作者：
Peng Tang;Huihuang Zhao*;Weiliang Meng;Yaonan Wang
期刊：
Expert Systems with Applications,2026年297:129344 ISSN：0957-4174
通讯作者：
Huihuang Zhao
作者机构：
[Peng Tang; Huihuang Zhao] Hengyang Normal University, College of Computer Science and Technology, Hengyang, 421002, Hunan, China;[Weiliang Meng] University of Chinese Academy of Sciences, School of Artificial Intelligence, Beijing, 100049, China;[Yaonan Wang] Hunan University, National Engineering Laboratory for Robot Visual Perception and Control Technology, Changsha, 410082, Hunan, China
通讯机构：
[Huihuang Zhao] H;Hengyang Normal University, College of Computer Science and Technology, Hengyang, 421002, Hunan, China
摘要：
Exciting achievements have been made in audio-driven face-talking head generation. However, while existing methods excel at generating a speaking head from a frontal identity image, generating a speaking head with a head pose does not yield satisfactory results when the identity is based on a side image of the face. To address this limitation, a concise and effective approach is proposed in this work. Our method generates efficient talking head videos using a side face image as the identity. It uses facial features and head posture to predict frontal keypoints, during which facial expression features are added as additional features. The method achieves impressive head motion effects and maintains identity information consistency during motion. Additionally, a Transform-based lip-sync expert discriminator is designed to guide the lip synchronization of the generated video. A cross-attention is employed to jointly learn the front and back frames, extracting contextual information that improves lip synchronization over long sequences. Our method is tested on several datasets, such as VoxCeleb2 and LRS2, and compared with multiple advanced methods, demonstrating its outstanding performance through experiments.
Exciting achievements have been made in audio-driven face-talking head generation. However, while existing methods excel at generating a speaking head from a frontal identity image, generating a speaking head with a head pose does not yield satisfactory results when the identity is based on a side image of the face. To address this limitation, a concise and effective approach is proposed in this work. Our method generates efficient talking head videos using a side face image as the identity. It uses facial features and head posture to predict frontal keypoints, during which facial expression features are added as additional features. The method achieves impressive head motion effects and maintains identity information consistency during motion. Additionally, a Transform-based lip-sync expert discriminator is designed to guide the lip synchronization of the generated video. A cross-attention is employed to jointly learn the front and back frames, extracting contextual information that improves lip synchronization over long sequences. Our method is tested on several datasets, such as VoxCeleb2 and LRS2, and compared with multiple advanced methods, demonstrating its outstanding performance through experiments.
作者机构：
College of Computer Science and Technology, Hengyang Normal University, Hengyang, 421002, China;Hunan Provincial Key Laboratory of Intelligent Information Processing and Application, Hengyang Normal University, Hengyang, 421002, China;Hunan Engineering Research Center of Cyberspace Security Technology and Applications, Hengyang Normal University, Hengyang, 421002, China;[Xingqi Yue; Qingling Song] College of Computer Science and Technology, Hengyang Normal University, Hengyang, 421002, China<&wdkj&>Hunan Provincial Key Laboratory of Intelligent Information Processing and Application, Hengyang Normal University, Hengyang, 421002, China<&wdkj&>Hunan Engineering Research Center of Cyberspace Security Technology and Applications, Hengyang Normal University, Hengyang, 421002, China
摘要：
Lightweight cryptography is crucial for securing data in resource-constrained IoT devices. However, most existing lightweight block ciphers suffer from slow diffusion, high hardware cost, or insufficient side-channel resistance, limiting their practical deployment. To address these challenges, this work proposes HDHL, a 64-bit block cipher that integrates a hybrid Generalized Feistel and Substitution-Permutation Network (GSP) structure. The design features a 64-bit sponge-based F -function for strong nonlinearity, lightweight AND-Rotation-XOR (AND-RX) operations for enhanced confusion, and an involutive linear layer that enables efficient encryption and decryption with a unified datapath. A compact LFSR-driven key schedule delivers four 16-bit sub-keys per round with near-ideal entropy. Experimental results show that HDHL achieves full diffusion in just two rounds, area occupies only 1756.71 gate equivalents (GE), and energy 7.82 μ J/bit at a 100 kHz test clock, outperforming CRAFT and SKINNY-64 in both area and energy. Security evaluation demonstrates strong resistance against differential, linear, integral, and algebraic attacks, with a seven-round differential probability below 2 − 70 . Fixed-random t -tests on a prototype ASIC confirm first-order side-channel resistance when lightweight masking or a two-share threshold implementation is applied. These results indicate that HDHL offers a balanced combination of security, efficiency, and implementation cost, making it a promising candidate for future low-power embedded applications.
Lightweight cryptography is crucial for securing data in resource-constrained IoT devices. However, most existing lightweight block ciphers suffer from slow diffusion, high hardware cost, or insufficient side-channel resistance, limiting their practical deployment. To address these challenges, this work proposes HDHL, a 64-bit block cipher that integrates a hybrid Generalized Feistel and Substitution-Permutation Network (GSP) structure. The design features a 64-bit sponge-based F -function for strong nonlinearity, lightweight AND-Rotation-XOR (AND-RX) operations for enhanced confusion, and an involutive linear layer that enables efficient encryption and decryption with a unified datapath. A compact LFSR-driven key schedule delivers four 16-bit sub-keys per round with near-ideal entropy. Experimental results show that HDHL achieves full diffusion in just two rounds, area occupies only 1756.71 gate equivalents (GE), and energy 7.82 μ J/bit at a 100 kHz test clock, outperforming CRAFT and SKINNY-64 in both area and energy. Security evaluation demonstrates strong resistance against differential, linear, integral, and algebraic attacks, with a seven-round differential probability below 2 − 70 . Fixed-random t -tests on a prototype ASIC confirm first-order side-channel resistance when lightweight masking or a two-share threshold implementation is applied. These results indicate that HDHL offers a balanced combination of security, efficiency, and implementation cost, making it a promising candidate for future low-power embedded applications.
