Original Training Research Articles

Multi-span optical power spectrum prediction using cascaded learning with one-shot end-to-end measurement

Scalable methods for optical transmission performance prediction using machine learning (ML) are studied in metro reconfigurable optical add-drop multiplexer (ROADM) networks. A cascaded learning framework is introduced to encompass the use of cascaded component models for end-to-end (E2E) optical path prediction augmented with different combinations of E2E performance data and models. Additional E2E optical path data and models are used to reduce the prediction error accumulation in the cascade. Off-line training (pre-trained prior to deployment) and transfer learning are used for component-level erbium-doped fiber amplifier (EDFA) gain models to ensure scalability. Considering channel power prediction, we show that the data collection process of the pre-trained EDFA model can be reduced to only 5% of the original training set using transfer learning. We evaluate the proposed method under three different topologies with field deployed fibers and achieve a mean absolute error of 0.16 dB with a single (one-shot) E2E measurement on the deployed 6-span system with 12 EDFAs.

Research on automatic segmentation and recognition of single characters in original rubbings based on intelligent recognition of oracle bone inscriptions

This paper proposes an effective segmentation and recognition method using YOLOv8 to solve the problem of segmentation and recognition of Oracle rubbings images. Draw sub-pixel edge contours through spot detection, corner detection, and edge detection (such as Canny and Sobel), and combine grayscale, opening operations, binarization, and Gaussian filtering to remove noise and repair contours, extract image features, and establish preprocessing Model. Use this model to preprocess the original oracle rubbing image and train it based on YOLOv8. The results show that this method can effectively segment and identify oracle bones characters.

Classification and Comparison of Data Augmentation Techniques

Data augmentation is a technical method that generates additional data samples by applying different transformations and processes to the original training data. It aims to increase the size and diversity of the dataset, thus improving the models' generalization ability. This paper provides a detailed introduction to geometric transformations, data mixing, automated data augmentation methods, and generative adversarial methods. Geometric transformations manipulate the spatial configuration of data samples to create new variants that retain the original content. Data mixing techniques blend multiple data samples to produce novel training examples, promoting robustness and reducing overfitting. Automated data augmentation methods leverage search algorithms to discover optimal augmentation policies that improve model performance without manual intervention. Generative adversarial methods employ generative adversarial networks to synthesize realistic data samples that enrich the training dataset. These methods find broad application in domains including object detection, image classification, natural language processing, and speech recognition, significantly improving model performance. The paper explores the advantages and limitations of data augmentation techniques in practical applications and identifies future research directions and challenges. By systematically classifying and selecting appropriate data augmentation methods for specific application scenarios, this review provides a theoretical foundation and practical guidance for real-world applications.

From Labquakes to Megathrusts: Scaling Deep Learning Based Pickers Over 15 Orders of Magnitude

AbstractThe application of machine learning techniques in seismology has greatly advanced seismological analysis, especially for earthquake detection and seismic phase picking. However, machine learning approaches still face challenges in generalizing to data sets that differ from their original training setting. Previous studies focused on retraining or transfer‐learning models for these scenarios, but require high‐quality labeled data sets. This paper demonstrates a new approach for augmenting already trained models without the need for additional training data. We propose four strategies—rescaling, model aggregation, shifting, and filtering—to enhance the performance of pre‐trained models on out‐of‐distribution data sets. We further devise various methodologies to ensemble the individual predictions from these strategies to obtain a final unified prediction result featuring prediction robustness and detection sensitivity. We develop an open‐source Python module quakephase that implements these methods and can flexibly process input continuous seismic data of any sampling rate. With quakephase and pre‐trained ML models from SeisBench, we perform systematic benchmark tests on data recorded by different types of instruments, ranging from acoustic emission sensors to distributed acoustic sensing, and collected at different scales, spanning from laboratory acoustic emission events to major tectonic earthquakes. Our tests highlight that rescaling is essential for dealing with small‐magnitude seismic events recorded at high sampling rates as well as larger magnitude events having long coda and remote events with long wave trains. Our results demonstrate that the proposed methods are effective in augmenting pre‐trained models for out‐of‐distribution data sets, especially in scenarios with limited labeled data for transfer learning.

