Gradient Backpropagation Research Articles

A dual-domain cone beam computed tomography sparse-view reconstruction method based on generative projection interpolation

To propose a dual-domain CBCT reconstruction framework (DualSFR-Net) based on generative projection interpolation to reduce artifacts in sparse-view cone beam computed tomography (CBCT) reconstruction. The proposed method DualSFR-Net consists of a generative projection interpolation module, a domain transformation module, and an image restoration module. The generative projection interpolation module includes a sparse projection interpolation network (SPINet) based on a generative adversarial network and a full-view projection restoration network (FPRNet). SPINet performs projection interpolation to synthesize full-view projection data from the sparse-view projection data, while FPRNet further restores the synthesized full-view projection data. The domain transformation module introduces the FDK reconstruction and forward projection operators to complete the forward and gradient backpropagation processes. The image restoration module includes an image restoration network FIRNet that fine-tunes the domain-transformed images to eliminate residual artifacts and noise. Validation experiments conducted on a dental CT dataset demonstrated that DualSFR-Net was capable to reconstruct high-quality CBCT images under sparse-view sampling protocols. Quantitatively, compared to the current best methods, the DualSFR-Net method improved the PSNR by 0.6615 and 0.7658 and increased the SSIM by 0.0053 and 0.0134 under 2-fold and 4-fold sparse protocols, respectively. The proposed generative projection interpolation-based dual-domain CBCT sparse-view reconstruction method can effectively reduce stripe artifacts to improve image quality and enables efficient joint training for dual-domain imaging networks in sparse-view CBCT reconstruction.

Driver behavior recognition based on dual-branch and deformable convolutional network method

Aiming at the task of driver behavior recognition in the car cockpit, this paper proposes a recognition method based on a dual-branch neural network. The main branch of the network model uses ResNet50 as the backbone network for feature extraction, and uses deformable convolution to adapt the model to the shape and position changes of the driver in the image. The auxiliary branch assists in updating the parameters of the backbone network during the gradient backpropagation process, so that the backbone network can better extract features that are conducive to driver behavior recognition, thereby improving the recognition performance of the model. The ablation experiment and comparative experiment results of the network model on the State Farm public dataset show that the recognition accuracy of the proposed network model can reach 96.23%, and the recognition effect is better for easily confused behavior categories. The research results are of great significance for understanding driver behavior in the car cockpit and ensuring driving safety.

OligoFormer: an accurate and robust prediction method for siRNA design.

RNA interference (RNAi) has become a widely used experimental approach for post-transcriptional regulation and is increasingly showing its potential as future targeted drugs. However, the prediction of highly efficient siRNAs (small interfering RNAs) is still hindered by dataset biases, the inadequacy of prediction methods, and the presence of off-target effects. To overcome these limitations, we propose an accurate and robust prediction method, OligoFormer, for siRNA design. OligoFormer comprises three different modules including thermodynamic calculation, RNA-FM module, and Oligo encoder. Oligo encoder is the core module based on the transformer encoder. Taking siRNA and mRNA sequences as input, OligoFormer can obtain thermodynamic parameters, RNA-FM embedding, and Oligo embedding through these three modules, respectively. We carefully benchmarked OligoFormer against six comparable methods on siRNA efficacy datasets. OligoFormer outperforms all the other methods, with an average improvement of 9% in AUC, 6.6% in PRC, 9.8% in F1 score, and 5.1% in PCC compared to the best method among them in our inter-dataset validation. We also provide a comprehensive pipeline with prediction of siRNA efficacy and off-target effects using PITA score and TargetScan score. The ablation study shows RNA-FM module and thermodynamic parameters improved the performance and accelerated convergence of OligoFormer. The saliency maps by gradient backpropagation and base preference maps show certain base preferences in initial and terminal region of siRNAs. The source code of OligoFormer is freely available on GitHub at: https://github.com/lulab/OligoFormer. Docker image of OligoFormer is freely available on the docker hub at https://hub.docker.com/r/yilanbai/oligoformer.