作者：
Xiaoqing Wan*;Hui Liu;Feng Chen;Kun Hu;Zhize Li
期刊：
Digital Signal Processing,2026年168:105561 ISSN：1051-2004
通讯作者：
Xiaoqing Wan
作者机构：
[Hui Liu; Feng Chen; Kun Hu; Zhize Li] Hengyang Normal University, College of Computer Science and Technology, 421002, Hengyang, China;Hunan Provincial Key Laboratory of Intelligent Information Processing and Application, 421002, Hengyang, China;[Xiaoqing Wan] Hengyang Normal University, College of Computer Science and Technology, 421002, Hengyang, China<&wdkj&>Hunan Provincial Key Laboratory of Intelligent Information Processing and Application, 421002, Hengyang, China
通讯机构：
[Xiaoqing Wan] H;Hengyang Normal University, College of Computer Science and Technology, 421002, Hengyang, China<&wdkj&>Hunan Provincial Key Laboratory of Intelligent Information Processing and Application, 421002, Hengyang, China
摘要：
In recent years, the combination of convolutional neural networks (CNNs) with transformers for spectral-spatial feature extraction and robust semantic modeling has greatly improved the performance in hyperspectral image (HSI) classification tasks. However, these methods often overlook frequency information; CNNs struggle to capture global dependencies due to limited receptive fields, and transformers tend to lose fine-grained local structures and high-frequency variations. To address these challenges, this paper proposes a Laplacian frequency aware hierarchical network (LFAH-Net). We first design the method employing a diversity frequency-aware transformer (DFAT) module alongside a multi-level frequency fusion block (MFFB) stack to explicitly separate and integrate high-frequency signals such as edges and textures, as well as low-frequency signals like spectral contours, thereby achieving cross-level frequency feature complementarity. Besides, we propose a spectral-spatial adaptive recalibration fusion (SSARF) module, specifically designed to correct misalignments and suppress noise in hyperspectral features. Finally, the multi-scale dilation convolution (MSDC) module utilizes dilated convolutions to capture both local and global contextual information, while the adaptive feature fusion (AFF) module adaptively recalibrates and fuses these features with the spectral representations from DFAT. Experimental results on four popular hyperspectral datasets demonstrate that our framework significantly outperforms several state-of-the-art methods.
In recent years, the combination of convolutional neural networks (CNNs) with transformers for spectral-spatial feature extraction and robust semantic modeling has greatly improved the performance in hyperspectral image (HSI) classification tasks. However, these methods often overlook frequency information; CNNs struggle to capture global dependencies due to limited receptive fields, and transformers tend to lose fine-grained local structures and high-frequency variations. To address these challenges, this paper proposes a Laplacian frequency aware hierarchical network (LFAH-Net). We first design the method employing a diversity frequency-aware transformer (DFAT) module alongside a multi-level frequency fusion block (MFFB) stack to explicitly separate and integrate high-frequency signals such as edges and textures, as well as low-frequency signals like spectral contours, thereby achieving cross-level frequency feature complementarity. Besides, we propose a spectral-spatial adaptive recalibration fusion (SSARF) module, specifically designed to correct misalignments and suppress noise in hyperspectral features. Finally, the multi-scale dilation convolution (MSDC) module utilizes dilated convolutions to capture both local and global contextual information, while the adaptive feature fusion (AFF) module adaptively recalibrates and fuses these features with the spectral representations from DFAT. Experimental results on four popular hyperspectral datasets demonstrate that our framework significantly outperforms several state-of-the-art methods.
摘要：
Currently, Camouflaged Object Detection (COD) methods often rely on single-view feature perception, which struggles to fully capture camouflaged objects due to environmental interference such as background clutter, lighting variations, and viewpoint changes. To address this, we propose the multi-view collaboration network (MCNet), inspired by human visual strategies for complex scene analysis. MCNet incorporates multiple perspectives for enhanced feature extraction. The global perception module takes the original, far, and near views, using different large-kernel convolutions and multi-head attention mechanisms for global feature embedding. In parallel, the local perception module processes the tilted, projected, and color-jittered views, extracting fine-grained local features through multi-branch deep convolutions and dilated convolutions. To facilitate deep interaction between global and local features, we introduce the hybrid interactive module, which explores the correlation of multi-view feature information and adaptively fuses features. For feature decoding, the dynamic pyramid shrinkage module integrates dynamic gated convolutions with a pyramid shrinkage mechanism, progressively aggregating semantic features through a hierarchical shrinking strategy and group fusion strategy. Experimental results on popular COD benchmark datasets show that MCNet outperforms 18 state-of-the-art methods.
Currently, Camouflaged Object Detection (COD) methods often rely on single-view feature perception, which struggles to fully capture camouflaged objects due to environmental interference such as background clutter, lighting variations, and viewpoint changes. To address this, we propose the multi-view collaboration network (MCNet), inspired by human visual strategies for complex scene analysis. MCNet incorporates multiple perspectives for enhanced feature extraction. The global perception module takes the original, far, and near views, using different large-kernel convolutions and multi-head attention mechanisms for global feature embedding. In parallel, the local perception module processes the tilted, projected, and color-jittered views, extracting fine-grained local features through multi-branch deep convolutions and dilated convolutions. To facilitate deep interaction between global and local features, we introduce the hybrid interactive module, which explores the correlation of multi-view feature information and adaptively fuses features. For feature decoding, the dynamic pyramid shrinkage module integrates dynamic gated convolutions with a pyramid shrinkage mechanism, progressively aggregating semantic features through a hierarchical shrinking strategy and group fusion strategy. Experimental results on popular COD benchmark datasets show that MCNet outperforms 18 state-of-the-art methods.
