Adversarial Classifier Research Articles

Leveraging Temporal Trends for Training Contextual Word Embeddings to Address Bias in Biomedical Applications: Development Study.

Women have been underrepresented in clinical trials for many years. Machine-learning models trained on clinical trial abstracts may capture and amplify biases in the data. Specifically, word embeddings are models that enable representing words as vectors and are the building block of most natural language processing systems. If word embeddings are trained on clinical trial abstracts, predictive models that use the embeddings will exhibit gender performance gaps. We aim to capture temporal trends in clinical trials through temporal distribution matching on contextual word embeddings (specifically, BERT) and explore its effect on the bias manifested in downstream tasks. We present TeDi-BERT, a method to harness the temporal trend of increasing women's inclusion in clinical trials to train contextual word embeddings. We implement temporal distribution matching through an adversarial classifier, trying to distinguish old from new clinical trial abstracts based on their embeddings. The temporal distribution matching acts as a form of domain adaptation from older to more recent clinical trials. We evaluate our model on 2 clinical tasks: prediction of unplanned readmission to the intensive care unit and hospital length of stay prediction. We also conduct an algorithmic analysis of the proposed method. In readmission prediction, TeDi-BERT achieved area under the receiver operating characteristic curve of 0.64 for female patients versus the baseline of 0.62 (P<.001), and 0.66 for male patients versus the baseline of 0.64 (P<.001). In the length of stay regression, TeDi-BERT achieved a mean absolute error of 4.56 (95% CI 4.44-4.68) for female patients versus 4.62 (95% CI 4.50-4.74, P<.001) and 4.54 (95% CI 4.44-4.65) for male patients versus 4.6 (95% CI 4.50-4.71, P<.001). In both clinical tasks, TeDi-BERT improved performance for female patients, as expected; but it also improved performance for male patients. Our results show that accuracy for one gender does not need to be exchanged for bias reduction, but rather that good science improves clinical results for all. Contextual word embedding models trained to capture temporal trends can help mitigate the effects of bias that changes over time in the training data.

Multi-source domain adversarial graph convolutional networks for rolling mill health states diagnosis under variable working conditions

As the rolling mill often encounters variable and complicated working conditions and shock loads, unsupervised domain adaptive (UDA) methods are imperative in its health monitoring. However, efforts of applying UDA methods on the rolling mill are negligible, and many existing approaches have constraints in domain adaptation, domain label, and data construction that prevent meaningful features from being extracted. Hence, a multi-source domain adversarial graph convolutional networks framework (MSDAGCNs) is presented to overcome these challenges and combine three essential elements to achieve cross-domain health states diagnosis under variable working conditions. First, a shared feature extract module is introduced to extract common features. Then, the features are input to a multi-source feature extract module to extract the data construction from the graphs generated by a graph construction module. Meanwhile, a multi-source domain adversarial classifier module is modeled to extract multi-source invariant features and classify them. After that, the local maximum mean discrepancy is employed to align the domain categories. Next, a task classifier module integrates the results of the multi-source classifier for reliable health state diagnosis. Results on the two cases can verify that the proposed MSDAGCNs can not only outperform other state-of-the-art methods, but also extract domain-invariant knowledge. Compared with the best-performing method, the proposed method can boost accuracy by 0.53% and 0.83% in the simplest task of the two case studies, respectively. Furthermore, the arrangement of sensors on the rolling mill is discussed to select the optimal location for collecting vibrations.

A deep transfer learning network with two classifiers based on sample selection for motor imagery brain-computer interface

The non-stationary of Motor Imagery (MI) electroencephalogram (EEG) signals makes the traditional machine learning methods ineffective in EEG recognition across time, which limits the practicability of Motor Imagery Brain-computer Interface(MI BCI). Transfer learning makes the classification model of non-stationary MI EEG signals collected and trained at different times universal. In this paper, we proposed a new deep transfer neural network model that can effectively utilize labeled EEG data from previous times, thereby achieving good recognition performance for a small number of labeled EEG signals at the current time. Firstly, a filter bank was used to filter the original motor imagery EEG, and the filtered EEG retained the key features of each domain through the domain specific attention module. Then designed two domain adaptation modules simultaneously. The first domain adaptation module selected similar source domain data, and the second domain adaptation module minimized the difference between the source and target domains. At last, two adversarial classifiers were used to improve the accuracy and robustness of the classification. The results of average classification accuracy, feature visualization, and paired t-test on two public datasets all indicated that the classification performance of the proposed method was significantly higher than other comparison methods. The experimental results on two public datasets showed that the proposed method performed better than the comparison methods when classifying only a small number of labeled MI EEG data, and can achieve higher classification accuracy. This method has practical application value in the rehabilitation of stroke patients.

