The proliferation of multimodal misinformation poses growing threats to public discourse and societal trust. While Large Vision-Language Models (LVLMs) have enabled recent progress in multimodal misinformation detection (MMD), the rise of generative AI (GenAI) tools introduces a new challenge: GenAI-driven news diversity, characterized by highly varied and complex content. We show that this diversity induces multi-level drift, comprising (1) model-level misperception drift, where stylistic variations disrupt a model’s internal reasoning, and (2) evidence-level drift, where expression diversity degrades the quality or relevance of retrieved external evidence. These drifts significantly degrade the robustness of current LVLM-based MMD systems. To systematically study this problem, we introduce DriftBench, a large-scale benchmark comprising 16,000 news instances across six categories of diversification. We design three evaluation tasks: (1) robustness of truth verification under multi-level drift; (2) susceptibility to adversarial evidence contamination generated by GenAI; and (3) analysis of reasoning consistency across diverse inputs. Experiments with six state-of-the-art LVLM-based detectors show substantial performance drops (average F1 ↓ 14.8%) and increasingly unstable reasoning traces, with even more severe failures under adversarial evidence injection. Our findings uncover fundamental vulnerabilities in existing MMD systems and suggest an urgent need for more resilient approaches in the GenAI era.

Tingchao Fu, Wenkai Wang, Fanxiao Li, Huadong Zhang, Jinhong Zhang, Dayang Li, Yunyun Dong, Renyang Liu, Wei Zhou. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Predicting drug-target binding affinity (DTA) is a key task for compound screening and drug discovery. Despite notable advances achieved by deep learning approaches, two critical challenges remain in DTA prediction: (i) appropriately and accurately representing the fine-grained features of drugs and targets; (ii) mitigating model's reliance on overly simple drug patterns while fully exploring the multi-level feature interactions between drugs and target. To address these challenges, we propose DILPA-DTA, a multimodal neural framework that employs spatial and hierarchical attention mechanisms to integrate the structural and sequence representations of drug molecules and target proteins. Specifically, we represent drugs as graphs and model proteins at multiple scales using sequence and structural information. Experiments on two benchmark datasets, together with case studies, show that DILPA-DTA outperforms state-of-the-art baselines. Furthermore, our model provides strong interpretability by accurately localizing key interacting atoms and residues involved in binding. The source code and datasets of DILPA-DTA are available at https://github.com/DILPA-DTA/DILPA-DTA.

Personalized cancer drug treatment is emerging as a frontier issue in modern medical research. Considering the genomic differences among cancer patients, determining the most effective drug treatment plan is a complex and crucial task. In response to these challenges, this study introduces the Adaptive Sparse Graph Contrastive Learning Network (ASGCL), an innovative approach to unraveling latent interactions in the complex context of cancer cell lines and drugs. The core of ASGCL is the GraphMorpher module, an innovative component that enhances the input graph structure via strategic node attribute masking and topological pruning. By contrasting the augmented graph with the original input, the model delineates distinct positive and negative sample sets at both node and graph levels. This dual-level contrastive approach significantly amplifies the model's discriminatory prowess in identifying nuanced drug responses. Leveraging a synergistic combination of supervised and contrastive loss, ASGCL accomplishes end-to-end learning of feature representations, substantially outperforming existing methodologies. Comprehensive ablation studies underscore the efficacy of each component, corroborating the model's robustness. Experimental evaluations further illuminate ASGCL's proficiency in predicting drug responses, offering a potent tool for guiding clinical decision-making in cancer therapy.

The traditional three-stage pruning pipeline is first to train an original dense network, then identify redundant parts of the network for pruning based on the evaluation metrics of the pruning algorithm, and finally fine-tuning the pruned model, which is a time-consuming and computationally expensive process. Traditional pruning algorithms are greedy and aggressive, which may cause many important network connections to be pruned incorrectly, resulting in significant performance degradation. In this paper, we propose an incremental soft pruning during training method with the following characteristics: 1) Given the pruning rate of the network, a trained sub-network, which has performance comparable to the original network, can be obtained after training. 2) We propose three incremental pruning rate growth functions and allow the network structure to be dynamically adjusted during training to avoid pruning important network connections. 3) During the network training process, we also introduce gradient restriction, which only updates important network connections to discover good sub-network structures better. Extensive experiments show that our method can achieve better results than previous methods in different datasets and network models.

