Adversarial Learning Research Articles

Enhancing zero-shot object detection with external knowledge-guided robust contrast learning

Zero-shot object detection aims to identify objects from unseen categories not present during training. Existing methods rely on category labels to create pseudo-features for unseen categories, but they face limitations in exploring semantic information and lack robustness. To address these issues, we introduce a novel framework, EKZSD, enhancing zero-shot object detection by incorporating external knowledge and contrastive paradigms. This framework enriches semantic diversity, enhancing discriminative ability and robustness. Specifically, we introduce a novel external knowledge extraction module that leverages attribute and relationship prompts to enrich semantic information. Moreover, a novel external knowledge contrastive learning module is proposed to enhance the model’s discriminative and robust capabilities by exploring pseudo-visual features. Additionally, we use cycle consistency learning to align generated visual features with original semantic features and adversarial learning to align visual features with semantic features. Collaboratively trained with contrast learning loss, cycle consistency loss, adversarial learning loss, and classification loss, our framework outperforms superior performance on the MSCOCO and Ship-43 datasets, as demonstrated in experimental results.

DACA: A domain adaptive fault diagnosis approach with class-aware based on cross-domain extreme imbalance data

Existing research on unsupervised domain adaptation (UDA) has primarily centered on mitigating differences in feature distribution across various domains. However, in real-world scenarios, natural data often exhibits not only a disparity in instance counts across different classes within the same domain but also across different domains, leading to imbalances at both the class and domain levels. To address this issue, this study introduces a novel scenario, termed extreme imbalance domain adaptation (EIDA), which encompasses both covariate shifts and labeling shifts. We propose a Domain Adaptation with Class-aware (DACA) framework for bearing fault diagnosis that aims to tackle inter-domain feature shifts as well as intra- and inter-domain category imbalances. Specifically, DACA enhances the recognition accuracy of minority categories without degrading the performance concerning the majority category, reduces the adverse effects of data imbalance during domain transitions, and strengthens the robustness of decision boundaries for improved spatial matching. This is achieved through four sub-modules: adversarial clustering learning, cross-domain category alignment, category-level reweighting, and category-level re-margining. The efficacy of DACA is demonstrated through simulations of three imbalance protocols in the EIDA scenario across two bearing datasets and one gearbox dataset, showcasing significant advancements in cross-domain diagnostic tasks.

Endo-atmospheric maneuver penetration strategy based on generative adversarial reinforcement learning

An intelligent endo-atmospheric penetration strategy based on generative adversarial reinforcement learning is proposed in this manuscript. Firstly, attack and defense adversarial models are established, and missile maneuver penetration problem is transformed into an optimal control problem, considering penetration, handover position and mid-terminal guidance velocity constraints. Then, Radau Pseudospectral method is adopted to generate data samples considering random perturbations. Furthermore, Generative Adversarial Imitation Learning Combined with Deep Deterministic Policy Gradient method (GAIL-DDPG) is designed, with internal process reward signals constructed to tackle long-term sparse reward in missile manuver penetration problem. Finally, penetration strategy is trained and verified. Simulation shows that using generative adversarial reinforcement learning, with sample library to learn expert experience in training early stage, the proposed method can quickly converge. Also, performance is further optimized with reinforcement learning exploration strategy in the later stage of training. Simulation shows that the proposed method has better engineering application ability compared with traditional reinforcement learning method.

Two mass-imbalanced atoms in a hard-wall trap: Deep learning integrability of many-body systems.

The study of integrable systems has led to significant advancements in our understanding of many-body physics. We design a series of numerical experiments to analyze the integrability of a mass-imbalanced two-body system through energy-level statistics and deep learning of wave functions. The level spacing distributions are fitted by a Brody distribution and the fitting parameter ω is found to separate the integrable and nonintegrable mass ratios by a critical line ω=0. The convolutional neural network built from the probability density images could identify the transition points between integrable and nonintegrable systems with high accuracy, yet in a much shorter computation time. A brilliant example of the network's ability is to identify a new integrable mass ratio 1/3 by learning from the known integrable case of equal mass, with a remarkable network confidence of 98.22%. The robustness of our neural networks is further enhanced by adversarial learning, where samples are generated by standard and quantum perturbations mixed in the probability density images and the wave functions, respectively.