ScDREAMER for atlas-level integration of single-cell datasets using deep generative model paired with adversarial classifier

Integration of heterogeneous single-cell sequencing datasets generated across multiple tissue locations, time, and conditions is essential for a comprehensive understanding of the cellular states and expression programs underlying complex biological systems. Here, we present scDREAMER (https://github.com/Zafar-Lab/scDREAMER), a data-integration framework that employs deep generative models and adversarial training for both unsupervised and supervised (scDREAMER-Sup) integration of multiple batches. Using six real benchmarking datasets, we demonstrate that scDREAMER can overcome critical challenges including skewed cell type distribution among batches, nested batch-effects, large number of batches and conservation of development trajectory across batches. Our experiments also show that scDREAMER and scDREAMER-Sup outperform state-of-the-art unsupervised and supervised integration methods respectively in batch-correction and conservation of biological variation. Using a 1 million cells dataset, we demonstrate that scDREAMER is scalable and can perform atlas-level cross-species (e.g., human and mouse) integration while being faster than other deep-learning-based methods.

Adversarially-Regularized Mixed Effects Deep Learning (ARMED) Models Improve Interpretability, Performance, and Generalization on Clustered (non-iid) Data.

Natural science datasets frequently violate assumptions of independence. Samples may be clustered (e.g., by study site, subject, or experimental batch), leading to spurious associations, poor model fitting, and confounded analyses. While largely unaddressed in deep learning, this problem has been handled in the statistics community through mixed effects models, which separate cluster-invariant fixed effects from cluster-specific random effects. We propose a general-purpose framework for Adversarially-Regularized Mixed Effects Deep learning (ARMED) models through non-intrusive additions to existing neural networks: 1) an adversarial classifier constraining the original model to learn only cluster-invariant features, 2) a random effects subnetwork capturing cluster-specific features, and 3) an approach to apply random effects to clusters unseen during training. We apply ARMED to dense, convolutional, and autoencoder neural networks on 4 datasets including simulated nonlinear data, dementia prognosis and diagnosis, and live-cell image analysis. Compared to prior techniques, ARMED models better distinguish confounded from true associations in simulations and learn more biologically plausible features in clinical applications. They can also quantify inter-cluster variance and visualize cluster effects in data. Finally, ARMED matches or improves performance on data from clusters seen during training (5-28% relative improvement) and generalization to unseen clusters (2-9% relative improvement) versus conventional models.

Sequential Variational Autoencoder with Adversarial Classifier for Video Disentanglement

In this paper, we propose a sequential variational autoencoder for video disentanglement, which is a representation learning method that can be used to separately extract static and dynamic features from videos. Building sequential variational autoencoders with a two-stream architecture induces inductive bias for video disentanglement. However, our preliminary experiment demonstrated that the two-stream architecture is insufficient for video disentanglement because static features frequently contain dynamic features. Additionally, we found that dynamic features are not discriminative in the latent space. To address these problems, we introduced an adversarial classifier using supervised learning into the two-stream architecture. The strong inductive bias through supervision separates dynamic features from static features and yields discriminative representations of the dynamic features. Through a comparison with other sequential variational autoencoders, we qualitatively and quantitatively demonstrate the effectiveness of the proposed method on the Sprites and MUG datasets.

Multi-label emotion classification based on adversarial multi-task learning

In this paper, we focus on the task of multi-label emotion classification and aim to tackle two problems of this task. First, few studies try to exploit the correlation among different emotions, which motivates us to introduce the task-specific information into the shared hidden layer. Second, the public multi-label emotion datasets for low-resource languages are limited. To overcome these problems, we propose a novel multi-task multi-label emotion classification. Our approach consists of three components: general representation module, emotion representation module and adversarial classifier. The model applies emotion descriptors to incorporate the correlation among different emotions, and then uses adversarial training to prevent too much emotion-relevant information from being injected into the shared layer. Extensive experiments demonstrate that our approach outperforms the state-of-the-art baselines across a variety of evaluation metrics, achieving macro-average F1 scores of 50.21%, 41.33% and 40.24% on the Chinese, English, and Indonesian datasets, respectively. In addition, an Indonesian dataset and an English one containing 4207 and 26,019 samples respectively, are constructed for the multi-label emotion classification task. The datasets will be publicly available and we believe this can support the future study of Indonesian multi-label emotion recognition resources, which are limited for the related research fields now. We make our codes and resources in this work publicly available on: https://github.com/GKLMIP/MLEC-AML.