Drug combination therapy is generally more effective than monotherapy in the field of cancer treatment. However, screening for effective synergistic combinations from a wide range of drug combinations is particularly important given the increase in the number of available drug classes and potential drug-drug interactions. Existing methods for predicting the synergistic effects of drug combinations primarily focus on extracting structural features of drug molecules and cell lines, but neglect the interaction mechanisms between cell lines and drug combinations. Consequently, there is a deficiency in comprehensive understanding of the synergistic effects of drug combinations. To address this issue, we propose a drug combination synergy prediction model based on multi-source feature interaction learning, named MFSynDCP, aiming to predict the synergistic effects of anti-tumor drug combinations. This model includes a graph aggregation module with an adaptive attention mechanism for learning drug interactions and a multi-source feature interaction learning controller for managing information transfer between different data sources, accommodating both drug and cell line features. Comparative studies with benchmark datasets demonstrate MFSynDCP's superiority over existing methods. Additionally, its adaptive attention mechanism graph aggregation module identifies drug chemical substructures crucial to the synergy mechanism. Overall, MFSynDCP is a robust tool for predicting synergistic drug combinations. The source code is available from GitHub at https://github.com/kkioplkg/MFSynDCP .

Abstract Image hiding is a task that hides secret images into cover images. The purposes of image hiding are to ensure the secret images are invisible to the human and the secret images can be recovered. The current state‐of‐the‐art steganography methods run the risk of secret information leakage. A safe image hiding network (SIHNet) is presented to reduce the leakage of secret information. Based on some phenomena of image hiding methods which use invertible neural network, a reversible secret image processing (SIP) module is proposed to make the secret images suitable for hiding and make the stego images leak less secret information. Besides, a reversible lost information hiding (LIH) module is used to hide the lost information into the cover images, thus the method can recover the secret images better than the method that uses random noise to replace the lost information. Experimental results show that SIHNet outperforms other state‐of‐the‐art methods on the PSNR and SSIM values of the recovered secret images and the stego images. Besides, residual images of other state‐of‐the‐art methods all contain information about secret images while residual images of SIHNet leak almost no secret information. Thus the method can prevent the listener of transmission channel from obtaining the information of the secret image through the residual image, which means SIHNet performs better in security than other state‐of‐the‐art methods.

Abstract Image steganography aims to hide secret data in the cover media for covert communication. Though many deep‐learning‐based image steganography methods have been presented, these approaches suffer from the inefficiency of building long‐distance connections between the cover and secret images, leading to noticeable modification traces and poor steganalysis resistance. To improve the visual imperceptibility of generated stego images, it is essential to establish a global correlation between the cover and secret images. In this way, the secret image can be dispersed throughout the cover image globally. To bridge this gap, a novel image steganography framework called HiiT is proposed, which takes advantage of CNN and Transformer to learn both the local and global pixel correlation in image hiding. Specifically, a new Transformer structure called Inception Transformer is proposed, which incorporates the Inception Net in the attention‐based Transformer architecture. The Inception Net can learn different scaled image features using multiple convolution kernels, while the attention mechanism can learn the global pixel correlation. By this, the proposed Inception Transformer learns the long‐distance pixel dependency between the cover and secret images. Furthermore, we propose a ‘Skip Connection’ mechanism in the proposed Inception Transformer, which merges the low‐level visual features and high‐level semantic features and achieves better model performance. In detail, The HiiT generates higher‐quality stego images with 45.46 PSNR and 0.9915 SSIM. Besides, accurately restored secret images achieve 47.27 PSNR and 0.9952 SSIM. Extensive experimental results show the proposed HiiT significantly improves the image‐hiding performance compared with state‐of‐the‐art methods.

Abstract Image hiding is a task that embeds secret images in digital images without being detected. The performance of image hiding has been greatly improved by using the invertible neural network. However, current image hiding methods are less robust in the face of Joint Photographic Experts Group (JPEG) compression. The secret image cannot be extracted from the stego image after JPEG compression of the stego image. Some methods show good robustness for some certain JPEG compression quality factors but poor robustness for other common JPEG compression quality factors. An image‐hiding network (RIHINNet) that is robust to all common JPEG compression quality factors is proposed. First of all, the loss function is redesigned; thus, the secret image is hidden as much as possible in the area that is less likely to be changed after JPEG compression. Second, the classifier is designed, which can help the model to select the extractor according to the range of JPEG compression degree. Finally, the interval robustness of the secret image extraction is improved through the design of a denoising module. Experimental results show that this RIHINNet outperforms other state‐of‐the‐art image‐hiding methods in the face of JPEG compressed noise with random compression quality factors, with more than 10 dB peak signal‐to‐noise ratio improvement in secret image recovery on ImageNet, COCO and DIV2K datasets.