Improving image-based weld defects detection under interference information via a novel semantic supervision strategy

ABSTRACT Weld defects of skin-skeleton structures are invisible and image-based visual inspections using deep learning neural networks are in demand. The main limitation of previous detection algorithms is interference information in images. The reflective light and uneven brightness distribution hinder algorithms to achieve higher detection reliability. In this work, a novel supervision strategy for defects detection algorithms was proposed to break the limitation. A new semantic gate convolution block was developed to help the neural networks to distinguish between targets and interference. The block utilised semantic segmentation labels and a gate function to control the information transmitted to the detection output. Additionally, a new category-pixel-accuracy loss function was adopted to improve the effectiveness of semantic features. The new function reduced the effects of negative pixels to avoid over emphasis on semantic completeness, thus the ineffective features of interference could be eliminated. The results indicated that the back-propagated gradients from the supervision part taught the backbone networks to focus on correct objects instead of interference. To testify the validity of the current method, the algorithms was verified in burns and collapse defects. The detection accuracy under the new semantic supervision reached 99.1%, which was superior to common CNN.

Spatial exchanging fusion network for RGB-T crowd counting

RGB-T crowd counting (RGB-T CC) aims to estimate the crowd population size utilizing the complementary information from visible and thermal images. Current deep models for RGB-T CC typically adopt a three-tier architecture, featuring a middle fusion layer that aggregates both RGB and thermal streams. However, we find that this dedicated fusion layer dominates the training process, causing under-optimization of both modal branches, which becomes the performance bottleneck in mainstream multi-modal counting models. To address this challenge, we propose a simple-yet-effective counting architecture, the Spatial Exchanging Fusion Network (SEFNet). It is built on a Dual Attention Guided Spatial Exchanging (DASE) mechanism, enabling direct extraction and exchange of modality-complementary features between modalities without the extra fusion branch employed in most existing works. This design ensures a more balanced gradient back-propagation over networks, attaining optimized representations in multi-modality fusion over prior models. Besides, the Modality Gradient Enhancement Module (MGEM) in SEFNet can effectively learn modality-specific crowd representations with two counting sub-tasks, dynamically achieving better gradient distribution and further enhancing optimization in both modalities. Extensive experiments demonstrate that SEFNet significantly outperforms state-of-the-art methods on mainstream benchmark datasets, and also exhibits promising generalization ability across various counting backbones and losses.

Linear diffusion noise boosted deep image prior for unsupervised sparse-view CT reconstruction

Objective. Deep learning has markedly enhanced the performance of sparse-view computed tomography reconstruction. However, the dependence of these methods on supervised training using high-quality paired datasets, and the necessity for retraining under varied physical acquisition conditions, constrain their generalizability across new imaging contexts and settings. Approach. To overcome these limitations, we propose an unsupervised approach grounded in the deep image prior framework. Our approach advances beyond the conventional single noise level input by incorporating multi-level linear diffusion noise, significantly mitigating the risk of overfitting. Furthermore, we embed non-local self-similarity as a deep implicit prior within a self-attention network structure, improving the model’s capability to identify and utilize repetitive patterns throughout the image. Additionally, leveraging imaging physics, gradient backpropagation is performed between the image domain and projection data space to optimize network weights. Main Results. Evaluations with both simulated and clinical cases demonstrate our method’s effective zero-shot adaptability across various projection views, highlighting its robustness and flexibility. Additionally, our approach effectively eliminates noise and streak artifacts while significantly restoring intricate image details. Significance. Our method aims to overcome the limitations in current supervised deep learning-based sparse-view CT reconstruction, offering improved generalizability and adaptability without the need for extensive paired training data.