An adversarial network reinforcement method for power data time series generation based on knowledge transfer and Rayleigh differential privacy

Abstract In order to solve the problem of sensitive information leakage of power data in machine learning and artificial intelligence processing, this paper proposes a method of civic-text data reinforcement based on time series generation adversarial network, which combines knowledge transfer and Rényi differential privacy protection framework to balance privacy and availability, while improving the accuracy of the model. First, the generative adversarial network generates training data that is similar to the real data but secure through adversarial learning, ensuring that even if the data is stolen, the original information cannot be inferred. Secondly, through the integration of recurrent neural network training and autoencoder, the dynamic and static characteristics of the power load data can be effectively captured, and the high dimension of the learning space can be reduced. In addition, knowledge distillation technology is used to guide the generator of training student models through a powerful teacher network of knowledge transfer, improving the diversity and quality of the generated data. Finally, from the four aspects of privacy, data similarity and diversity, and practicability, the practical application value and potential of the framework for electricity consumption data are comprehensively evaluated.

BeTAIL: Behavior Transformer Adversarial Imitation Learning From Human Racing Gameplay

BeTAIL: Behavior Transformer Adversarial Imitation Learning From Human Racing Gameplay

Multi-source subdomain negative transfer suppression and multiple pseudo-labels guidance alignment: a method for fault diagnosis under cross-working conditions

Extensive researches have been conducted on transfer learning based fault diagnosis. However, negative information transfer may arise due to significant differences in the subdomain distribution across multiple source domains (MSDs). Most existing methods focus solely on the impact of subdomains from a single source domain (SSD) on the target domain (TD). Therefore, this paper proposed a novel multi-stage alignment multi-source subdomain adaptation (MAMSA) method. The global feature extractor is designed to extract domain-invariant features. Three domain-specific feature extractors capture high-level fault features from different domains with a customized adaptation strategy, which combines adversarial learning and distribution alignment based on multiple pseudo-label-guided local maximum mean discrepancy (MP-LMMD) to learn subdomain-invariant features. MP-LMMD utilizes pseudo-labels generated from all classifiers in the TD to guide the alignment of subdomains, suppressing negative transfer from the source domains (SDs). The experimental results indicate that the MAMSA method has excellent capabilities to suppress negative transfer, and the diagnostic performance can be greatly promoted with MAMSA under cross-working conditions.

Adversarial EM for variational deep learning: Application to semi-supervised image quality enhancement in low-dose PET and low-dose CT

In positron emission tomography (PET) and X-ray computed tomography (CT), reducing radiation dose can cause significant degradation in image quality. For image quality enhancement in low-dose PET and CT, we propose a novel theoretical adversarial and variational deep neural network (DNN) framework relying on expectation maximization (EM) based learning, termed adversarial EM (AdvEM). AdvEM proposes an encoder–decoder architecture with a multiscale latent space, and generalized-Gaussian models enabling datum-specific robust statistical modeling in latent space and image space. The model robustness is further enhanced by including adversarial learning in the training protocol. Unlike typical variational-DNN learning, AdvEM proposes latent-space sampling from the posterior distribution, and uses a Metropolis–Hastings scheme. Unlike existing schemes for PET or CT image enhancement which train using pairs of low-dose images with their corresponding normal-dose versions, we propose a semi-supervised AdvEM (ssAdvEM) framework that enables learning using a small number of normal-dose images. AdvEM and ssAdvEM enable per-pixel uncertainty estimates for their outputs. Empirical analyses on real-world PET and CT data involving many baselines, out-of-distribution data, and ablation studies show the benefits of the proposed framework.