作者机构：
[Li, Lang] Hengyang Normal Univ, Coll Comp Sci & Technol, Hengyang 421002, Peoples R China.;Hengyang Normal Univ, Hunan Prov Key Lab Intelligent Informat Proc & App, Hengyang 421002, Peoples R China.
通讯机构：
[Li, L ] H;Hengyang Normal Univ, Coll Comp Sci & Technol, Hengyang 421002, Peoples R China.
关键词：
Side-channel analysis;Sample correlation locally;Deep learning;Kernel density estimation;Profiling analysis
摘要：
Label distribution learning techniques can significantly enhance the effectiveness of side-channel analysis. However, this method relies on using probability density functions to estimate the relationships between labels. The settings of parameters play a crucial role in the impact of the attacks. This study introduces a non-parametric statistical method to calculate the distribution between labels, specifically employing smoothing with the Gaussian kernel function and adjusting bandwidth. Then, the aggregation of the results from each label processed by the Gaussian kernel facilitates a hypothesis-free estimation of the label distribution. This method accurately represents the actual leakage distribution, speeding up guess entropy convergence. Secondly, we exploit similarities between profiling traces, proposing an analysis scheme for sample correlation locally of label distribution learning. Furthermore, Signal to-Noise Ratio (SNR) is employed to re-extract and reduce dataset dimensions to 500 power consumption points, resulting in noise reduction. Our results showcase the successful training of 800 profiling traces using our method for sample correlation locally of label distribution learning, with the findings indicating its exceptional performance.
Label distribution learning techniques can significantly enhance the effectiveness of side-channel analysis. However, this method relies on using probability density functions to estimate the relationships between labels. The settings of parameters play a crucial role in the impact of the attacks. This study introduces a non-parametric statistical method to calculate the distribution between labels, specifically employing smoothing with the Gaussian kernel function and adjusting bandwidth. Then, the aggregation of the results from each label processed by the Gaussian kernel facilitates a hypothesis-free estimation of the label distribution. This method accurately represents the actual leakage distribution, speeding up guess entropy convergence. Secondly, we exploit similarities between profiling traces, proposing an analysis scheme for sample correlation locally of label distribution learning. Furthermore, Signal to-Noise Ratio (SNR) is employed to re-extract and reduce dataset dimensions to 500 power consumption points, resulting in noise reduction. Our results showcase the successful training of 800 profiling traces using our method for sample correlation locally of label distribution learning, with the findings indicating its exceptional performance.
通讯机构：
[Zhao, HH ] H;Hengyang Normal Univ, Coll Comp Sci & Technol, Hengyang 421002, Hunan, Peoples R China.
关键词：
Text to image;Image generation;Generative Adversarial Network;Attention
摘要：
Text-to-image generation is a challenging and significant research task. It aims to synthesize high-quality images that match the given descriptive statements. Existing methods still have problems in generating semantic information fusion insufficiently, and the generated images cannot represent the descriptive statements properly. Therefore, A novel method named EMF-GAN (Efficient Multilayer Fusion Generative Adversarial Network) is proposed. It uses a Multilayer Fusion Module (MF Module) and Efficient Multi-Scale Attention Module (EMA Module) to fuse the semantic information into the feature maps gradually. It realizes the full utilization of the semantic information and obtains high-quality realistic images. Extensive experimental results show that our EMF-GAN is highly competitive in image generation quality and semantic consistency . Compared with the state-of-the-art methods, EMF-GAN shows significant performance improvement on both CUB (FID from 14.81 to 10.74) and COCO (FID from 19.32 to 16.86) datasets. It can generate photorealistic images with richer details and text-image consistency. Code can be found at https://github.com/zxcnmmmmm/EMF-GAN-master .
Text-to-image generation is a challenging and significant research task. It aims to synthesize high-quality images that match the given descriptive statements. Existing methods still have problems in generating semantic information fusion insufficiently, and the generated images cannot represent the descriptive statements properly. Therefore, A novel method named EMF-GAN (Efficient Multilayer Fusion Generative Adversarial Network) is proposed. It uses a Multilayer Fusion Module (MF Module) and Efficient Multi-Scale Attention Module (EMA Module) to fuse the semantic information into the feature maps gradually. It realizes the full utilization of the semantic information and obtains high-quality realistic images. Extensive experimental results show that our EMF-GAN is highly competitive in image generation quality and semantic consistency . Compared with the state-of-the-art methods, EMF-GAN shows significant performance improvement on both CUB (FID from 14.81 to 10.74) and COCO (FID from 19.32 to 16.86) datasets. It can generate photorealistic images with richer details and text-image consistency. Code can be found at https://github.com/zxcnmmmmm/EMF-GAN-master .