Enhanced vision-transformer integrating with semi-supervised transfer learning for state of health and remaining useful life estimation of lithium-ion batteries

The state of health (SOH) and remaining useful life (RUL) of lithium-ion batteries are crucial for health management and diagnosis. However, most data-driven estimation methods heavily rely on scarce labeled data, while traditional transfer learning faces challenges in handling domain shifts across various battery types. This paper proposes an enhanced vision-transformer integrating with semi-supervised transfer learning for SOH and RUL estimation of lithium-ion batteries. A depth-wise separable convolutional vision-transformer is developed to extract local aging details with depth-wise convolutions and establishes global dependencies between aging information using multi-head attention. Maximum mean discrepancy is employed to initially reduce the distribution difference between the source and target domains, providing a superior starting point for fine-tuning the target domain model. Subsequently, the abundant aging data of the same type as the target battery are labeled through semi-supervised learning, compensating for the source model's limitations in capturing target battery aging characteristics. Consistency regularization incorporates the cross-entropy between predictions with and without adversarial perturbations into the gradient backpropagation of the overall model. In particular, across the experimental groups 13–15 for different types of batteries, the root mean square error of SOH estimation was less than 0.66 %, and the mean relative error of RUL estimation was 3.86 %. Leveraging extensive unlabeled aging data, the proposed method could achieve accurate estimation of SOH and RUL.

Physics-driven deep-learning for marine CSEM data inversion

Marine controlled-source electromagnetic (MCSEM) inversion plays a crucial role in hydrocarbon exploration and pre-drill reservoir evaluation. Deep learning techniques have been widely used in geophysical inversions. Although they work on theoretical data well, their performance on survey data needs to be improved. Since no constraint of physical laws is applied in the training phase, the trained neural network often exhibits large errors when extended to new datasets with different distributions from the train set. To solve this problem, we add a differentiable marine EM forward operator at the end of the neural network that maps the network-predicted results back to the response data. We incorporate a data error term to the loss function and the gradient of data error with respect to model parameters in the gradient back-propagation process so that we can successfully introduce the physical law constraints into the network training process. Experiments on synthetic data validate the effectiveness of our Physics-driven Deep Neural Network (PhyDNN) inversions. It performs significantly better than the conventional DNN as it can recover the model accurately while maintaining data fitting. Tests on theoretical data with different noise levels further demonstrate the superiority of our PhyDNN, which can achieve stable inversions under high noise levels. Moreover, we use the t-distributed stochastic neighbor embedding (t-SNE) algorithm to analyze the similarity between the train sets and real data. The results show that the real data falls within the data distribution of the train sets, ensuring the credibility of the inversion results. Finally, we use PhyDNN to invert an EM survey dataset acquired over a deep-sea sedimentary basin. The inversion results match well Occam's inversions, indicating that our physics-driven network has enhanced the data adaptability and overcome the limitation of conventional DNN in handling new data.

New strategy to optimize in-situ fenton oxidation for TPH contaminated soil remediation via artificial neural network approach

In-situ remediation of total petroleum hydrocarbon (TPH) contaminated soils via Fenton oxidation is a promising approach. However, determining the proper injection amount of H2O2 and Fe source over the Fenton reaction in the complex geological conditions for in-situ TPH soil remediation remains a daunting challenge. Herein, we introduced a practical and novel approach using soft computational models, a multilayer perception artificial neural network (MPLNN), for predicting the TPH removal performance. In this study, we conducted 48 sets of TPH removal experiments using Fenton oxidation to determine the TPH removal performance of a wide range of different ground conditions and generated 336 data points. As a result, a negative Pearson correlation coefficient was obtained in the Fe injection mass and the natural presence of Fe mineral in the soil, indicating that the excess of Fe could significantly retarded the TPH removal performance in the Fenton reaction. In addition, the MPLNN model with 6-6-1 training using Scaled conjugate gradient backpropagation (SCG) with tangent sigmoid as the transfer function demonstrated a high accuracy for TPH removal prediction with the correlation determination of 0.974 and mean square error value of 0.0259. The optimized MPLNN model achieved less than 20% error for predicting TPH removal performance in actual TPH-contaminated soil via Fenton oxidation. Hence, the proposed MPLNN can be useful in improving the Fenton oxidation of TPH removal performance in-situ soil remediation.

