Zhenguang Wang, Bo Li, Wenhui Tan, Peng Cao, Yang Wang, Jia Duan, Fei Wang, Osmar Zaiane. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

In recent years, transformer-based methods have achieved remarkable progress in hyperspectral image classification (HSIC). However, they often rely heavily on extensive training samples to achieve optimal performance. Moreover, these methods frequently fail to adequately capture diverse local spectral–spatial correlations and multi-granular features inherent in hyperspectral images (HSIs). Crucially, existing approaches often overlook the pivotal role of the target center pixel. Their attention mechanisms tend to focus on irrelevant background regions, thereby reducing feature discriminability and degrading classification accuracy. To address these challenges, we propose a novel Center-Focused Transformer (CFT) framework that seamlessly integrates multi-scale spectral–spatial fusion for HSIC. Our framework comprises three key components. First, the Spectral–Spatial Fusion (SSF) mechanism integrates local and global dependencies by employing PCA alongside a Superpixel Graph Feature Extraction (SGFE) block. Second, the Multi-Granular Feature Enhancement (MGFE) approach strengthens spectral–spatial interactions through patch augmentation, a HybridConv block, and a Multi-Scale CBAM (MS-CBAM) block. Finally, the Focus Center Transformer (FCT) strategy explicitly emphasizes the importance of the central pixel for precise classification by incorporating Gaussian Positional Embedding (GPE) and cross-layer aggregation. Extensive experiments on four public datasets demonstrate that the proposed CFT consistently outperforms state-of-the-art methods, highlighting its potential for practical engineering applications.

Trigger-Induced Recommendation (TIR), which aims to predict user interest based on a trigger item, has gained considerable traction on e-commerce platforms. Current TIR methods typically analyze user intent by integrating explicit interest in the trigger item and implicit interest derived from user historical behaviors. However, these methods often overlook the contextual information and occurring scenarios related to the trigger, resulting in an undue emphasis on isolated trigger items and a consequently restrictive understanding of users' short-term intentions. To address these challenges, we propose a novel scenario-aware multi-trigger induction method featuring three key enhancements: (1) The Context Modeling Network learns contextual information associated with the trigger during the request, improving the understanding of users' real intentions regarding the trigger item; (2) The Multi-Trigger Learning Network introduces user latent triggers from various scenarios to uncover users' potential external preferences; (3) The Scenario Induction Network captures the characteristics of the scenarios in which triggers occur and performs induction to yield scenario-aware user intentions prediction. We validate our approach through experiments on multiple industrial datasets, demonstrating the model's effectiveness. Furthermore, we have integrated the model into an online advertising system, achieving a 5.46% improvement in Click-Through Rate (CTR).

Accurate Click-Through Rate (CTR) prediction is vital for search engines and recommendation systems, yet it is often hindered by the ''cold start problem'', which arises from insufficient historical data for new users. Recent approaches have sought to tackle this by training encoder-decoder networks on data from warm users to generate virtual behavior embeddings for cold users. However, these methods have shortcomings in terms of simplistic encoding techniques for warm user behaviors and direct utilization of virtual behavior embeddings, leading to limitations in user interest expression and generalization. To address these challenges, we propose a novel method that leverages Graph Isomorphism Networks (GIN) for cohort modeling within CTR prediction. GIN effectively captures high-order user-item interactions, providing a more nuanced understanding of users' diverse interests. Additionally, the cohort modeling strategy minimizes deviations in constructed embeddings, enhancing the model's generalization abilities. We validate our approach through experiments on public and industrial datasets, demonstrating significant improvements for both warm and cold users compared to existing methodologies. Furthermore, we implemented the GIN Cohort Modeling (GINCM) in a large-scale online advertising system, optimizing for both pre-computation and real-time processing to reduce latency. The implementation yields notable enhancements of 2.13% in CTR and Revenue Per Mille(RPM), showcasing the practical effectiveness and real-world applicability of our model.

In response to the strong correlation between the chaotic system state and initial state and parameters in traditional chaotic encryption algorithms, which may lead to periodicity in chaotic sequences, the chaos long short-term memory (Chaos-LSTM) model is constructed by combining chaotic systems with LSTM neural networks. The chaos sequence proliferation (CSP) algorithm is constructed to address the problem that the limited computational accuracy of computers can lead to periodicity in long chaotic sequences, making them unsuitable for encrypting objects with large amounts of data. By combining the Chaos-LSTM model and CSP algorithm, a geographic information encryption system is proposed. First, the Chaos-LSTM model is used to output chaotic sequences with high spectral entropy (SE) complexity. Then, a shorter chaotic sequence is selected and proliferated using the CSP algorithm to generate chaotic proliferation sequences that match the encrypted object; a randomness analysis is conducted and testing is performed on it. Finally, using geographic images as encryption objects, the chaotic proliferation sequence, along with the scrambling and diffusion algorithms, are combined to form the encryption system, which is implemented on the ZYNQ platform. The system’s excellent confidentiality performance and scalability are proved by software testing and hardware experiments, making it suitable for the confidentiality peers of various encryption objects with outstanding application value.