摘要：
Deep learning-based side-channel attacks (DL-SCA) are favored for their strong key recovery capabilities. However, their implementation is based on the attacker being able to manipulate a cloned device to build an attack model, which means that the attacker needs to know secret information in advance. The non-profiled side-channel attacks (NP-SCA) methods can complete the key recovery without knowing the secret information. Differential Deep Learning Analysis (DDLA) is the first NP-DLSCA method proposed in CHES2019, and several improved versions appeared later. In these methods, the bad quality of the raw traces, such as noise, random delay, etc., is often ignored, which limits the efficiency of key recovery. In this work, the conditional generative adversarial network (CGAN) is introduced and a novel framework NPSCA-CGAN is proposed to optimize traces in non-profiled SCA scenarios. We apply CGAN in non-profiled attacks and use plaintext to do trace labeling that optimizes the raw traces by training the generator to learn the label traces. The convolutional module and plaintext feature are added to the generator network to adapt various countermeasures. Moreover, a new traces quality evaluation metric average relative signal-to-noise ratio (AR-SNR) is proposed for non-profiled attack scenarios, which can directly reflect the performance of the traces in practical attack. The method is applied to unprotected, unaligned, and masked traces respectively. The experimental results indicate that it can enormously optimize the quality of the traces and improve the efficiency of non-profiled side-channel attacks.
摘要：
The objective of image-based virtual try-on is to seamlessly integrate clothing onto a target image, generating a realistic representation of the character in the specified attire. However, existing virtual try-on methods frequently encounter challenges, including misalignment between the body and clothing, noticeable artifacts, and the loss of intricate garment details. To overcome these challenges, we introduce a two-stage high-resolution virtual try-on framework that integrates an attention mechanism, comprising a garment warping stage and an image generation stage. During the garment warping stage, we incorporate a channel attention mechanism to effectively retain the critical features of the garment, addressing challenges such as the loss of patterns, colors, and other essential details commonly observed in virtual try-on images produced by existing methods. During the image generation stage, with the aim of maximizing the utilization of the information proffered by the input image, the input features undergo double sampling within the normalization procedure, thereby enhancing the detail fidelity and clothing alignment efficacy of the output image. Experimental evaluations conducted on high-resolution datasets validate the effectiveness of the proposed method. Results demonstrate significant improvements in preserving garment details, reducing artifacts, and achieving superior alignment between the clothing and body compared to baseline methods, establishing its advantage in generating realistic and high-quality virtual try-on images.
The objective of image-based virtual try-on is to seamlessly integrate clothing onto a target image, generating a realistic representation of the character in the specified attire. However, existing virtual try-on methods frequently encounter challenges, including misalignment between the body and clothing, noticeable artifacts, and the loss of intricate garment details. To overcome these challenges, we introduce a two-stage high-resolution virtual try-on framework that integrates an attention mechanism, comprising a garment warping stage and an image generation stage. During the garment warping stage, we incorporate a channel attention mechanism to effectively retain the critical features of the garment, addressing challenges such as the loss of patterns, colors, and other essential details commonly observed in virtual try-on images produced by existing methods. During the image generation stage, with the aim of maximizing the utilization of the information proffered by the input image, the input features undergo double sampling within the normalization procedure, thereby enhancing the detail fidelity and clothing alignment efficacy of the output image. Experimental evaluations conducted on high-resolution datasets validate the effectiveness of the proposed method. Results demonstrate significant improvements in preserving garment details, reducing artifacts, and achieving superior alignment between the clothing and body compared to baseline methods, establishing its advantage in generating realistic and high-quality virtual try-on images.
期刊：
Expert Systems with Applications,2025年272:126693 ISSN：0957-4174
通讯作者：
Chen, WH
作者机构：
[Yan, Li; Chen, Wenhui; Zhao, Huihuang; Yang, Yanqing; Wang, Weijie] Hengyang Normal Univ, Coll Comp Sci & Technol, Hengyang 421002, Peoples R China.;[Chen, Wenhui; Yang, Yanqing] Hengyang Normal Univ, Hunan Prov Key Lab Intelligent Informat Proc & App, Hengyang 421002, Peoples R China.;[Zhao, Huihuang] Hunan Univ, Natl Engn Lab Robot Visual Percept & Control Techn, Hengyang, Peoples R China.
通讯机构：
[Chen, WH ] H;Hengyang Normal Univ, Coll Comp Sci & Technol, Hengyang 421002, Peoples R China.;Hengyang Normal Univ, Hunan Prov Key Lab Intelligent Informat Proc & App, Hengyang 421002, Peoples R China.
关键词：
Time series floating point data;Lossless compression;Internet of things;Compression algorithm;Heuristic genetic algorithm
摘要：
The processing of large volumes of time series data across various fields presents significant challenges, particularly when it comes to effectively managing floating-point numbers. Current dual precision floating-point lossless compression algorithms often struggle to deliver exceptional performance on diverse datasets, highlighting their inherent limitations. To address this issue, we propose a novel method called the Heuristic Genetic Algorithm Parameter Optimizer for Lossless Compression of Time Series Floating Point Data (HGA-ACTF). This method features a highly effective parameter optimizer designed specifically for compression algorithms that utilize leading zeros. The combination of our parameter optimizer and the HGA-ACTF algorithm strategy has been proven to outperform existing leading compression algorithms across multiple fields. This approach not only enhances the compression ratio but also significantly reduces both compression and decompression times. In our comparative study, we evaluated the HGA-ACTF algorithm against eleven well-performing algorithms and a variant of the algorithm, integrating our parameter optimizer and algorithmic strategy into other adaptable algorithms, and demonstrating notable improvements. Experimental results indicate that the HGA-ACTF algorithm achieves an average compression ratio improvement of 38.87%, with some datasets showing improvements of up to 54.36%. Our approach effectively addresses the transmission and storage of time series data, significantly reducing the overhead associated with data processing. The code can be found at https://github.com/wwj10/HGA-ACTF .