Integrating Machine Learning for Improved Prediction of Temperature and Moisture in Pavement Granular Layers

Abstract Pavement temperature and moisture content within the base and subgrade layers affect the load-bearing capacity of the pavement and dominate the pavement performance in cold regions. Accurately predicting pavement temperature and moisture content can improve pavement design and management. Conventional approaches, including numerical and statistical models, have been implemented to predict pavement temperature and soil moisture content. However, they have weaknesses, such as being only suitable for warm regions or only for predicting pavement temperature within the asphalt layer. Furthermore, none of them can simultaneously predict the pavement temperature and moisture content. To address this issue, data collected from an instrumented test road in Alberta, Canada, were used to train a model to predict the daily average pavement temperature and moisture content at various depths through three parameters, namely depth, day of the year, and air temperature. The MATLAB toolbox, Neural Net Fitting, was used, and the performance of three built-in algorithms, Levenberg–Marquardt, Bayesian regularization, and Scaled conjugate gradient backpropagation, was compared. The model with Bayesian regularization showed the highest accuracy, with an R2 value of 0.99 and a root mean square error (RMSE) of 1.49°C for pavement temperature prediction, and an R2 value of 0.95 and an RMSE of 0.025 m3/m3 for moisture content prediction. The model developed in this research is the first to simultaneously estimate pavement temperature and moisture content, so its performance was separately compared with two existing models in the literature. The artificial neural network (ANN) model shows higher accuracy than the two existing models, so it was found that the ANN could be a robust method for pavement temperature and moisture content prediction at various depths.

Vessel-targeted compensation of deformable motion in interventional cone-beam CT

The present standard of care for unresectable liver cancer is transarterial chemoembolization (TACE), which involves using chemotherapeutic particles to selectively embolize the arteries supplying hepatic tumors. Accurate volumetric identification of intricate fine vascularity is crucial for selective embolization. Three-dimensional imaging, particularly cone-beam CT (CBCT), aids in visualization and targeting of small vessels in such highly variable anatomy, but long image acquisition time results in intra-scan patient motion, which distorts vascular structures and tissue boundaries. To improve clarity of vascular anatomy and intra-procedural utility, this work proposes a targeted motion estimation and compensation framework that removes the need for any prior information or external tracking and for user interaction. Motion estimation is performed in two stages: (i) a target identification stage that segments arteries and catheters in the projection domain using a multi-view convolutional neural network to construct a coarse 3D vascular mask; and (ii) a targeted motion estimation stage that iteratively solves for the time-varying motion field via optimization of a vessel-enhancing objective function computed over the target vascular mask. The vessel-enhancing objective is derived through eigenvalues of the local image Hessian to emphasize bright tubular structures. Motion compensation is achieved via spatial transformer operators that apply time-dependent deformations to partial angle reconstructions, allowing efficient minimization via gradient backpropagation. The framework was trained and evaluated in anatomically realistic simulated motion-corrupted CBCTs mimicking TACE of hepatic tumors, at intermediate (3.0 mm) and large (6.0 mm) motion magnitudes. Motion compensation substantially improved median vascular DICE score (from 0.30 to 0.59 for large motion), image SSIM (from 0.77 to 0.93 for large motion), and vessel sharpness (0.189 mm−1 to 0.233 mm−1 for large motion) in simulated cases. Motion compensation also demonstrated increased vessel sharpness (0.188 mm−1 before to 0.205 mm−1 after) and reconstructed vessel length (median increased from 37.37 to 41.00 mm) on a clinical interventional CBCT. The proposed anatomy-aware motion compensation framework presented a promising approach for improving the utility of CBCT for intra-procedural vascular imaging, facilitating selective embolization procedures.