Dual constraints and adversarial learning for fair recommenders

Recommender systems, which are consist of common artificial intelligence technology, have a profound impact on the lifestyles of people. However, recent studies have demonstrated that recommender systems have fairness problems which means that some people with certain attributes are treated unfairly. A fair recommender means that users with different attributes achieve the same recommender accuracy. In particular, the recommender systems completely rely on users’ behavior data for preferences learning, which leads to a high probability of unfair problems because that the behavior data usually contains sensitive information of users. Unfortunately, there are a few studies exploring unfair problem in recommender systems. To alleviate this problem, we present a novel fairness-aware recommender with dual fairness constraints (FRFC) to improve fairness in recommendations and protect the user’s sensitive information from being exposed. This model has several advantages: one advantage is that an adversarial-based graph neural network (GNN) is proposed to prevent the target user being infected by sensitive features of neighbor users; another advantage is that two fairness constraints are proposed to solve the problems of adversarial classifier failures in whole data and unfair ranking losses. With this design, the FRFC model can effectively filter out users’ sensitive information and give users of different attributes the same training opportunities, which is helpful for making a fair recommendation. Finally, extensive experiments demonstrate that the proposed model can significantly improve the fairness of recommendation results.

Adversarial Risk via Optimal Transport and Optimal Couplings

Modern machine learning algorithms perform poorly on adversarially manipulated data. Adversarial risk quantifies the error of classifiers in adversarial settings; adversarial classifiers minimize adversarial risk. In this paper, we analyze adversarial risk and adversarial classifiers from an optimal transport perspective. We show that the optimal adversarial risk for binary classification with 0-1 loss is determined by an optimal transport cost between the probability distributions of the two classes. We develop optimal transport plans (probabilistic couplings) for univariate distributions such as the normal, the uniform, and the triangular distribution. We also derive optimal adversarial classifiers in these settings. Our analysis leads to algorithm-independent fundamental limits on adversarial risk, which we calculate for several real-world datasets. We extend our results to general loss functions under convexity and smoothness assumptions.

Agreement-Discrepancy-Selection: Active Learning with Progressive Distribution Alignment

In active learning, the ignorance of aligning unlabeled samples' distribution with that of labeled samples hinders the model trained upon labeled samples from selecting informative unlabeled samples. In this paper, we propose an agreement-discrepancy-selection (ADS) approach, and target at unifying distribution alignment with sample selection by introducing adversarial classifiers to the convolutional neural network (CNN). Minimizing classifiers' prediction discrepancy (maximizing prediction agreement) drives learning CNN features to reduce the distribution bias of labeled and unlabeled samples, while maximizing classifiers' discrepancy highlights informative samples. Iterative optimization of agreement and discrepancy loss calibrated with an entropy function drives aligning sample distributions in a progressive fashion for effective active learning. Experiments on image classification and object detection tasks demonstrate that ADS is task-agnostic, while significantly outperforms the previous methods when the labeled sets are small.

PSAT-GAN: Efficient Adversarial Attacks Against Holistic Scene Understanding.

Recent advances in deep neural networks (DNNs) have facilitated high-end applications, including holistic scene understanding (HSU), in which many tasks run in parallel with the same visual input. Following this trend, various methods have been proposed to use DNNs to perform multiple vision tasks. However, these methods are task-specific and less effective when considering multiple HSU tasks. End-to-end demonstrations of adversarial examples, which generate one-to-many heterogeneous adversarial examples in parallel from the same input, are scarce. Additionally, one-to-many mapping of adversarial examples for HSU usually requires joint representation learning and flexible constraints on magnitude, which can render the prevalent attack methods ineffective. In this paper, we propose PSAT-GAN, an end-to-end framework that follows the pipeline of HSU. It is based on a mixture of generative models and an adversarial classifier that employs partial weight sharing to learn a one-to-many mapping of adversarial examples in parallel, each of which is effective for its corresponding task in HSU attacks. PSAT-GAN is further enhanced by applying novel adversarial and soft-constraint losses to generate effective perturbations and avoid studying transferability. Experimental results indicate that our method is efficient in generating both universal and image-dependent adversarial examples to fool HSU tasks under either targeted or non-targeted settings.