Unsupervised domain adaptive building semantic segmentation network by edge-enhanced contrastive learning

Unsupervised domain adaptation (UDA) is a weakly supervised learning technique that classifies images in the target domain when the source domain has labeled samples, and the target domain has unlabeled samples. Due to the complexity of imaging conditions and the content of remote sensing images, the use of UDA to accurately extract artificial features such as buildings from high-spatial-resolution (HSR) imagery is still challenging. In this study, we propose a new UDA method for building extraction, the contrastive domain adaptation network (CDANet), by utilizing adversarial learning and contrastive learning techniques. CDANet consists of a single multitask generator and dual discriminators. The generator employs a region and edge dual-branch structure that strengthens its edge extraction ability and is beneficial for the extraction of small and densely distributed buildings. The dual discriminators receive the region and edge prediction outputs and achieve multilevel adversarial learning. During adversarial training processing, CDANet aligns the cross-domain of similar pixel features in the embedding space by constructing the regional pixelwise contrastive loss. A self-training (ST) strategy based on pseudolabel generation is further utilized to address the target intradomain discrepancy. Comprehensive experiments are conducted to validate CDANet on three publicly accessible datasets, namely the WHU, Austin, and Massachusetts. Ablation experiments show that the generator network structure, contrastive loss and ST strategy all improve the building extraction accuracy. Method comparisons validate that CDANet achieves superior performance to several state-of-the-art methods, including AdaptSegNet, AdvEnt, IntraDA, FDANet and ADRS, in terms of F1 score and mIoU.

A medical image classification method based on self-regularized adversarial learning.

Deep learning (DL) techniques have been extensively applied in medical image classification. The unique characteristics of medical imaging data present challenges, including small labeled datasets, severely imbalanced class distribution, and significant variations in imaging quality. Recently, generative adversarial network (GAN)-based classification methods have gained attention for their ability to enhance classification accuracy by incorporating realistic GAN-generated images as data augmentation. However, the performance of these GAN-based methods often relies on high-quality generated images, while large amounts of training data are required to train GAN models to achieve optimalperformance. In this study, we propose an adversarial learning-based classification framework to achieve better classification performance. Innovatively, GAN models are employed as supplementary regularization terms to support classification, aiming to address the challenges describedabove. The proposed classification framework, GAN-DL, consists of a feature extraction network (F-Net), a classifier, and two adversarial networks, specifically a reconstruction network (R-Net) and a discriminator network (D-Net). The F-Net extracts features from input images, and the classifier uses these features for classification tasks. R-Net and D-Net have been designed following the GAN architecture. R-Net employs the extracted feature to reconstruct the original images, while D-Net is tasked with the discrimination between the reconstructed image and the original images. An iterative adversarial learning strategy is designed to guide model training by incorporating multiple network-specific loss functions. These loss functions, serving as supplementary regularization, are automatically derived during the reconstruction process and require no additional dataannotation. To verify the model's effectiveness, we performed experiments on two datasets, including a COVID-19 dataset with 13958 chest x-ray images and an oropharyngeal squamous cell carcinoma (OPSCC) dataset with 3255 positron emission tomography images. Thirteen classic DL-based classification methods were implemented on the same datasets for comparison. Performance metrics included precision, sensitivity, specificity, and -score. In addition, we conducted ablation studies to assess the effects of various factors on model performance, including the network depth of F-Net, training image size, training dataset size, and loss function design. Our method achieved superior performance than all comparative methods. On the COVID-19 dataset, our method achieved , , , and in terms of precision, sensitivity, specificity, and -score, respectively. It achieved across all these metrics on the OPSCC dataset. The study to investigate the effects of two adversarial networks highlights the crucial role of D-Net in improving model performance. Ablation studies further provide an in-depth understanding of ourmethodology. Our adversarial-based classification framework leverages GAN-based adversarial networks and an iterative adversarial learning strategy to harness supplementary regularization during training. This design significantly enhances classification accuracy and mitigates overfitting issues in medical image datasets. Moreover, its modular design not only demonstrates flexibility but also indicates its potential applicability to various clinical contexts and medical imaging applications.