A dual-domain cone beam computed tomography reconstruction framework with improved differentiable domain transform for cone-angle artifact correction

We propose a dual-domain cone beam computed tomography (CBCT) reconstruction framework DualCBR-Net based on improved differentiable domain transform for cone-angle artifact correction. The proposed CBCT dual-domain reconstruction framework DualCBR-Net consists of 3 individual modules: projection preprocessing, differentiable domain transform, and image post-processing. The projection preprocessing module first extends the original projection data in the row direction to ensure full coverage of the scanned object by X-ray. The differentiable domain transform introduces the FDK reconstruction and forward projection operators to complete the forward and gradient backpropagation processes, where the geometric parameters correspond to the extended data dimension to provide crucial prior information in the forward pass of the network and ensure the accuracy in the gradient backpropagation, thus enabling precise learning of cone-beam region data. The image post-processing module further fine-tunes the domain-transformed image to remove residual artifacts and noises. The results of validation experiments conducted on Mayo's public chest dataset showed that the proposed DualCBR-Net framework was superior to other comparison methods in terms of artifact removal and structural detail preservation. Compared with the latest methods, the DualCBR-Net framework improved the PSNR and SSIM by 0.6479 and 0.0074, respectively. The proposed DualCBR-Net framework for cone-angle artifact correction allows effective joint training of the CBCT dual-domain network and is especially effective for large cone-angle region.

Financial constraints prediction to lead socio-economic development: An application of neural networks to the Italian market

This study applies a neural network framework to optimize the classification of firms and to predict their difficulties in collecting external financial resources in the short term. In detail, we adopt replicated bootstrapped algorithms optimized on sensitivity and specificity as error measures and we propose a comparative analysis to identify the best-performing one. According to our results, the Conjugate gradient backpropagation with Fletcher-Reeves updates (i.e., CGF) is the best-performing algorithm, with sensitivity equal to 74.41 % and specificity equal to 70.11 %. Then, we use this algorithm and its weights to provide a classification of the Italian manufacturing industry in 2019, identifying the geographical areas in which firms under financial constraints are located, as well as the most critical industrial sectors. Based on this evidence, and considering the implementation of a cohesion policy, we highlight interventions by policy makers to support firms’ access to the capital market, fostering their investments and the consequent socio-economic development.

Reduced Geostatistical Approach With a Fourier Neural Operator Surrogate Model for Inverse Modeling of Hydraulic Tomography

AbstractIn this study, we propose a new approach for inverse modeling of hydraulic tomography (HT) using the Fourier neural operator (FNO) as a surrogate forward model. FNO is a deep learning model that directly parameterizes the integral kernel in Fourier space to learn solution operators for parametric partial differential equations (PDEs). We trained a highly accurate FNO to learn the solution operators of PDEs in HT, which can efficiently generate multiple channels of hydraulic head fields corresponding to specific pumping tests in new parameter fields through a single forward pass. The trained FNO surrogate model is integrated with the reduced geostatistical approach (RGA) to provide a powerful tool, named RGA‐FNO, for inverse modeling of Gaussian random fields. RGA utilizes principal components to effectively reduce problem dimensions and encode geostatistical information. FNO benefits from deep learning automatic differentiation, enabling efficient backpropagation of gradients during inverse modeling. Our results show that RGA‐FNO can efficiently produce satisfactory estimations for random transmissivity fields with an exponential covariance function, which is non‐differentiable and visually non‐smooth. The RGA‐FNO application on steady‐state multi‐well HT is better than the single‐well transient pumping test for inverse modeling. The data required for inverse modeling applications of RGA‐FNO increases with the variance of the random transmissivity fields. FNO and other neural operator models have the potential to play an important role in groundwater modeling, especially in inverse modeling and experimental design, which often requires a large number of forward simulations.