A Reliable Aggregation Method Based on Threshold Additive Secret Sharing in Federated Learning with Quality of Service (QoS) Support

Federated learning is a decentralized privacy-preserving mechanism that allows multiple clients to collaborate without exchanging their datasets. Instead, they jointly train a model by uploading their own gradients. However, recent research has shown that attackers can use clients’ gradients to reconstruct the original training data, compromising the security of federated learning. Thus, there has been an increasing number of studies aiming to protect gradients using different techniques. One common technique is secret sharing. However, it has been shown in previous research that when using secret sharing to protect gradient privacy, the original gradient cannot be reconstructed when one share is lost or a server is damaged, causing federated learning to be interrupted. In this paper, we propose an approach that involves using additive secret sharing for federated learning gradient aggregation, making it difficult for attackers to easily access clients’ original gradients. Additionally, our proposed method ensures that any server damage or loss of gradient shares are unlikely to impact the federated learning operation, within a certain probability. We also added a membership level system, allowing members of varying levels to ultimately obtain models with different accuracy levels.

Noise-Aware Active Learning to Develop High-Temperature Shape Memory Alloys with Large Latent Heat.

Shape memory alloys (SMAs) with large latent heat absorbed/released during phase transformation at elevated temperatures benefit their potential application on thermal energy storage (TES) in high temperature environment like power plants, etc. The desired alloys can be designed quickly by searching the vast component space of doped NiTi-based SMAs via data-driven method, while be challenging with the noisy experimental data. A noise-aware active learning strategy is proposed to accelerate the design of SMAs with large latent heat at elevated phase transformation temperatures based on noisy data. The optimal noise level is estimated by minimizing the model error with incorporation of a range of noise levels as noise hyper-parameters into the noise-aware Kriging model. The employment of this strategy leads to the discovery of the alloy with latent heat of -36.08 J g-1, 9.2% larger than the best value (-33.04 J g-1) in the original training dataset within another four experiments. Additionally, the alloy represents high austenite finish temperature (481.71°C) and relatively small hysteresis. This promotes the latent heat TES application of SMAs in high temperature circumstance. It is expected that the noise-aware approach can be convenient for the accelerated materials design via the data-driven method with noisydata.

Validation of numerical method of unsteady Reynolds-averaged Navier–Stokes coupled with discrete phase model and optimization for the snow-resistance performance of bogies of a high-speed train

This paper presents an investigation aimed at enhancing the snow resistance of bogies in a three-car grouped alpine high-speed train (HST) that operates in a snowy region. The study employs a combination of the unsteady Reynolds-averaged Navier–Stokes model, coupled with the discrete phase model. This modeling approach has been validated through wind tunnel tests, including flow field tests, two-phase flow tests, and snow-ice tests. The study comprehensively compares the characteristics of the wind-snow flow, spatial snow concentration, and the distribution of snow on the train's underbody between two scenarios: the original train's underbody structure and the multifactor collaborative optimization scheme (MCOS). The results indicate that the MCOS case facilitates the escape of snow particles from the bogie cavities by weakening the upward and reverse flows. Furthermore, the MCOS case significantly reduces the concentration of snow particles around the heat-producing elements and decreases the accumulation of snow on both the lower and upper surfaces of the bogies. As a result, it reduces snow accumulation on the bogies by 50.6%, 56.4%, 60.8%, and 62.4% at train speeds of 200, 250, 300, and 350 km/h, respectively. In summary, this research provides valuable insights for improving the snow resistance of HSTs operating in snowy regions, with potential applications in enhancing railway transportation safety and efficiency.

Subgraph Structure Membership Inference Attacks against Graph Neural Networks

Graph Neural Networks (GNNs) have been widely applied to various applications across different domains. However, recent studies have shown that GNNs are susceptible to the membership inference attacks (MIAs) which aim to infer if some particular data samples were included in the model’s training data. While most previous MIAs have focused on inferring the membership of individual nodes and edges within the training graph, we introduce a novel form of membership inference attack called the Structure Membership Inference Attack (SMIA) which aims to determine whether a given set of nodes corresponds to a particular target structure, such as a clique or a multi-hop path, within the original training graph. To address this issue, we present novel black-box SMIA attacks that leverage the prediction outputs generated by the target GNN model for inference. Our approach involves training a three-label classifier, which, in combination with shadow training, aids in enabling the inference attack. Our extensive experimental evaluation of three representative GNN models and three real-world graph datasets demonstrates that our proposed attacks consistently outperform three baseline methods, including the one that employs the conventional link membership inference attacks to infer the subgraph structure. Additionally, we design a defense mechanism that introduces perturbations to the node embeddings thus influencing the corresponding prediction outputs by the target model. Our defense selectively perturbs dimensions within the node embeddings that have the least impact on the model's accuracy. Our empirical results demonstrate that the defense effectiveness of our approach is comparable with two established defense techniques that employ differential privacy. Moreover, our method achieves a better trade-off between defense strength and the accuracy of the target model compared to the two existing defense methods.