Подход к настройке CatBoost для продвинутого обнаружения атак DoS и DDoS в сетевом трафике

In the ever-evolving landscape of network security, the sophistication of cyber-attacks, especially Denial of Service (DoS) and Distributed Denial of Service (DDoS) attacks, poses a formidable challenge to intrusion detection systems. Recognizing the longstanding application of CatBoost in various domains, this study explores its novel optimization for network intrusion detection, a critical area in need of advanced solutions. Leveraging the strengths of CatBoost in handling categorical data and imbalanced datasets, we meticulously adapt the classifier to meet the complex demands of distinguishing between DoS, DDoS, and benign traffic within the comprehensive CICIDS2017 and CSE-CIC-IDS2018 datasets. This research is an attempt to refine the learning efficiency and detection capabilities of CatBoost through the implementation of advanced feature selection and data preparation, contributing to the field by improving detection accuracy within real-time intrusion detection systems. The results show a notable improvement in performance, underscoring the classifier's role in advancing cybersecurity measures. Furthermore, the study paves the way for future exploration into adversarial machine learning and automated feature engineering, fortifying the resilience and adaptability of intrusion detection systems against the backdrop of a rapidly changing cyber threat landscape. These efforts provide solid approaches to address the current challenges in network security, signaling a move towards more refined and dependable intrusion detection methods.

ASD-GANNet: A Generative Adversarial Network-Inspired Deep Learning Approach for the Classification of Autism Brain Disorder.

The classification of a pre-processed fMRI dataset using functional connectivity (FC)-based features is considered a challenging task because of the set of high-dimensional FC features and the small dataset size. To tackle this specific set of FC high-dimensional features and a small-sized dataset, we propose here a conditional Generative Adversarial Network (cGAN)-based dataset augmenter to first train the cGAN on computed connectivity features of NYU dataset and use the trained cGAN to generate synthetic connectivity features per category. After obtaining a sufficient number of connectivity features per category, a Multi-Head attention mechanism is used as a head for the classification. We name our proposed approach "ASD-GANNet", which is end-to-end and does not require hand-crafted features, as the Multi-Head attention mechanism focuses on the features that are more relevant. Moreover, we compare our results with the six available state-of-the-art techniques from the literature. Our proposed approach results using the "NYU" site as a training set for generating a cGAN-based synthetic dataset are promising. We achieve an overall 10-fold cross-validation-based accuracy of 82%, sensitivity of 82%, and specificity of 81%, outperforming available state-of-the art approaches. A sitewise comparison of our proposed approach also outperforms the available state-of-the-art, as out of the 17 sites, our proposed approach has better results in the 10 sites.

Open-set domain adaptive fault diagnosis based on supervised contrastive learning and a complementary weighted dual adversarial network

In an actual industrial environment, the complex working environment of mechanical equipment may lead to new faults in the target domain, called the open-set domain adaptation problem. Recently, open-set adaptive fault diagnosis has been extensively employed. However, most studies not only require pre-set fixed thresholds to identify unknown class features but also ignore the learning of discriminable features under specific tasks, which affects the diagnostic performance. Hence, this paper proposes a complementary weighted dual adversarial network combined with supervised contrastive learning (CWDAN-SCL) to address the open-set cross-different working fault diagnosis of bearings. Specifically, a novel complementary weighted adversarial learning strategy is designed using supervised classification and uncertainty measurement to effectively control the participation of target domain features in the domain adaptation process and achieve the alignment of shared class fault features between the source and target domains. Moreover, an adaptive unknown fault separation module is designed using an adversarial learning method to construct a hyperplane between shared and unknown class fault features in the target domain to identify unknown class faults accurately. Additionally, a supervised contrastive loss term is designed based on contrastive learning and label knowledge to improve the aggregation of fault features of the same class and enhance the model’s generalization ability in target domain diagnosis tasks. Subsequently, the efficacy and advancement of the proposed method are substantiated through experimentation on two datasets. The experimental results illustrate that the average diagnostic performance of the proposed method is 91.73 %. This study contributes a dependable diagnostic approach for ascertaining the health status of rotating machinery equipment.