ClsGAN: Selective Attribute Editing Model based on Classification Adversarial Network

Attribution editing has achieved remarkable progress in recent years owing to the encoder–decoder structure and generative adversarial network (GAN). However, it remains challenging to generate high-quality images with accurate attribute transformation. Attacking these problems, the work proposes a novel selective attribute editing model based on classification adversarial network (referred to as ClsGAN) that shows good balance between attribute transfer accuracy and photo-realistic images. Considering that the editing images are prone to be affected by original attribute due to skip-connection in encoder–decoder structure, an upper convolution residual network (referred to as Tr-resnet) is presented to selectively extract information from the source image and target label. In addition, to further improve the transfer accuracy of generated images, an attribute adversarial classifier (referred to as Atta-cls) is introduced to guide the generator from the perspective of attribute through learning the defects of attribute transfer images. Experimental results on CelebA demonstrate that our ClsGAN performs favorably against state-of-the-art approaches in image quality and transfer accuracy. Moreover, ablation studies are also designed to verify the great performance of Tr-resnet and Atta-cls.

Generative adversarial classifier for handwriting characters super-resolution

Generative Adversarial Networks (GAN) receive great attention recently due to its excellent performance in image generation, transformation, and super-resolution. However, less emphasis or study has been put on GAN for classification with super-resolution. Moreover, though GANs may fabricate images which perceptually looks realistic, they usually fabricate some fake details especially in character data; this would impose further difficulties when they are input for classification. In this paper, we propose a novel Generative Adversarial Classifier (GAC) for low-resolution handwriting character recognition. Specifically, we design an additional classifier component in GAC, leading to a novel three-player GAN model which is not only able to generate high-quality super-resolved images, but also favorable for classification. Experimental results show that our proposed method can obtain remarkable performance in handwriting characters with 8 × super-resolution, achieving new state-of-the-art on benchmark dataset CASIA-HWDB1.1, and MNIST.

ASNM Datasets: A Collection of Network Attacks for Testing of Adversarial Classifiers and Intrusion Detectors

In this paper, we present three datasets that have been built from network traffic traces using ASNM features, designed in our previous work. The first dataset was built using a state-of-the-art dataset called CDX 2009, while the remaining two datasets were collected by us in 2015 and 2018, respectively. These two datasets contain several adversarial obfuscation techniques that were applied onto malicious as well as legitimate traffic samples during the execution of particular TCP network connections. Adversarial obfuscation techniques were used for evading machine learning-based network intrusion detection classifiers. Further, we showed that the performance of such classifiers can be improved when partially augmenting their training data by samples obtained from obfuscation techniques. In detail, we utilized tunneling obfuscation in HTTP(S) protocol and non-payload-based obfuscations modifying various properties of network traffic by, e.g., TCP segmentation, re-transmissions, corrupting and reordering of packets, etc. To the best of our knowledge, this is the first collection of network traffic metadata that contains adversarial techniques and is intended for non-payload-based network intrusion detection and adversarial classification. Provided datasets enable testing of the evasion resistance of arbitrary classifier that is using ASNM features.

Style-Neutralized Pattern Classification Based on Adversarially Trained Upgraded U-Net

Traditional machine learning approaches usually hold the assumption that data for model training and in real applications are created following the identical and independent distribution (i.i.d.). However, several relevant research topics have demonstrated that such condition may not always describe the real scenarios. One particular case is that the patterns are equipped with diverse and changeable style information. In this paper, a novel classification framework named Style Neutralization Generative Adversarial Classifier (SN-GAC), based on an upgraded U-Net architecture, and trained adversarially with the Generative Adversarial Network (GAN) framework, is introduced to accomplish the classification in such disparate and inconsistent data information case. The generative model in SN-GAC neutralizes style information from the original style-discriminative patterns (style-source) by building the mapping function from them to their style-free counterparts (corresponding standard examples, standard-target). A well-learned generator in the SN-GAC framework is capable of producing the targeted style-neutralized data (generated-target), satisfying the i.i.d. condition. Additionally, SN-GAC is trained adversarially, where an independent discriminator is used to surveil and supervise the training progress of the above-mentioned generator by distinguishing between the real and the generated. Simultaneously, an auxiliary classifier is also embedded in the discriminator to assign the correct class label of both the real and generated data. This process proves effective to aid the generator to produce high-quality human-readable style-neutralized patterns. It will then be further fine-tuned for the sake of promoting the final classification performance. Extensive experiments have adequately demonstrated the effectiveness of the proposed SN-GAC framework: it outperforms several relevant state-of-the-art baselines on two empirical data sets in the non-i.i.d. data classification task.