HIERARCHICAL NEURAL NETWORKS IN PREDICTING THE PROPERTIES OF OIL AND GAS RESERVOIRS BASED ON WELL AND SEISMIC DATA

This paper describes a technique for hierarchical neural networks based on the nearest neighbor method with preliminary clustering of the original training dataset and construction of a search cluster decision tree. This method is a promising alternative to neural network technologies with deep learning and has quite a few advantages: high learning rate, identification of objects with a low degree of similarity, and the ability to generalize and retrain. As shown by testing the hierarchical neural network method on real data from the West Siberian oil and gas province, predicting the oil saturation in the Vikulov suite interval is much faster and more efficient than inversion approaches to quantitative interpretation of seismic data while achieving fairly similar geological results. This characterizes the proposed method of hierarchical neural networks as an effective tool for the quantitative interpretation of seismic data to solve geological problems.

Improving Localization in Wireless Sensor Networks for the Internet of Things Using Data Replication-Based Deep Neural Networks.

Localization is one of the most challenging problems in wireless sensor networks (WSNs), primarily driven by the need to develop an accurate and cost-effective localization system for Internet of Things (IoT) applications. While machine learning (ML) algorithms have been widely applied in various WSN-based tasks, their effectiveness is often compromised by limited training data, leading to issues such as overfitting and reduced accuracy, especially when the number of sensor nodes is low. A key strategy to mitigate overfitting involves increasing both the quantity and diversity of the training data. To address the limitations posed by small datasets, this paper proposes an intelligent data augmentation strategy (DAS)-based deep neural network (DNN) that enhances the localization accuracy of WSNs. The proposed DAS replicates the estimated positions of unknown nodes generated by the Dv-hop algorithm and introduces Gaussian noise to these replicated positions, creating multiple modified datasets. By combining the modified datasets with the original training data, we significantly increase the dataset size, which leads to a substantial reduction in normalized root mean square error (NRMSE). The experimental results demonstrate that this data augmentation technique significantly improves the performance of DNNs compared to the traditional Dv-hop algorithm at a low number of nodes while maintaining an efficient computational cost for data augmentation. Therefore, the proposed method provides a scalable and effective solution for enhancing the localization accuracy of WSNs.

Geometry of Textual Data Augmentation: Insights from Large Language Models

Data augmentation is crucial for enhancing the performance of text classification models when labelled training data are scarce. For natural language processing (NLP) tasks, large language models (LLMs) are able to generate high-quality augmented data. But a fundamental understanding of the reasons for their effectiveness remains limited. This paper presents a geometric and topological perspective on textual data augmentation using LLMs. We compare the augmentation data generated by GPT-J with those generated through cosine similarity from Word2Vec and GloVe embeddings. Topological data analysis reveals that GPT-J generated data maintains label coherence. Convex hull analysis of such data represented by their two principal components shows that they lie within the spatial boundaries of the original training data. Delaunay triangulation reveals that increasing the number of augmented data points that are connected within these boundaries correlates with improved classification accuracy. These findings provide insights into the superior performance of LLMs in data augmentation. A framework for predicting the usefulness of augmentation data based on geometric properties could be formed based on these techniques.

An updated AL-Base reveals ranked enrichment of immunoglobulin light chain variable genes in AL amyloidosis.

Each monoclonal antibody light chain associated with AL amyloidosis has a unique sequence. Defining how these sequences lead to amyloid deposition could facilitate faster diagnosis and lead to new treatments. Light chain sequences are collected in the Boston University AL-Base repository. Monoclonal sequences from AL amyloidosis, multiple myeloma and the healthy polyclonal immune repertoire were compared to identify differences in precursor gene use, mutation frequency and physicochemical properties. AL-Base now contains 2,193 monoclonal light chain sequences from plasma cell dyscrasias. Sixteen germline precursor genes were enriched in AL amyloidosis, relative to multiple myeloma and the polyclonal repertoire. Two genes, IGKV1-16 and IGLV1-36, were infrequently observed but highly enriched in AL amyloidosis. The number of mutations varied widely between light chains. AL-associated κ light chains harbored significantly more mutations compared to multiple myeloma and polyclonal sequences, whereas AL-associated λ light chains had fewer mutations. Machine learning tools designed to predict amyloid propensity were less accurate for new sequences than their original training data. Rarely-observed light chain variable genes may carry a high risk of AL amyloidosis. New approaches are needed to define sequence-associated risk factors for AL amyloidosis. AL-Base is a foundational resource for such studies.