Abstract Image steganography is the art of concealing secret information within images to prevent detection. In deep‐learning‐based image steganography, a common practice is to fuse the secret image with the cover image to directly generate the stego image. However, not all features are equally critical for data hiding, and some insignificant ones may lead to the appearance of residual artifacts in the stego image. In this article, a novel network architecture for image steganography with hybrid attention mechanism based on generative adversarial network is introduced. This model consists of three subnetworks: a generator for generate stego images, an extractor for extracting the secret images, and a discriminator to simulate the detection process, which aids the generator in producing more realistic stego images. A specific hybrid attention mechanism (HAM) module is designed that effectively fuses information across channel and spatial domains, facilitating adaptive feature refinement within deep image representations. The experimental results suggest that the HAM module not only enhances the image quality during both the steganography and extraction processes but also improves the model's undetectability. Stego images are mixed with varying levels of noise in the training process, which can further improve robustness. Finally, it is verified that the model outperforms current steganography approaches on three datasets and exhibits good undetectability.

Single Domain Generalization (SDG) is a more realistic setting than Domain Generalization (DG) and Domain Adaptation (DA). It aims to train a domain-agnostic model in the presence of a single source domain to perform well on arbitrary unseen target domains. To facilitate the practical deployment of remote sensing image segmentation in the real world, we propose a novel SDG method, termed Category Consistency on Domain Randomization (CCDR). To expand the coverage of a single source domain, CCDR implements a simple yet effective data generation module to perform domain randomization of texture and style information, since texture divergences from different geographical environments or phenological periods, and style divergences from different illumination or weather conditions, are the main causes of the domain shift in remote sensing. And in addition to emphasizing inter- and intra-class relationships within each domain via multi-domain supervised learning, CCDR further draws inspiration from the triple loss to enhance the semantic correlation across the source domain and the generated auxiliary domain, which makes the segmentation model more sensitive to class discriminative information and better adapted to unseen target domains. In comparison to other state-of-the-art SDG methods, CCDR can synthesize the more effective auxiliary domain, and perform more reliable classification on the unseen target domains without the sophisticated training pipeline and cumbersome data generation process. And the experimental results on two public remote sensing datasets demonstrate that CCDR has remarkable advantages in remote sensing image segmentation tasks. And the code is publicly available at: https://github.com/LCB1970/CCDR.

There is a large amount of out-of-distribution data (OOD) in remote sensing, which hinders high-accuracy segmentation models under the assumption of independent identical distribution (i.i.d.) from stable and reliable performance in real-world remote sensing applications. And Domain Adaptation (DA) is presented to seamlessly extend classifiers to the label-scarce target domain in the presence of the label-sufficient source domain with different data distributions. However, given that the domain shift, i.e. the distribution difference between the two domains, is more serious in remote sensing images, the current DA methods for image segmentation in Computer Vision (CV) typically perform unsatisfactorily in remote sensing, even suffering from the negative domain alignment. To this end, this paper proposes the Self-Training Adversarial Domain Adaptation (STADA) method for remote sensing image segmentation, which not only performs adversarial learning to extract domain-invariant features, but also implements Self-Training using pseudo-labels in the target domain denoised by the conditional adversarial loss for classifier adaptation. The ISPRS and WHU datasets are employed to conduct extensive experiments to investigate the effectiveness of STADA and the specific effect of its each DA component. And the experimental results demonstrate that STADA outperforms other state-of-the-art DA methods in the remote sensing image segmentation task.