The processing of large volumes of time series data across various fields presents significant challenges, particularly when it comes to effectively managing floating-point numbers. Current dual precision floating-point lossless compression algorithms often struggle to deliver exceptional performance on diverse datasets, highlighting their inherent limitations. To address this issue, we propose a novel method called the Heuristic Genetic Algorithm Parameter Optimizer for Lossless Compression of Time Series Floating Point Data (HGA-ACTF). This method features a highly effective parameter optimizer designed specifically for compression algorithms that utilize leading zeros. The combination of our parameter optimizer and the HGA-ACTF algorithm strategy has been proven to outperform existing leading compression algorithms across multiple fields. This approach not only enhances the compression ratio but also significantly reduces both compression and decompression times. In our comparative study, we evaluated the HGA-ACTF algorithm against eleven well-performing algorithms and a variant of the algorithm, integrating our parameter optimizer and algorithmic strategy into other adaptable algorithms, and demonstrating notable improvements. Experimental results indicate that the HGA-ACTF algorithm achieves an average compression ratio improvement of 38.87%, with some datasets showing improvements of up to 54.36%. Our approach effectively addresses the transmission and storage of time series data, significantly reducing the overhead associated with data processing. The code can be found at https://github.com/wwj10/HGA-ACTF .
作者机构：
College of Computer Science and Technology, Hengyang Normal University, Hengyang, China;Hunan Provincial Key Laboratory of Intelligent Information Processing and Application, Hengyang Normal University, Hengyang, China;[Yezhou (59366878200); Lang (36070638700); Yu (57202403406)] College of Computer Science and Technology, Hengyang Normal University, Hengyang, China<&wdkj&>Hunan Provincial Key Laboratory of Intelligent Information Processing and Application, Hengyang Normal University, Hengyang, China
通讯机构：
[Lang Li] C;College of Computer Science and Technology, Hengyang Normal University, Hengyang, China<&wdkj&>Hunan Provincial Key Laboratory of Intelligent Information Processing and Application, Hengyang Normal University, Hengyang, China
摘要：
Deep learning algorithms are increasingly employed to exploit side-channel information, such as power consumption and electromagnetic leakage from hardware devices, significantly enhancing attack capabilities. However, relying solely on power traces for side-channel information often requires adequate domain knowledge. To address this limitation, this work proposes a new attack scheme. Firstly, a Convolutional Neural Network (CNN)-based plaintext-extended bilinear feature fusion model is designed. Secondly, multi-model intermediate layers are fused and trained, yielding in the increase of the amount of effective information and generalization ability. Finally, the model is employed to predict the output probability of three public side-channel datasets (e.g. ASCAD, AES
$$\_$$
HD, and AES
$$\_$$
RD), and analyze the recovery key guessing entropy for each key to efficiently assess attack efficiency. Experimental results showcase that the plaintext-extended bilinear feature fusion model can effectively enhance the Side-Channel Attack (SCA) capabilities and prediction performance. Deploying the proposed method, the number of traces required for a successful attack on the ASCAD
$$\_$$
R dataset is significantly reduced to less than 914, representing an 70.5% reduction in traces compared to the network in Convolutional Neural Network-Visual Geometry Group (CNNVGG16) with plaintext, which incorporating plaintext features before the fully connected layer. Compared to existing solutions, the proposed scheme requires only 80% of the power traces for the attack mask design using only 75 epochs. As a result, the power of the proposed method is well proved through the different experiments and comparison processes.
摘要：
Hyperspectral images (HSIs) contain rich spectral and spatial information, motivating the development of a novel circulant singular spectrum analysis (CiSSA) and multiscale local ternary pattern fusion method for joint spectral-spatial feature extraction and classification. Due to the high dimensionality and redundancy in HSIs, principal component analysis (PCA) is used during preprocessing to reduce dimensionality and enhance computational efficiency. CiSSA is then applied to the PCA-reduced images for robust spatial pattern extraction via circulant matrix decomposition. The spatial features are combined with the global spectral features from PCA to form a unified spectral-spatial feature set (SSFS). Local ternary pattern (LTP) is further applied to the principal components (PCs) to capture local grayscale and rotation-invariant texture features at multiple scales. Finally, the performance of the SSFS and multiscale LTP features is evaluated separately using a support vector machine (SVM), followed by decision-level fusion to combine results from each pipeline based on probability outputs. Experimental results on three popular HSIs show that, under 1% training samples, the proposed method achieves 95.98% accuracy on the Indian Pines dataset, 98.49% on the Pavia University dataset, and 92.28% on the Houston2013 dataset, outperforming several traditional classification methods and state-of-the-art deep learning approaches.
摘要：
Deep learning-assisted template attack (DLATA) is a novel side-channel attack (SCA) method proposed by Lichao Wu at CHES2022. It utilizes a triplet network to assist template attacks (TA), avoiding the redundant training and hyperparameter tuning required in traditional DL-based SCA methods. However, the training of the triplet network requires a large number of power samples due to its unique structure. We propose a new optimization scheme, in which the transfer learning (TL) technology is used to train multiple models on several similar datasets with fewer power traces, to mitigation the problem. The approach allows us to leverage pre-trained models to product a new mode on the another target dataset by fine-tuning weights so that significantly reduce the training cost for the triplet network while maintaining attack effectiveness. We remould the structure and dimensionality of similar datasets so that the models trained on them can perform effective transfer learning for training on the target dataset. Concretely, some of parameters and features obtained from pretraining can be used directly for the target task, while the rest only require fine-tuning. Evaluation and experimental validation on the public ASCAD dataset demonstrate that our method achieves or even surpasses the performance of the original method with a 90% reduction in the training set. These findings highlight the effectiveness of the proposed TL strategy in achieving robust attack performance in low-sample training environments.