Retinal disease diagnosis with unsupervised Grad-CAM guided contrastive learning

To alleviate the reliance on expensive annotations, contrastive learning techniques have been applied to diagnose diseases from various types of medical images. However, popular contrastive learning methods, which generate positive pairs by random cropping, face challenges in diagnosing retinal diseases, due to the retinal disease’s characteristic that retinal lesions are tiny and randomly distributed in abnormal fundus images. These lesions may be missed after random cropping, resulting in semantically inconsistent positive pairs that hinder the effectiveness of contrastive learning for retinal disease diagnosis. To address this issue, we propose a novel unsupervised gradient-weighted class activation mapping (Grad-CAM) strategy to roughly locate lesions, thereby suppressing or even eliminating semantically inconsistent positive pairs. Specifically, we develop a gradient-weighted Class Activation Map guided Contrastive Learning (CAMCL) method with two branches for the Grad-CAM based instance discrimination task and the k-nearest neighbors (KNN)-based cluster-wise discrimination task, respectively. By minimizing the KNN loss, the cluster-wise discrimination branch learns high-level representations containing class semantic information. This is then followed by gradient back-propagation to generate Grad-CAM heatmaps from unlabeled data. The generated heatmaps can highlight class-discriminative regions in abnormal fundus images (e.g., retinal lesions) to identify semantically consistent positive pairs while suppressing inconsistent ones. The semantically consistent positive pairs are then input to the instance discrimination task for contrastive learning. In this manner, the semantic inconsistency problem is relieved, and the improved contrastive learning pipeline can be effectively used for retinal disease diagnosis. Experimental results on five retinal disease classification datasets show that our model surpasses other contrastive learning methods, indicating a promising approach for clinical application.

Deep Learning Based Over-the-Air Training of Wireless Communication Systems without Feedback.

In trainable wireless communications systems, the use of deep learning for over-the-air training aims to address the discontinuity in backpropagation learning caused by the channel environment. The primary methods supporting this learning procedure either directly approximate the backpropagation gradients using techniques derived from reinforcement learning, or explicitly model the channel environment by training a generative channel model. In both cases, over-the-air training of transmitter and receiver requires a feedback channel to sound the channel environment and obtain measurements of the learning objective. The use of continuous feedback not only demands extra system resources but also makes the training process more susceptible to adversarial attacks. Conversely, opting for a feedback-free approach to train the models over the forward link, exclusively on the receiver side, could pose challenges to reliably end the training process without intermittent testing over the actual channel environment. In this article, we propose a novel method for the over-the-air training of wireless communication systems that does not require a feedback channel to train the transmitter and receiver. Random samples are transmitted through the channel environment to train a mixture density network to approximate the channel distribution on the receiver side of the network. The transmitter and receiver models are trained with the resulting channel model, and the transmitter can be deployed after training. We show that the block error rate measurements obtained with the simulated channel are suitable for monitoring as a stopping criterion during the training process. The resulting method is demonstrated to have equivalent performance to the end-to-end autoencoder training on small message sequences.

A novel consolidation analysis framework: universal function approximators regularized by physical principles

Analytical and numerical techniques are widely used to analyze and interpret soil consolidation problems. An important limitation is the requirement for significant geotechnical knowledge and expertise to derive “true” solutions. This study proposes an alternative: a universal function approximator regularized with known physical principles. The proposed approach here advances previous work to solve one-dimensional consolidation considering both self-weight and large strains, as well as two-dimensional consolidation by vertical drains. Initial and boundary conditions of the studied consolidation problems are first strongly and weakly enforced for comparison. A neural network is adopted here as the function approximator. To boost prediction accuracy, a novel strategy is proposed to adaptively sample data points for training. An estimate of epistemic uncertainty is achieved using the confidence interval of ensembled multiple outputs. The results show that the proposed approach accurately predicts the behaviors of complex consolidation processes. Results also indicate that regularization using weak physical constraints can alleviate the imbalance of back-propagated gradients of different loss terms and, in turn, achieve higher accuracy. The proposed method is generic, mesh-free, more robust, and can be applied to a wide range of geotechnical problems.