Prior image-based generative adversarial learning for multi-material decomposition in photon counting computed tomography

BackgroundPhoton counting detector computed tomography (PCD-CT) is a novel promising technique providing higher spatial resolution, lower radiation dose and greater energy spectrum differentiation, which create more possibilities to improve image quality. Multi-material decomposition is an attractive application for PCD-CT to identify complicated materials and provide accurate quantitative analysis. However, limited by the finite photon counting rate in each energy window of photon counting detector, the noise problem hinders the decomposition of high-quality basis material images. MethodsTo address this issue, an end-to-end multi-material decomposition network based on prior images is proposed in this paper. First, the reconstructed images corresponding to the full spectrum with less noise are introduced as prior information to improve the overall signal-to-noise ratio of the data. Then, a generative adversarial network is designed to mine the relationship between reconstructed images and basis material images based on the information interaction of material decomposition. Furthermore, a weighted edge loss is introduced to adapt to the structural differences of different basis material images. ResultsTo verify the performance of the proposed method, simulation and real studies are carried out. In simulation study of structured fibro-glandular tissue model, the results show that the proposed method decreased the root mean square error by 67 % and 26 % on adipose, 66 % and 28 % on fibroglandular, 52 % and 8 % on calcification, compared to butterfly network and dual interactive Wasserstein generative adversarial network. ConclusionExperimentally, the proposed method shows certain advantages over other methods on noise suppression effect, detail retention ability and decomposition accuracy.

Causality-inspired Domain Expansion network for single domain generalization

Most existing single domain generalization (SDG) methods focus on extending the distribution of the training data to achieve good generalizability in the target domain by synthesizing out-of-domain data. However, in practice, the generated data are insufficient to cover all possible shifts. When the target distributions are not covered, the spurious associations between irrelevant features and labels captured by a model will deteriorate the generalization performance. To alleviate this problem, in this paper, we first analyze that augmenting the source domain with samples of new styles (e.g., different backgrounds) can mitigate spurious associations between irrelevant features and labels from a causal perspective, which can reduce the interference of non-causal features on the classification labels then propose a novel Causality-inspired Domain Expansion Network (CDEN) for SDG. Specifically, CDEN contains a domain augmentation subnetwork and a representation learning subnetwork. The former uses a special style generator module AdaIN∗ to generate out-of-domain data with the same semantic information but different styles from the source domain data to help the model capture true associations between discriminative features and labels. The latter introduces a novel class-level contrastive learning, instance-level contrastive learning, and invariant risk minimization to help the model further eliminate spurious associations. CDEN progressively generates diverse style data and learns domain invariant representations containing few spurious associations by optimizing the two subnetworks in an adversarial learning manner. We conduct extensive experiments on three benchmark datasets to evaluate the superiority of CDEN, comparing it to fifteen state-of-the-art methods.

Inductive State-Relabeling Adversarial Active Learning with Heuristic Clique Rescaling.

Active learning (AL) is to design label-efficient algorithms by labeling the most representative samples. It reduces annotation cost and attracts increasing attention from the community. However, previous AL methods suffer from the inadequacy of annotations and unreliable uncertainty estimation. Moreover, we find that they ignore the intra-diversity of selected samples, which leads to sampling redundancy. In view of these challenges, we propose an inductive state-relabeling adversarial AL model (ISRA) that consists of a unified representation generator, an inductive state-relabeling discriminator, and a heuristic clique rescaling module. The generator introduces contrastive learning to leverage unlabeled samples for self-supervised training, where the mutual information is utilized to improve the representation quality for AL selection. Then, we design an inductive uncertainty indicator to learn the state score from labeled data and relabel unlabeled data with different importance for better discrimination of instructive samples. To solve the problem of sampling redundancy, the heuristic clique rescaling module measures the intra-diversity of candidate samples and recurrently rescales them to select the most informative samples. The experiments conducted on eight datasets and two imbalanced scenarios show that our model outperforms the previous state-of-the-art AL methods. As an extension on the cross-modal AL task, we apply ISRA to the image captioning and it also achieves superior performance.