Deep learning-assisted template attack (DLATA) is a novel side-channel attack (SCA) method proposed by Lichao Wu at CHES2022. It utilizes a triplet network to assist template attacks (TA), avoiding the redundant training and hyperparameter tuning required in traditional DL-based SCA methods. However, the training of the triplet network requires a large number of power samples due to its unique structure. We propose a new optimization scheme, in which the transfer learning (TL) technology is used to train multiple models on several similar datasets with fewer power traces, to mitigation the problem. The approach allows us to leverage pre-trained models to product a new mode on the another target dataset by fine-tuning weights so that significantly reduce the training cost for the triplet network while maintaining attack effectiveness. We remould the structure and dimensionality of similar datasets so that the models trained on them can perform effective transfer learning for training on the target dataset. Concretely, some of parameters and features obtained from pretraining can be used directly for the target task, while the rest only require fine-tuning. Evaluation and experimental validation on the public ASCAD dataset demonstrate that our method achieves or even surpasses the performance of the original method with a 90% reduction in the training set. These findings highlight the effectiveness of the proposed TL strategy in achieving robust attack performance in low-sample training environments.
摘要：
With the rapid development of generative models , there is an increasing demand for universal fake image detectors. In this paper, we investigate the problem of fake image detection for the synthesis of generative models to detect fake images from multiple generative methods . Recent research methods explore the benefits of pre-trained models and mainly adopt a fixed paradigm of training additional classifiers separately, but we find that the fixed paradigm hinders the full learning of forgery features, leading to insufficient representation learning in the detector. The main reason is that the fixed paradigm pays too much attention to global features and neglects local features , which limits the ability of the model to perceive image details and leads to some information loss or confusion. In this regard, based on the pre-trained visual language space, our method introduces two core designs. First, we designed a Deep Window Triple Attention (DWTA) module. A similar dense sliding window strategy is adopted to capture multi-scale local abnormal patterns, and the sensitivity to generated artifacts is enhanced through the triple attention mechanism . Secondly, we proposed a Cross-Space Feature Alignment (CSFA) module. A two-way interactive channel between global features and local features is established, and the alignment loss function is used to achieve semantic alignment of cross-modal feature spaces. The aligned features are adaptively fused through a gating mechanism to obtain the final adaptive forged features. Experiments demonstrate that our method, when trained solely on ProGAN data, achieves superior cross-generator generalization: it attains an average accuracy of 94.7% on unseen GANs and generalizes to unseen diffusion models with 94% accuracy, surpassing existing methods by 2.1%. The source code is available at https://github.com/long2580h/GLFAFormer .
With the rapid development of generative models , there is an increasing demand for universal fake image detectors. In this paper, we investigate the problem of fake image detection for the synthesis of generative models to detect fake images from multiple generative methods . Recent research methods explore the benefits of pre-trained models and mainly adopt a fixed paradigm of training additional classifiers separately, but we find that the fixed paradigm hinders the full learning of forgery features, leading to insufficient representation learning in the detector. The main reason is that the fixed paradigm pays too much attention to global features and neglects local features , which limits the ability of the model to perceive image details and leads to some information loss or confusion. In this regard, based on the pre-trained visual language space, our method introduces two core designs. First, we designed a Deep Window Triple Attention (DWTA) module. A similar dense sliding window strategy is adopted to capture multi-scale local abnormal patterns, and the sensitivity to generated artifacts is enhanced through the triple attention mechanism . Secondly, we proposed a Cross-Space Feature Alignment (CSFA) module. A two-way interactive channel between global features and local features is established, and the alignment loss function is used to achieve semantic alignment of cross-modal feature spaces. The aligned features are adaptively fused through a gating mechanism to obtain the final adaptive forged features. Experiments demonstrate that our method, when trained solely on ProGAN data, achieves superior cross-generator generalization: it attains an average accuracy of 94.7% on unseen GANs and generalizes to unseen diffusion models with 94% accuracy, surpassing existing methods by 2.1%. The source code is available at https://github.com/long2580h/GLFAFormer .
作者机构：
[Wang, Fangyan; Jiao, Ge; Yue, Guowen] Hengyang Normal Univ, Coll Comp Sci & Technol, Hengyang 421002, Hunan, Peoples R China.;[Jiao, Ge] Hengyang Normal Univ, Hunan Prov Key Lab Intelligent Informat Proc & App, Hengyang 421002, Hunan, Peoples R China.;[Jiao, Ge] Hengyang Normal Univ, Hunan Engn Res Ctr Cyberspace Secur Technol & Appl, Hengyang 421002, Hunan, Peoples R China.
通讯机构：
[Jiao, G ] H;Hengyang Normal Univ, Coll Comp Sci & Technol, Hengyang 421002, Hunan, Peoples R China.;Hengyang Normal Univ, Hunan Prov Key Lab Intelligent Informat Proc & App, Hengyang 421002, Hunan, Peoples R China.;Hengyang Normal Univ, Hunan Engn Res Ctr Cyberspace Secur Technol & Appl, Hengyang 421002, Hunan, Peoples R China.
关键词：
Camouflaged object detection;Self-supervised learning;Reconstruction;Adapter;Segment anything model
摘要：
Current Camouflaged Object Detection (COD) methods primarily rely on a direct mapping from image to mask. However, due to the inherent semantic and structural gap between the image and its corresponding mask, the learned feature representations often exhibit poor generalization ability. To address this issue, we propose a novel intra-domain dual reconstruction framework, termed InDReCT, which reformulates the image-to-mask prediction as a cross-domain transfer task by simultaneously reconstructing both the input image and its corresponding mask. Within this framework, semantic knowledge is transferred through two reconstruction processes from different domains: image reconstruction (appearance domain) and mask reconstruction (structure domain), and is eventually integrated back into the image-to-mask prediction task. This dual reconstruction mechanism implicitly guides the network to extract hidden appearance semantics from image-to-image reconstruction and explicit structural information from mask-to-mask reconstruction, thereby enhancing the model’s generalization capability. Extensive experiments on three benchmark COD datasets and four downstream tasks demonstrate that InDReCT consistently outperforms state-of-the-art methods in both detection accuracy and generalization ability. Notably, on the widely-used COD10K dataset, InDReCT achieves a Mean E-measure ( E m ) of 95.6 %, surpassing the latest state-of-the-art model CamoDiffusion by 1.6 %. Code and models will be publicly available at: https://github.com/KungFuProgrammerle/InDReCT .