Variational Adversarial Defense: A Bayes Perspective for Adversarial Training.

Various methods have been proposed to defend against adversarial attacks. However, there is a lack of enough theoretical guarantee of the performance, thus leading to two problems: First, deficiency of necessary adversarial training samples might attenuate the normal gradient's back-propagation, which leads to overfitting and gradient masking potentially. Second, point-wise adversarial sampling offers an insufficient support region for adversarial data and thus cannot form a robust decision-boundary. To solve these issues, we provide a theoretical analysis to reveal the relationship between robust accuracy and the complexity of the training set in adversarial training. As a result, we propose a novel training scheme called Variational Adversarial Defense. Based on the distribution of adversarial samples, this novel construction upgrades the defend scheme from local point-wise to distribution-wise, yielding an enlarged support region for safeguarding robust training, thus possessing a higher promising to defense attacks. The proposed method features the following advantages: 1) Instead of seeking adversarial examples point-by-point (in a sequential way), we draw diverse adversarial examples from the inferred distribution; and 2) Augmenting the training set by a larger support region consolidates the smoothness of the decision boundary. Finally, the proposed method is analyzed via the Taylor expansion technique, which casts our solution with natural interpretability.

Learning to Generate Temporal Origin-destination Flow Based-on Urban Regional Features and Traffic Information

Origin-destination (OD) flow contains population mobility information between every two regions in the city, which is of great value in urban planning and transportation management. Nevertheless, the collection of OD flow data is extremely difficult due to the hindrance of privacy issues and collection costs. Significant efforts have been made to generate OD flow based on urban regional features, e.g., demographics, land use, and so on, since spatial heterogeneity of urban function is the primary cause that drives people to move from one place to another. On the other hand, people travel through various routes between OD, which will have effects on urban traffic, e.g., road travel speed and time. These effects of OD flows reveal the fine-grained spatiotemporal patterns of population mobility. Few works have explored the effectiveness of incorporating urban traffic information into OD generation. To bridge this gap, we propose to generate real-world daily temporal OD flows enhanced by urban traffic information in this paper. Our model consists of two modules: Urban2OD and OD2Traffic . In the Urban2OD module, we devise a spatiotemporal graph neural network to model the complex dependencies between daily temporal OD flows and regional features. In the OD2Traffic module, we introduce an attention-based neural network to predict urban traffic based on OD flow from the Urban2OD module. Then, by utilizing gradient backpropagation, these two modules are able to enhance each other to generate high-quality OD flow data. Extensive experiments conducted on real-world datasets demonstrate the superiority of our proposed model over the state of the art.

Towards transferable adversarial attacks on vision transformers for image classification

The deployment of high-performance Vision Transformer (ViT) models has garnered attention from both industry and academia. However, their vulnerability to adversarial examples highlights security risks for scenarios such as intelligent surveillance, autonomous driving, and fintech regulation. As a black-box attack technique, transfer attacks leverage a surrogate model to generate transferable adversarial examples to attack a target victim model, which mainly focuses on a forward (input diversification) and a backward (gradient modification) approach. However, both approaches are currently implemented straightforwardly and limit the transferability of surrogate models. In this paper, we propose a Forward-Backward Transferable Adversarial Attack framework (FBTA) that can generate highly transferable adversarial examples against different models by fully leveraging ViT’s distinctive intermediate layer structures. In the forward inference process of FBTA, we propose a Dropout-based Transferable Attack (DTA) approach to diversify the intermediate states of ViT models, simulating an ensemble learning effect; in the backward process, a Backpropagation Gradient Clipping (BGC) method is designed to refine the gradients within intermediate layers of ViT models intricately. Extensive experiments on state-of-the-art ViTs and robust CNNs demonstrate that our FBTA framework achieves an average performance improvement of 2.79% compared to state-of-the-art transfer-based attacks, offering insights for the comprehension and defense against transfer attacks.