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

Due to more abundant data sources, more various objects of interest, and more time-consuming annotations, there is a large amount of out-of-distribution (OOD) data in the remote sensing field, on which the performance of high-accuracy image segmentation models trained under ideal experimental conditions generally degrades dramatically. Domain adaptation (DA) consequently comes into being, which aims to learn the predictor for the label-scarce target domain of interest with the help of the label-sufficient source domain in the presence of the distribution difference, namely, domain shift, between the two domains. However, the off-the-shelf DA methods for image segmentation not only struggle to cope with the more complex domain shift problems in remote sensing imagery but also almost cannot process heterogeneous data directly without information loss. While the current heterogeneous DA methods mostly still rely on some supervision information from the target domain, which is typically inaccessible in the real world. To overcome these drawbacks, we propose the multilevel heterogeneous unsupervised DA (UDA) method, termed MHDA, which unifies the instance-level DA based on cycle consistency, the feature-level DA based on contrastive learning, and the decision-level DA based on task consistency into a framework to more effectively handle the complex domain shift and heterogeneous data. After that, extensive DA experiments are conducted on the International Society for Photogrammetry and Remote Sensing (ISPRS) dataset, the BigCity dataset constructed by ourselves, and the Wuhan University (WHU) dataset, to explore the effect of each module in MHDA, the necessity of heterogeneous DA, and the effectiveness of multilevel DA. And the results demonstrate that MHDA can achieve superior performance on the remote sensing image segmentation task, compared with several state-of-the-art DA methods.

Most cloud providers offer Deep Learning as a Service (DLaaS) for different business, science and engineering domains. However, it is known that deep neural networks (DNNs) are vulnerable to adversarial examples, which can cause well-trained DNN models to misbehave by injecting human-imperceptible perturbations to the query input data. Securing deep learning as a service becomes a critical challenge in mitigating such adversarial input perturbations, and enhancing the robustness of DNNs. In this article, we report two important facts: First, most adversarial perturbations are high frequency signals or are added to high frequency signals. Second, due to Frequency Principle that neural networks overly pay attention to fit the low frequency signals during training, the models could be easily misled by the high frequency signals of adversarial examples. These facts consequently contribute to the vulnerability of DNNs service in the Cloud. We conjecture that the more robust the neural networks are in learning from high frequency signals, the more resilient these neural networks are against adversarial perturbed examples. We propose a novel method for generating high-frequency-enhanced adversarial examples, which is achieved by a high-pass filter in the frequency domain via Fourier Transform. This method enhances the learning ability for high frequency signals and ameliorates to over-fit useless low frequency signals. In order to improve the robustness of DNNs service under such signal frequency attacks, we propose a multi-modal collaborative adversarial training framework, named as MMCAT, which uses the multi-modal information of the input images for cross-modal collaborative training, delivering excellent extension for effectively learning of multi-modal image information. Extensive experiments show that under strong adaptive frequency attacks, the DNNs service trained with the proposed MMCAT method achieve superior performance and robustness over the state-of-the-art adversarial training approaches.

The robustness and security of deep neural network (DNN) models have received much attention in recent years. In-depth research on adversarial example generation methods that make DNN models make wrong judgments and decisions will facilitate further research on more comprehensive and practical adversarial defense methods. Most existing adversarial example generation methods focus too much on attack performance and design adversarial noise at the pixel level, resulting in the generated adversarial examples with redundant noise and evident perturbations. In this paper, we try to find the well-designed perturbations at the feature-level and propose a novel deep reversible network-based imperceptible adversarial examples generation method called RIA. Experimental results show that RIA can obtain more natural adversarial examples without losing attack performance and reducing redundant noise based on well-designed feature maps. To the best of our knowledge, in the white-box attack method research, this work is the first attempt to directly add perturbations to feature maps and use an reversible network to generate adversarial examples based on the perturbed feature maps.

Multi-focus image fusion can overcome the issues that optical lens imaging cannot focus multiple targets simultaneously due to the depth of field limitation. In this paper, we propose a generative adversarial network (DDF-GAN) which consists of a generator and two discriminators to directly generate fused images without decision maps and post-processing. In training process, the source all-in-focus image and the fused image generated by generator are used as input to one of the dual discriminators. Meanwhile, the gradient map of the fused image and the source gradient map of the source all-in-focus image are used as input to another discriminator. An adversarial relationship is established to enhance texture details of fused image. In addition, we create a data set and use it as the main training set for the proposed model. Abundant experiments were carried out to verify the availability of our method. Experimental results prove that our method has advantages in subjective visual perception of human and quantitative measurement.