Current Camouflaged Object Detection (COD) methods primarily rely on a direct mapping from image to mask. However, due to the inherent semantic and structural gap between the image and its corresponding mask, the learned feature representations often exhibit poor generalization ability. To address this issue, we propose a novel intra-domain dual reconstruction framework, termed InDReCT, which reformulates the image-to-mask prediction as a cross-domain transfer task by simultaneously reconstructing both the input image and its corresponding mask. Within this framework, semantic knowledge is transferred through two reconstruction processes from different domains: image reconstruction (appearance domain) and mask reconstruction (structure domain), and is eventually integrated back into the image-to-mask prediction task. This dual reconstruction mechanism implicitly guides the network to extract hidden appearance semantics from image-to-image reconstruction and explicit structural information from mask-to-mask reconstruction, thereby enhancing the model’s generalization capability. Extensive experiments on three benchmark COD datasets and four downstream tasks demonstrate that InDReCT consistently outperforms state-of-the-art methods in both detection accuracy and generalization ability. Notably, on the widely-used COD10K dataset, InDReCT achieves a Mean E-measure ( E m ) of 95.6 %, surpassing the latest state-of-the-art model CamoDiffusion by 1.6 %. Code and models will be publicly available at: https://github.com/KungFuProgrammerle/InDReCT .
摘要：
Classifying hyperspectral images (HSIs) is a key challenge in remote sensing, with convolutional neural networks (CNNs) and transformer models becoming leading techniques in this area. CNNs, while effective, often struggle to adequately capture intricate semantic features, and increasing network depth leads to significantly higher computational costs. Conversely, transformers, despite their efficacy in modeling spectral-spatial dependencies, introduce significant computational overhead due to their complexity. Mamba, leveraging the state space model (SSM), presents a compelling alternative that efficiently captures long-range dependencies in HSIs while ensuring computational efficiency with linear complexity. To improve the classification performance of HSIs by simultaneously extracting rich local and global spatial-spectral features, as well as deep semantic features, while reducing the computational complexity of the model, this paper proposes an innovative hybrid large selective kernel and convolutional additive self-attention model (HLSK-CASMamba) for HSI classification. First, we design a feature extraction module that combines a 3D convolution layer, a 2D convolution layer, and a large selective kernel (LSK) network, enabling the efficient extraction of both depth-related and spatial details information from HSIs. Second, we propose a novel CASMamba model, with its core module, CAS-VSSM, combining convolutional additive self-attention (CAS) and the vision state-space sequence model (VSSM). This fusion leverages the local feature extraction of convolutions, spatial dependency modeling of self-attention, and long-range dependency handling of VSSM, enhancing the capture of both local and global context while ensuring computational efficiency. Finally, we incorporate the KANLinear module to replace the traditional linear layer, enhancing sample label acquisition. Extensive evaluations on three popular HSIs show that, under 10% training samples, the proposed method achieves 99.57% accuracy on the Houston 2013 dataset, 99.96% on the Botswana dataset, and 99.92% on the University of Pavia dataset, outperforming various existing advanced techniques.
通讯机构：
[Li, L ] H;Hengyang Normal Univ, Coll Comp Sci & Technol, Hunan Prov Key Lab Intelligent Informat Proc & App, Hengyang 421002, Peoples R China.
关键词：
Feistel structure;Lai-Massey structure;light-weight S-box;light-weight S-box;differential power analysis;differential power analysis
摘要：
Lightweight S-boxes have a significant impact on cryptographic algorithms for resource-constrained devices in the Internet of Things (IoT). Previous research on such S-oxes has focused on low area and good cryptographic properties, which neglects the resistance to side-channel analysis, especially differential power analysis (DPA). The reVisited transparency order (VTO) is one of the best indicators for evaluating S-boxes against DPA attacks so far. Therefore, this paper presents a scheme for designing 6-bit S-boxes. The designed S-boxes are suitable for lightweight block ciphers and have a certain ability to resist DPA attacks. Specifically, we first identify 23 3-bit S-boxes with the optimal cryptographic properties and the smallest hardware area through systematic screening. Then, the paper investigates the constructions of lightweight 6-bit S-boxes based on traditional algorithm structures. Finally, a novel structure named F-LM-F is proposed for designing lightweight 6-bit S-boxes by combining Feistel structure and Lai-Massey structure. It has been proven through comparison that the S-boxes under F-LM-F structure achieve fewer fixed points and lower VTO than the 6-bit S-box of BipBip cipher, with a 17.61% reduction in hardware area and a 13.58% decrease in CPU cycles.