UAV air combat autonomous trajectory planning method based on robust adversarial reinforcement learning

The poor robustness of the air combat autonomous trajectory planning strategy (ATP) trained through vanilla reinforcement learning (RL) methods is attributed to its dependence on the source environment. When this ATP is transformed from the source to the target environment, it may fail to lead to an effective dogfight victory due to disturbances, which can potentially threaten the flight safety of the unmanned combat aerial vehicle (UCAV). In order to enhance the robustness of ATP, we propose a robust RL-based method for generating an ATP strategy. This method can effectively consider uncertainties in the state and potential model misspecifications. First, a jointly trained adversary is designed to apply disturbances in the environment, forming a two-player zero-sum game with the ATP agent. To solve this game problem and learn a robust ATP, the robust adversarial RL method (RARL) is applied. Additionally, in order to prevent non-convergence of strategies resulting from the introduction of the RARL algorithm, a curriculum learning method is proposed that emulates the way human pilots learn, gradually progressing from simpler to more challenging courses. Finally, a quantitative method for evaluating the robustness of the ATP is proposed, taking into consideration the differences in ATP performance between the source and target combat environments. The results of robustness estimations demonstrate that the RARL algorithm, which is proposed herein, effectively enhances the ATP's robustness.

VeriBin: A Malware Authorship Verification Approach for APT Tracking through Explainable and Functionality-Debiasing Adversarial Representation Learning

VeriBin: A Malware Authorship Verification Approach for APT Tracking through Explainable and Functionality-Debiasing Adversarial Representation Learning

Adversarial Machine Learning in Industry: A Systematic Literature Review

Adversarial Machine Learning (AML) discusses the act of attacking and defending Machine Learning (ML) Models, an essential building block of Artificial Intelligence (AI). ML is applied in many software-intensive products and services and introduces new opportunities and security challenges. AI and ML will gain even more attention from the industry in the future, but threats caused by already-discovered attacks specifically targeting ML models are either overseen, ignored, or mishandled. Current AML research investigates attack and defense scenarios for ML in different industrial settings with a varying degree of maturity with regard to academic rigor and practical relevance. However, to the best of our knowledge, a synthesis of the state of academic rigor and practical relevance is missing. This literature study reviews studies in the area of AML in the context of industry, measuring and analyzing each study’s rigor and relevance scores. Overall, all studies scored a high rigor score and a low relevance score, indicating that the studies are thoroughly designed and documented but miss the opportunity to include touch points relatable for practitioners.

Neural Schrödinger bridge for unpaired real-world image deraining

Given the significant differences between domains, current unpaired learning methods struggle to accurately map the relationship between rainy and clear images. To this end, we introduce a neural Schrödinger bridge (NSB) for unpaired real-world image deraining, which utilizes the stochastic differential equations to capture the mapping relationships between rainy and clear domains. Meanwhile, we frame the deraining process as a Lagrangian problem using the Kullback-Leibler divergence between the data distribution and the model distribution. Additionally, by leveraging the capabilities of the contrastive language-image pre-training model (CLIP), our research shows that the CLIP prior helps differentiate between rainy and clear images. Building on this, we reformulate the Schrödinger bridge problem as a series of adversarial learning tasks using both image and prompt representations. To our knowledge, our approach is the first to use the Schrödinger bridge in unpaired image deraining. Extensive experiments show that our proposed NSB model outperforms existing unpaired deraining methods in both quantitative and qualitative evaluations.