Image registration based on phase correlation has drawn extensive attention due to its high accuracy and efficiency. However, due to changes of image content, non-linear gray difference and other noises of image pairs, the line fitting of phase angle points acquired by the Singular Value Decomposition (SVD) and 1-D phase unwrapping is also an intractable problem in the process of phase correlation image registration. In this letter, we propose to a probability-guided Random Sample Consensus(RANSAC), namely utilizing a probability to guide the hypothesis search of RANSAC to fit the line accurately and efficiently. The probability of each phase angle point is predicted by a deep convolution neural network of ProbNet we build and the parameters of network are optimized effectively by integrating probability-guided RANSAC into an end-to-end trainable displacement estimation pipeline. The qualitative experiment is carried out to illustrate the effectiveness of the proposed method. And in the quantitative experiments, two competitive methods of LO-RANSAC and Least Square fitting (LSQ) and the Naive RANSAC method are brought in to compare. The experimental result illustrates that the proposed method has an increase in the success rate of displacement estimation and in efficiency.

BACKGROUND: Lung cancer is one of the most common types of cancer, among which lung adenocarcinoma accounts for the largest proportion. Currently, accurate staging is a prerequisite for effective diagnosis and treatment of lung adenocarcinoma. Previous research has used mainly single-modal data, such as gene expression data, for classification and prediction. Integrating multi-modal genetic data (gene expression RNA-seq, methylation data and copy number variation) from the same patient provides the possibility of using multi-modal genetic data for cancer prediction. A new machine learning method called gcForest has recently been proposed. This method has been proven to be suitable for classification in some fields. However, the model may face challenges when applied to small samples and high-dimensional genetic data. RESULTS: In this paper, we propose a multi-weighted gcForest algorithm (MLW-gcForest) to construct a lung adenocarcinoma staging model using multi-modal genetic data. The new algorithm is based on the standard gcForest algorithm. First, different weights are assigned to different random forests according to the classification performance of these forests in the standard gcForest model. Second, because the feature vectors generated under different scanning granularities have a diverse influence on the final classification result, the feature vectors are given weights according to the proposed sorting optimization algorithm. Then, we train three MLW-gcForest models based on three single-modal datasets (gene expression RNA-seq, methylation data, and copy number variation) and then perform decision fusion to stage lung adenocarcinoma. Experimental results suggest that the MLW-gcForest model is superior to the standard gcForest model in constructing a staging model of lung adenocarcinoma and is better than the traditional classification methods. The accuracy, precision, recall, and AUC reached 0.908, 0.896, 0.882, and 0.96, respectively. CONCLUSIONS: The MLW-gcForest model has great potential in lung adenocarcinoma staging, which is helpful for the diagnosis and personalized treatment of lung adenocarcinoma. The results suggest that the MLW-gcForest algorithm is effective on multi-modal genetic data, which consist of small samples and are high dimensional.

In remote sensing community, accurate image registration is the basement of the subsequent application of remote sensing images. Phase correlation based image registration has drawn extensive attention due to its high accuracy and high efficiency. But the effect of image border corrupted its registration accuracy and success rate. Currently, the main solution is blurring off the border of image by weighting window function with reference and sensed image. However, the approach also inevitably filters out non-border information of an image, which is useful to image registration based on phase correlation. In this paper, another way of thinking to eliminate the effect of image border is proposed, namely decomposing the image into two images, one is periodic image and the other is smooth image. Engulfing the original image by the periodic one has no direct effect on the image border due to applying Fourier Transform. The smooth image is analogous to an error image, which has little information except at the border. The novel algorithm of eliminating the image border can improve the success rate and registration accuracy of phase correlation based image registration. To illustrate its superiority, we showed the corresponding magnitude of Fourier Transform of image visually and compared two measurements with other three state-of-the-art algorithms quantitatively.

Image registration is a critical process for the various applications in the remote sensing community, and its accuracy greatly affects the results of the subsequent applications. Image registration based on phase correlation has been widely concerned due to its robustness to gray differences and efficiency. After calculating the normalized cross-relation matrix Q, the most commonly used approach is fitting the 2-D phase plane that passes through the origin, but it needs to remove contaminated spectrum carefully and the corresponding parameters are empirical. In fact, the phase correlation matrix is rank one for a noise-free translation model. This property simplifies the matching problem to finding the best rank-one approximation of the normalized cross-relation matrix. We develop a novel algorithm that performs the rank-one matrix factorization on the phase correlation matrix by assuming its noise as mixture of Gaussian (MoG) distributions. The MoG model is a general approximator for any continuous distribution, and hence is able to model a wide range of noise distribution. The parameters of the MoG model can be evaluated under the framework of maximum likelihood estimation by using an expectation-maximization method, and the subspace is calculated with standard methods. The advantages of the algorithm, high accuracy, and robustness to aliasing, noise, gray difference, and occlusions are illustrated by a series of simulated and real-image experiments.