentation using the DrU-Net architecture obtained 99% for accuracy, 98% for precision, 98% for recalls, 98% for F1-Score, and 96.1% for IoU. The classification results of the segmented images using the DResNet architecture obtained 91% for accuracy, 86% for precision, 85% for recalls, and 84% for F1-Score. The performance results of DrU-Net architecture were excellent and robust in image segmentation. Unfortunately, the average performance of DResNet in classification was still below 90%. These results indicate that Dres-Net performs well in classifying lung disorders in 3 labels, namely Covid, Normal, and Pneumonia, but still needs improvement.

High-Content Image-Based Screening and Deep Learning for the Detection of Anti-Inflammatory Drug Leads.

We developed a high-content image-based screen that utilizes the pro-inflammatory stimulus lipopolysaccharide (LPS) and murine macrophages (RAW264.7) with the goal of enabling the identification of novel anti-inflammatory lead compounds. We screened 2,259 bioactive compounds with annotated mechanisms of action (MOA) to identify compounds that block the LPS-induced phenotype in macrophages. We utilized a set of seven fluorescence microscopy probes to generate images that were used to train and optimize a deep neural network classifier to distinguish between unstimulated and LPS-stimulated macrophages. The top hits from the deep learning classifier were validated using a linear classifier trained on individual cells and subsequently investigated in a multiplexed cytokine secretion assay. All 12 hits significantly modulated the expression of at least one cytokine upon LPS stimulation. Seven of these were allosteric inhibitors of the mitogen-activated protein kinase kinase (MEK1/2) and showed similar effects on cytokine expression. This deep learning morphological assay identified compounds that modulate the innate immune response to LPS and may aid in identifying new anti-inflammatory drug leads.

Automatic detection of artifacts and improved classification models for emotional activity detection from multimodal physiological data

This manuscript proposes an automated artifacts detection and multimodal classification system for human emotion analysis from human physiological signals. First, multimodal physiological data, including the Electrodermal Activity (EDA), electrocardiogram (ECG), Blood Volume Pulse (BVP) and respiration rate signals are collected. Second, a Modified Compressed Sensing-based Decomposition (MCSD) is used to extract the informative Skin Conductance Response (SCR) events of the EDA signal. Third, raw features (edge and sharp variations), statistical and wavelet coefficient features of EDA, ECG, BVP, respiration and SCR signals are obtained. Fourth, the extracted raw features, statistical and wavelet coefficient features from all physiological signals are fed into the parallel Deep Convolutional Neural Network (DCNN) to reduce the dimensionality of feature space by removing artifacts. Fifth, the fused artifact-free feature vector is obtained for neutral, stress and pleasure emotion classes. Sixth, an artifact-free feature vector is used to train the Random Forest Deep Neural Network (RFDNN) classifier. Then, a trained RFDNN classifier is applied to classify the test signals into different emotion classes. Thus, leveraging the strengths of both RF and DNN algorithms, more comprehensive feature learning using multimodal psychological data is achieved, resulting in robust and accurate classification of human emotional activities. Finally, an extensive experiment using the Wearable Stress and Affect Detection (WESAD) dataset shows that the proposed system outperforms other existing human emotion classification systems using physiological data.

Mapping integrated crop-livestock systems in Brazil with planetscope time series and deep learning

Accurate mapping of crops with high spatiotemporal resolution plays a critical role in achieving the Sustainable Development Goals (SDGs), especially in the context of integrated crop-livestock systems (ICLS). Stakeholders can make informed decisions and implement targeted strategies to achieve multiple SDGs related to agriculture, rural development, and sustainable livelihoods by understanding the spatial dynamics of these systems. Accurate information on the extent of ICLS derived from multitemporal remote sensing and emerging map techniques such as deep learning can help in the implementation of sustainable agricultural practices. However, far too little attention has been paid to ICLS map accuracy because it may not be at the forefront of research agendas compared to those of other agricultural practices. This paper aims to map ICLS using high spatiotemporal resolution imagery and deep learning neural network classifiers at two different sites located in Brazil. The pipeline involves four interpretation approaches based on the ICLS class: evaluating deep neural network classifiers with different image composition intervals, explaining commission and omission errors, evaluating the temporal transferability of the method, and evaluating the influence of variables. The study area consists of two locations in São Paulo (study site 1, SS1) and Mato Grosso state (study site 2, SS2), Brazil. We derived nine spectral variables from PlanetScope (PS) images and four metrics through object-based image analysis (OBIA) using two time intervals, 10 and 15 days, to generate the image compositions. These input variables were used in three deep neural network classifiers: convolutional neural network in one dimension (Conv1D), long short-term memory (LSTM), and LSTM with a fully convolutional network (LSTM-FCN). Our results showed that mapping dynamic land use such as ICLS is possible by using high-spatiotemporal-resolution imagery and deep neural network classifiers. The 15-day LSTM-FCN classifier returned the highest map accuracies for both sites, with the following class-level accuracies: producer accuracy (PA) = 97.0% and user accuracy (UA) = 97.0% for SS1 and PA = 82.0% and UA = 96.5% for SS2. Meanwhile, we found map uncertainties arising from the diverse crop calendars and spectro-temporal similarities between ICLS and other land use. The best approaches revealed that temporal generalization was suitable for mapping ICLS, but some classifiers could not generalize due to the inherent characteristics of the class. Most variables were considered efficient for predicting ICLS, although spectral indices revealed better functional relationships, while the PS bands had a lower influence on the predictions. The accuracies achieved with the proposed method represent promising opportunities for the sufficiently accurate mapping of ICLS and other complex crop activities.