Graph network-based human movement prediction for socially-aware robot navigation in shared workspaces

AbstractMethods for socially-aware robot path planning are increasingly needed as robots and humans increasingly coexist in shared industrial spaces. The practice of clearly separated zones for humans and robots in shop floors is transitioning towards spaces where both humans and robot operate, often collaboratively. To allow for safer and more efficient manufacturing operations in shared workspaces, mobile robot fleet path planning needs to predict human movement. Accounting for the spatiotemporal nature of the problem, the present work introduces a spatiotemporal graph neural network approach that uses graph convolution and gated recurrent units, together with an attention mechanism to capture the spatial and temporal dependencies in the data and predict human occupancy based on past observations. The obtained results indicate that the graph network-based approach is suitable for short-term predictions but the rising uncertainty beyond short-term would limit its applicability. Furthermore, the addition of learnable edge weights, a feature exclusive to graph neural networks, enhances the predictive capabilities of the model. Adding workspace context-specific embeddings to graph nodes has additionally been explored, bringing modest performance improvements. Further research is needed to extend the predictive capabilities beyond the range of scenarios captured through the original training, and towards establishing standardised benchmarks for testing human motion prediction in industrial environments.

Feature-Based Dataset Fingerprinting for Clustered Federated Learning on Medical Image Data

ABSTRACT Federated Learning (FL) allows multiple clients to train a common model without sharing their private training data. In practice, federated optimization struggles with sub-optimal model utility because data is not independent and identically distributed (non-IID). Recent work has proposed to cluster clients according to dataset fingerprints to improve model utility in such situations. These fingerprints aim to capture the key characteristics of clients’ local data distributions. Recently, a mechanism was proposed to calculate dataset fingerprints from raw client data. We find that this fingerprinting mechanism comes with substantial time and memory consumption, limiting its practical use to small datasets. Additionally, shared raw data fingerprints can directly leak sensitive visual information, in certain cases even resembling the original client training data. To alleviate these problems, we propose a Feature-based dataset FingerPrinting mechanism (FFP). We use the MedMNIST database to develop a highly realistic case study for FL on medical image data. Compared to existing methods, our proposed FFP reduces the computational overhead of fingerprint calculation while achieving similar model utility. Furthermore, FFP mitigates the risk of raw data leakage from fingerprints by design.

Deep‐Learning‐Assisted Affinity Classification for Humoral Immunoprotein Complexes

Immunoglobulins are important building blocks in biology and biotechnology. With the emergence of comprehensive deep‐learning approaches, there are enormous opportunities for fast and accurate methods of classification of protein–protein interactions to arise. Herein, widely accessible image classification algorithms for species‐specific typification of a range of different immunoglobulin G (IgG) complexes are repurposed. Droplets of various immunoglobulins mixed with a B‐cell superantigen (SAg) (recombinant staphylococcal Protein A) are deposited onto hydrophobic polymer substrates and the resulting protein stains are imaged using polarized light microscopy. A comprehensive study based on 23 745 images finds that the pretrained convolutional neural network (CNN) InceptionV3 not only successfully categorizes IgGs from four different species but also predicts their binding affinity to Protein A: averaged over 36 binding pairs, the following are observed: 1) an overall accuracy of 81.4%, 2) the highest prediction accuracy for human IgG, the antibody with the highest binding affinity for Protein A, and 3) that the classification accuracy regarding the various IgG/Protein A ratios generally correlates with the binding strength of the protein–protein–complex as determined via circular dichroism spectroscopy. In addition, the CNN pretrained with IgG/Protein A stain images has been challenged with a new set of images using a different superantigen (SAg, Protein G). Despite the use of the unknown SAg, the CNN correctly classifies the images of human IgG and Protein G as indicated by a 94% accuracy over the various molar binding ratios. These findings are noteworthy because they demonstrate that appropriately pretrained CNNs can be used for the prediction of protein–protein interactions beyond the scope of the original training set. Aided by deep‐learning methods, simple stains of mixed protein solutions may serve as accurate predictors of the strength of protein–protein interactions with relevance to protein engineering, self‐aggregation, or protein stability in complex media.