Dynamic anomaly detection of space targets is essential for space situational awareness. With the intrinsic Range Doppler imaging mechanism, the spatio-temporal feature of ISAR sequences has a great potential in dynamic representation. To address this, we propose a novel dynamic anomaly detection method using Spatio-Temporal Graph Convolutional Networks (STGCN) to capture spatio-temporal features from inverse synthetic aperture radar (ISAR) image sequences. By constructing a skeleton model according to the universal geometry of space satellites, our method captures node correlations in both temporal and spatial dimensions from continuous ISAR frames, thereby enabling reliable dynamic target recognition. The effectiveness and superiority of this approach are demonstrated through comparative experiments.

By unfolding traditional optimization algorithms into the form of neural networks, the unfolding network methods have attracted more and more attention in sparse inverse synthetic aperture radar (ISAR) imaging because of their high reconstruction performance and good interpretability. However, existing unfolding network methods mainly focus on 1-D sparse ISAR imaging and cannot be directly applied to 2-D sparse ISAR data. For this reason, a novel learned 2D-two-step iterative shrinkage/thresholding algorithm (L-2D-TwISTA) is proposed for high-efficiency and high-accuracy 2-D sparse ISAR imaging. Specifically, each stage of L-2D-TwISTA corresponds to an iterative solution step of the developed 2D-TwISTA approach. Moreover, a complex-valued (CV) residual network is designed in L-2D-TwISTA to improve training efficiency and solve the nonlinear problem of proximal mapping of the 2D-TwISTA more effectively. The experimental results of real-measured data confirm that the L-2D-TwISTA can realize high-performance 2-D sparse ISAR imaging.

With the widespread application of digital assets and the rapid development of blockchain technology, achieving secure and efficient transactions between different blockchain networks has become an urgent challenge. Existing cross-chain methods impose limitations, e.g., hash lock technology exhibits low scalability, side-chain technology is overly complex, and notary schemes pose centralization risks. In order to address these issues, we propose a cross-chain transaction solution based on proxy re-signature technology. To tackle the centralization concerns, we employ proxy re-signature technology to decentralize the authority of notaries among transaction participants, minimizing the potential for centralization at a low cost. We further present an extended scheme that can guarantee more requirements such as higher transaction amount and shorter transaction time. Furthermore, we analyze and compare the signature technologies used in the proposed solution and provide a systematic proof of our approach’s security.

Optimal hyper-parameter selection for low-rank and sparse matrix decomposition (LRSMD) in synthetic aperture radar jamming suppression is usually challenging. This letter proposes an effective approach to LRSMD jamming suppression by exploiting the polarized similarity. A polarimetric ratio function is established to straightforwardly determine the rank hyper-parameter. The polarimetric ratio function is defined as the energy ratio of the decomposed low-rank jamming component to the sparse target signal, which is consistent across multiple polarized channels due to the equal jamming gain distinguished from the target polarized scattering. The algorithm optimally exploits the polarized similarity of jamming components to determine the rank hyper-parameter. It provides enhanced robustness and accuracy in SAR jamming removal, confirmed by synthetic experiments.

The attributed scattering center (ASC) can provide a concise physical description of radar targets, which has broad applications for physical-induced radar recognition. However, the ASC extraction is of great difficulty due to its high-dimensional parametric modeling. In this letter, a modified Alternative Direction Method of Multipliers (ADMM) unfolded-net is proposed for ASC decomposition of synthetic aperture radar (SAR) targets. By transforming the ASC decomposition challenge into a sparse reconstruction problem, the modified ADMM-net can fast attain ASC parameters as well as reconstruct a complete signal from sparse observations. Instead of tremendous iterations with fixed hyper-parameters as traditional methods, the modified ADMM-net unrolls several iterations into a few layers with learned hyper-parameters. Therefore, the proposed algorithm can realize fast and precise ASC decomposition. Experimental results confirm its effectiveness and superiority.

Abnormal dynamics awareness of space targets is critical for space surveillance. With the intrinsic range-Doppler projection mechanism, an inverse synthetic aperture radar (ISAR) image sequence naturally has crucial potential for abnormal dynamics interpretation on uncooperative targets. In this paper, we develop an automated deep neural network architecture for abnormal dynamic recognition of space targets from ISAR image sequences. With the accommodation of the stacked sparse autoencoder (SSAE) network joint sparsity constraint, the geometrical feature flow of an ISAR image sequence is learned to represent the target dynamics concisely. Then, the abnormal dynamic recognition is turned into a sequential classification task by exploiting the encoded feature flow through the long-short time memory (LSTM) network. Extensive experiment results confirm the superiority of the proposal in both precision and efficiency aspects.