Automatic Detection of Covid-19 in CT Images Using an Optimized Deep Neural Network Classifier

An exact and quick finding of Covid-19 patients plays a significant part in the initial period of medicinal treatment and prevention. Automatic recognition of COVID-19 cases using CT images may help lessen the effect of this infection on human civilization. In recent years, computer vision is the major solution for diagnosing the covid-19 disease by using CT images. Besides, many researchers had presented efficient artificial intelligence techniques for identifying Covid-19 disease. Nevertheless, the accuracy and time consumption of the model further to be improved. Thus, in this work, we propose a Covid-19 disease classification using a hybrid deep neural network with handcrafted features. The proposed approach consists of three stages namely, pre-processing, handcrafted feature extraction, and classification. Initially, the images are given to the pre-processing stage to remove the noise present in the input images. Then, we extract the hand-crafted features (Gray-Level Co-Occurrence Matrix (GLCM)) from each image. After that, the pre-processed image is given to the input of the optimized deep neural network classifier to classify an image as normal or abnormal. The proposed optimized deep neural network is a combination of a convolution neural network (CNN) and adaptive artificial jelly optimization (A2JO) algorithm. To enhance the performance of classifier accuracy, the extracted handcrafted features are fused with the hybrid deep neural network. The performance of the proposed approach was analyzed based on different metrics and performance compared with other techniques.

Wearable Sensors-Based Human Activity Recognition with Deep Convolutional Neural Network and Fuzzy Classification

Wearable Sensors-Based Human Activity Recognition with Deep Convolutional Neural Network and Fuzzy Classification

Detecting Adversarial Examples on Deep Neural Networks With Mutual Information Neural Estimation

Despite achieving exceptional performance, deep neural networks (DNNs) suffer from the harassment caused by adversarial examples, which are produced by corrupting clean examples with tiny perturbations. Many powerful defense methods have been presented such as training data augmentation and input reconstruction which, however, usually rely on the prior knowledge of the targeted models or attacks. A clean example and its adversarial version are very similar but have different high-level representations in a victim model. If we can obtain a space in which the representations of similar examples are also similar, then adversarial examples can be picked out by comparing the representations of input examples in this space and the high-level space of the victim model. Inspired by this, we propose a novel approach for detecting adversarial images, which can protect any pre-trained DNN classifiers and resist an endless stream of new attacks. Specifically, we first adopt a dual autoencoder to project images to a latent space. The dual autoencoder uses the self-supervised learning to ensure that small modifications to samples do not significantly alter their latent representations. Next, the mutual information neural estimation is utilized to enhance the discrimination of the latent representations. We then leverage the prior distribution matching to regularize the latent representations. To easily compare the representations of examples in the two spaces, and not rely on the prior knowledge of the targeted model, a simple fully connected neural network is used to embed the learned representations into an eigenspace, which is consistent with the output eigenspace of the targeted model. Through the distribution similarity of an input example in the two eigenspaces, we can judge whether the input example is adversarial or not. Extensive experiments on MNIST, CIFAR-10, and ImageNet show that the proposed method has superior defense performance and transferability than state-of-the-arts.

Comprehensive analysis of the heterogeneous computing performance of DNNs under typical frameworks on cloud and edge computing platforms

In this paper, nearly 40 commonly used deep neural network(DNN) models are selected, and their cross-platform and cross-inference frameworks are deeply analysed. The main metrics of accuracy, the total number of model parameters, the computational complexity, the accuracy density, the inference time, the memory consumption and other related parameters are used to measure their performance. The heterogeneous computing experiment is implemented on both the Google Colab cloud computing platform and the Jetson Nano embedded edge computing platform. The obtained performance is compared with that of two previous computing platforms: a workstation equipped with an NVIDIA Titan X Pascal and an embedded system based on an NVIDIA Jetson TX1 board. In addition, on the Jetson Nano embedded edge computing platform, different inference frameworks are investigated to evaluate the inference efficiency of the DNN models. Regression models are established to characterize the variation in the computing performance of different DNN classification algorithms so that the inference results of unknown models can be estimated. ANOVA methods are proposed to quantify the differences between models. The experimental results have important guiding significance for the better selection, deployment and application of DNN models in practice. Codes are available at this https URL https://github.com/Foreverzfy/Model-Test.