Warning! This is not your typical mentor training: an invitation to explore a video-based mentor training to prepare mentor teachers for their roles in the classroom

PurposeThe purpose of this conceptual paper is to define and explore the roles of mentors, the responsibilities and even misconceptions of their position in partnership schools, the characteristics of effective mentors, the gaps that exist in current mentor training and the need for updated and forward-thinking flexible and accessible mentor training as it relates to improvements in University Teacher Preparation Programs and their partnership schools.Design/methodology/approachThe authors detail the process of interviewing a diverse population of mentors and teacher candidates (TCs) who have participated in our professional development (partnership) schools. See Appendices A and B for parallel questions posed during each interview process. Following the interviews, data were gathered in both text and video-based formats to create mentor training video modules to improve the current state of mentor training in our program. Themes were identified following an analysis of both interview intakes, and modules were created to align with these themes.FindingsFollowing the first rounds of implementation, the authors have reflected and noted that a need for an even more diverse population of both mentors and interns to be interviewed is necessary moving forward. The authors do note an appreciation by mentors and university partners in the quality, flexibility and accessibility that this new video module-based mentor training program provides.Practical implicationsThe video-based mentor training modules that the authors detail honor the needs of both new and returning mentors. Via the interviews with teacher candidates, mentors are able to empathize and understand how to be better mentors to their future TCs. Via the scenarios and questions that follow each thematically driven module, mentors are able to independently reflect on their current practices and ways to improve their roles. At any point in the internship year, mentors can return to the video training modules to review and thus improve their practice.Social implicationsMentors, university supervisors, teacher candidates and site coordinators (those individuals who identify mentors in their buildings) note improvement in their ability to communicate effectively as a result of being presented with video modules and reflections about the role of mentors from both mentors and TCs.Originality/valueThe authors were called on by leadership in the College of Education at the University to create this original mentor training video module. It is unique to the College of Education at the University. Great value exists in its accessibility, adaptability (we can upload new videos at any time) and representation of both the mentor and TC perceptions, suggestions and experiences in our program.

Deep learning with uncertainty estimation for automatic tumor segmentation in PET/CT of head and neck cancers: impact of model complexity, image processing and augmentation

Objective. Target volumes for radiotherapy are usually contoured manually, which can be time-consuming and prone to inter- and intra-observer variability. Automatic contouring by convolutional neural networks (CNN) can be fast and consistent but may produce unrealistic contours or miss relevant structures. We evaluate approaches for increasing the quality and assessing the uncertainty of CNN-generated contours of head and neck cancers with PET/CT as input. Approach. Two patient cohorts with head and neck squamous cell carcinoma and baseline 18F-fluorodeoxyglucose positron emission tomography and computed tomography images (FDG-PET/CT) were collected retrospectively from two centers. The union of manual contours of the gross primary tumor and involved nodes was used to train CNN models for generating automatic contours. The impact of image preprocessing, image augmentation, transfer learning and CNN complexity, architecture, and dimension (2D or 3D) on model performance and generalizability across centers was evaluated. A Monte Carlo dropout technique was used to quantify and visualize the uncertainty of the automatic contours. Main results. CNN models provided contours with good overlap with the manually contoured ground truth (median Dice Similarity Coefficient: 0.75–0.77), consistent with reported inter-observer variations and previous auto-contouring studies. Image augmentation and model dimension, rather than model complexity, architecture, or advanced image preprocessing, had the largest impact on model performance and cross-center generalizability. Transfer learning on a limited number of patients from a separate center increased model generalizability without decreasing model performance on the original training cohort. High model uncertainty was associated with false positive and false negative voxels as well as low Dice coefficients. Significance. High quality automatic contours can be obtained using deep learning architectures that are not overly complex. Uncertainty estimation of the predicted contours shows potential for highlighting regions of the contour requiring manual revision or flagging segmentations requiring manual inspection and intervention.

Improving 18F-FDG PET Quantification Through a Spatial Normalization Method.