Model protection is vital when deploying Convolutional Neural Networks (CNNs) for commercial services, due to the massive costs of training them. In this work, we propose a selective encryption (SE) algorithm to protect CNN models from unauthorized access, with a unique feature of providing hierarchical services to users. Our algorithm firstly selects important model parameters via the proposed Probabilistic Selection Strategy (PSS). It then encrypts the most important parameters with the designed encryption method called Distribution Preserving Random Mask (DPRM), so as to maximize the performance degradation by encrypting only a very small portion of model parameters. We also design a set of access permissions, using which different amount of most important model parameters can be decrypted. Hence, different levels of model performance can be naturally provided for users. Experimental results demonstrate that the proposed scheme could effectively protect the classification model VGG19 by merely encrypting 8% parameters of convolutional layers. We also implement the proposed model protection scheme in the denoising model DnCNN, showcasing the hierarchical denoising services.

Deep neural networks (DNNs) have been shown to be vulnerable against adversarial examples (AEs), which are maliciously designed to cause dramatic model output errors. In this work, we reveal that normal examples (NEs) are insensitive to the fluctuations occurring at the highly-curved region of the decision boundary, while AEs typically designed over one single domain (mostly spatial domain) exhibit exorbitant sensitivity on such fluctuations. This phenomenon motivates us to design another classifier (called dual classifier) with transformed decision boundary, which can be collaboratively used with the original classifier (called primal classifier) to detect AEs, by virtue of the sensitivity inconsistency. When comparing with the state-of-the-art algorithms based on Local Intrinsic Dimensionality (LID), Mahalanobis Distance (MD), and Feature Squeezing (FS), our proposed Sensitivity Inconsistency Detector (SID) achieves improved AE detection performance and superior generalization capabilities, especially in the challenging cases where the adversarial perturbation levels are small. Intensive experimental results on ResNet and VGG validate the superiority of the proposed SID.

Model protection is vital when deploying Convolutional Neural Networks (CNNs) for commercial services, due to the massive costs of training them. In this work, we propose a selective encryption (SE) algorithm to protect CNN models from unauthorized access, with a unique feature of pro-viding hierarchical services to users. Our algorithm firstly selects important model parameters via the proposed Probabilistic Selection Strategy (PSS). It then encrypts the most important parameters with the designed encryption method called Distribution Preserving Random Mask (DPRM), so as to maximize the performance degradation by encrypting only a very small portion of model parameters. We also design a set of access permissions, using which different amount of most important model parameters can be decrypted. Hence, different levels of model performance can be naturally provided for users. Experimental results demonstrate that the proposed scheme could effectively protect the classification model VGG19 by merely encrypting 8% parameters of convolutional layers. We also implement the proposed model protection scheme in the denoising model DnCNN, showcasing the hierarchical denoising services.

Nowadays, cloud computing platforms are becoming increasingly prevalent and readily available, providing alternative and economic services for resource-constrained clients to perform large-scale computations. This work addresses the problem of secure outsourcing of large-scale nonnegative matrix factorization (NMF) to a cloud in a way that the client can verify the correctness of the results with small overhead. The protection of the input matrix is achieved by a random permutation and scaling encryption mechanism. By exploiting the iterative nature of NMF computation, we propose a single-round verification strategy, which can be proved to be quite effective. Theoretical and experimental results are provided to show the superior performance of the proposed scheme.

Cloud computing platforms are becoming increasingly prevalent and readily available nowadays, providing us alternative and economic services for resource-constrained clients to perform large-scale computation. In this work, we address the problem of secure outsourcing of large-scale nonnegative matrix factorization (NMF) to a cloud in a way that the client can verify the correctness of results with small overhead. The input matrix protection is achieved by a lightweight, permutation-based encryption mechanism. By exploiting the iterative nature of NMF computation, we propose a single-round verification strategy, which can be proved to be effective. Both theoretical and experimental results are given to demonstrate the superior performance of our scheme.

As training samples are not always identically distributed with the clutter in the cell under test (CUT) in heterogeneous environments, the estimated clutter covariance matrix for space-time adaptive processing (STAP) is not accurate, which degrades the performance of STAP. To improve the performance of STAP in heterogeneous environments, this letter proposes a novel training sample selection algorithm to estimate the covariance matrix. Based on the subaperture smoothing techniques, subapertures' covariance matrices are estimated, which are used to measure the similarities between the clutter covariance matrix of the CUT and the clutter covariance matrices of the training samples. Training samples whose clutter covariance matrices are similar to that of the CUT are selected, leading to a better estimation of the clutter covariance matrix, and the performance of STAP improves. Experimental results confirm the performance of the proposed algorithm.

In this letter, a novel polarimetric target decomposition (PTD) method based on the attributed scattering center (ASC) model is proposed for man-made targets in synthetic aperture radar (SAR) images. By extracting attributed parameters, polarimetric characteristics of targets can be exploited by performing PTD on the extracting parameters of ASCs instead of pixels in conventional PTD algorithms. As a result, the integrity of target components is enhanced, leading to a reliable analysis on the polarimetric scattering mechanisms of SAR targets. In the proposal, an attributed parameters extraction method based on joint exploitation of multiple polarimetric channels and a target discriminating method based on a constant-false-alarm threshold are developed to improve its robustness in strong noise scenarios. Experimental results confirm the effectiveness of the proposed algorithm.