Quantification of 18F-FDG PET images is useful for accurate diagnosis and evaluation of various brain diseases, including brain tumors, epilepsy, dementia, and Parkinson disease. However, accurate quantification of 18F-FDG PET images requires matched 3-dimensional T1 MRI scans of the same individuals to provide detailed information on brain anatomy. In this paper, we propose a transfer learning approach to adapt a pretrained deep neural network model from amyloid PET to spatially normalize 18F-FDG PET images without the need for 3-dimensional MRI. Methods: The proposed method is based on a deep learning model for automatic spatial normalization of 18F-FDG brain PET images, which was developed by fine-tuning a pretrained model for amyloid PET using only 103 18F-FDG PET and MR images. After training, the algorithm was tested on 65 internal and 78 external test sets. All T1 MR images with a 1-mm isotropic voxel size were processed with FreeSurfer software to provide cortical segmentation maps used to extract a ground-truth regional SUV ratio using cerebellar gray matter as a reference region. These values were compared with those from spatial normalization-based quantification methods using the proposed method and statistical parametric mapping software. Results: The proposed method showed superior spatial normalization compared with statistical parametric mapping, as evidenced by increased normalized mutual information and better size and shape matching in PET images. Quantitative evaluation revealed a consistently higher SUV ratio correlation and intraclass correlation coefficients for the proposed method across various brain regions in both internal and external datasets. The remarkably good correlation and intraclass correlation coefficient values of the proposed method for the external dataset are noteworthy, considering the dataset's different ethnic distribution and the use of different PET scanners and image reconstruction algorithms. Conclusion: This study successfully applied transfer learning to a deep neural network for 18F-FDG PET spatial normalization, demonstrating its resource efficiency and improved performance. This highlights the efficacy of transfer learning, which requires a smaller number of datasets than does the original network training, thus increasing the potential for broader use of deep learning-based brain PET spatial normalization techniques for various clinical and research radiotracers.

Translation, cross-cultural adaptation and validation of the universal Welch Emotional Connection Screen using primary and bilingual Spanish-speaking coders of videotaped mother-child interactions

IntroductionUsing clear explicit translatable language, we translated the Welch Emotional Connection Screen into a new universal language instrument, the English uWECS. In this study, we had two aims: Aim 1 was to establish concurrent validity of the uWECS by comparing scores coded by primary Spanish-speaking coders using the Spanish translation of the uWECS to scores coded by bilingual, secondary Spanish-speaking coders using the oWECS. Aim 2 was to establish the criterion-related validity in terms of oWECS and uWECS performance in tracking change in autonomic emotional connection (AEC) during the course of an intervention among preschool aged children.MethodsWe created a library of 52 five-minute Spanish-speaking mother-child videos that were collected during a randomized controlled trial of Mother-Child Emotional Preparation intervention (MCEP). The videos were collected at two time points, at enrollment and at a 6-month follow-up. The subsample of Primary Spanish-Speaking dyads from the MCEP study were coded by two independent teams of coders. We trained primary English-speaking (bilingual Spanish) coders on the oWECS, using the original training program. A different team of primary Spanish-speaking coders coded the same cases using the novel uWECS guide and trained briefly for reliability with the Spanish uWECS translation materials.ResultsWe found that the Spanish oWECS and Spanish uWECS ratings from the baseline and 6-month follow-up observations were robustly correlated, with intraclass correlations ranging from .81 to .84 and all p-values<.001, thus demonstrating sound concurrent validity for the uWECS. The oWECS and uWECS scores also achieved parallel results when evaluating the efficacy of the MCEP for primary Spanish-speaking dyads. Both the AEC scores of the oWECS [F(1, 27) = 4.31, p < .05] and the scores of the uWECS [F(1,27) = 4.06, p < .05] similarly demonstrated significant change post intervention, thus demonstrating sound criterion-related validity of the uWECS.DiscussionThese findings demonstrate that the uWECS can be used to measure parent/child AEC in linguistically diverse populations and cultures.

An intrusion detection method combining variational auto-encoder and generative adversarial networks

Deep learning is a crucial research area in network security, particularly when it comes to detecting network attacks. While some deep learning algorithms have shown promising results in distinguishing between normal and abnormal traffic, identifying different types of imbalanced anomalous traffic data is still a challenging task at present. To enhance the detection performance of unbalanced anomalous flows, we propose a new intrusion detection architecture based on a variational auto-encoder (VAE) and generative adversarial networks (GAN) in this research. Firstly, we present the VAE-WGAN model, which combines the advantages of VAE and GAN and enables us to generate data with predefined labels to balance the original training dataset. In the intrusion detection phase, we use a hybrid neural network model based on stacked Long Short-Term Memory (LSTM) and Multi-Scale Convolutional Neural Network (MSCNN). Stacked LSTM and MSCNN networks can extract network characteristics at different depths and scales, and subsequent feature fusion is used to increase network attack detection rates. Finally, the results from the NSL-KDD and AWID datasets indicate that the proposed network intrusion detection model improves the accuracy of network attack detection. The model outperforms other existing intrusion detection approaches in terms of accuracy, precision, recall, and f1-score, obtaining 83.45% accuracy and 83.69% f1-score on the NSL-KDD dataset. Moreover, it attains an accuracy and f1-score exceeding 98.9% on the AWID dataset.