作者机构：
College of Computer Science and Technology, Hengyang Normal University, Hengyang, China;Hunan Provincial Key Laboratory of Intelligent Information Processing and Application, Hengyang Normal University, Hengyang, China;[Zhiying Feng; Lang Li] College of Computer Science and Technology, Hengyang Normal University, Hengyang, China<&wdkj&>Hunan Provincial Key Laboratory of Intelligent Information Processing and Application, Hengyang Normal University, Hengyang, China
通讯机构：
[Lang Li] C;College of Computer Science and Technology, Hengyang Normal University, Hengyang, China<&wdkj&>Hunan Provincial Key Laboratory of Intelligent Information Processing and Application, Hengyang Normal University, Hengyang, China
摘要：
The deep learning-assisted template attack (DLATA) is an advanced side-channel attack (SCA) technique that employs a triplet network to embed side-channel features efficiently within a template attack framework. This method utilizes a triplet network to efficiently embed input data into a template attack (TA) in a single training iteration. Although the triplet network is highly effective in feature extraction, its performance deteriorates significantly in high-noise environments, thereby highlighting the need for further improvements. The triplet network’s embedded representations are refined using an SVM classifier, which significantly enhances classification accuracy and robustness in high-dimensional spaces. This study proposes an enhanced DLATA framework that integrates multi-domain feature fusion using wavelet transform (WT) and support vector machine (SVM) techniques to address these challenges. This approach is referred to as MD-DLATA. Specifically, Daubechies 4 (db4) is chosen as the mother wavelet function for multi-resolution feature extraction on side-channel data, allowing the triplet network to learn within a denoised, high-quality feature space. The triplet network’s embedded representations are further refined using an SVM classifier with a radial basis function (RBF) kernel. Experiments on the publicly available ASCAD dataset confirm that our method exhibits strong resistance to noise. Across various noise levels, our method reduces the number of traces required to achieve Guessing Entropy (GE) < 1 by 17-29% compared to the original method.
摘要：
Epicanthus refers to the longitudinal curved skin folds that cover the medial canthus, which affect aesthetics due to covering the medial canthus angle and lacrimal mound. Various surgical methods exist for correcting epicanthus, each with its own set of advantages and disadvantages, and lacking a standardized operational protocol, making it difficult for beginners to master and for clinical promotion.This article aims to explore a standardized and simplified five-step procedure for treating epicanthus and report our clinical experience and effectiveness. A retrospective analysis was conducted from October 2019 to September 2022 at the Burn and Plastic Surgery Department of the Second Affiliated Hospital of South China University. A consistent team of doctors utilized a five-step method to correct the medial canthus in 306 patients with epicanthus. All patients were followed up for more than 6 months. We observed 306 patients and used iris diameter as a reference value to subjectively evaluate the clinical effect through photo evaluation and scar scoring.Objective evaluation of clinical efficacy was achieved through the inter canthal distance (ICD) and palpebral fissure length (PFL). The study included 295 females and 11 males, with an average follow-up time of 14.2 months.The average increase rate of PFL is 14.9%, and the average reduction rate of ICD is 8.6%. Two cases of bleeding and swelling were promptly treated, and no long-term complications were left. 85 cases of scar hyperplasia were treated with KELO-COTE® silicone gel, triamcinolone injection, and appropriate laser therapy in combination, and the scars gradually resolved after 12 months. 4 cases of recurrence and 2 cases of asymmetry underwent reoperation. Observing the satisfaction and effectiveness rate of 306 patients, the overall satisfaction and effectiveness rate reached over 95%. About 96.40% of patients were satisfied with the surgery and would recommend it to their family and friends. The paired t-test was used for statistical analysis, and the results showed statistical significance. The five-step method for correcting epicanthus proves to be a simple, efficient, and reliable technique that is easily mastered by beginners. It boasts high patient satisfaction and carries a low risk of scar formation.
摘要：
With the rapid advancement of digital technology and the pervasive use of media information, ensuring image security and efficient transmission has become increasingly critical. Therefore, this paper proposes a novel 4D hyperchaotic system incorporating a compression encryption algorithm that integrates generalized Fibonacci matrices and heterogeneous scrambling. Firstly, the proposed scheme constructs a novel hyperchaotic system, and its dynamic characteristics are analyzed to validate its rich dynamical behavior, exhibiting high randomness and sensitivity to initial conditions. Secondly, during the compression phase, chaotic sequences regulate image compression, aiming to minimize the number of parameters involved in the encryption process. Thirdly, in the multi-channel heterogeneous scrambling method, the RGB channels are individually subjected to 2D non-equal-length Arnold scrambling, pseudo-random permutation, and higher-order Peano curve fractal scrambling at the pixel level. Furthermore, the scrambling parameters for each channel are dynamically governed by independent chaotic sequences. Finally, this paper proposes a cross-channel nonlinear diffusion algorithm leveraging a 3D dynamic Fibonacci matrix. Through the construction of a spatially coupled encryption scheme, the proposed method ensures triple-layer protection at the pixel, channel, and spatial levels. Experimental results and performance analysis indicate that, at a compression ratio of 0.5, the PSNR exceeds 30 dB, while the SSIM remains above 0.90, reflecting high image reconstruction quality. In addition, the NPCR and the UACI reach approximately 99.60% and 33.46%, respectively. These results confirm that the proposed compression-encryption scheme is highly secure and demonstrates strong resilience against both differential and brute-force attacks.
摘要：
Rice is a staple food for nearly half the global population and, with rising living standards, the demand for high-quality grain is increasing. Chalkiness, a key determinant of appearance quality, requires accurate detection for effective quality evaluation. While traditional 2D imaging has been used for chalkiness detection, its inherent inability to capture complete 3D morphology limits its suitability for precision agriculture and breeding. Although micro-CT has shown promise in 3D chalk phenotype analysis, high-throughput automated 3D detection for multiple grains remains a challenge, hindering practical applications. To address this, we propose a high-throughput 3D chalkiness detection method using micro-CT and VSE-UNet. Our method begins with non-destructive 3D imaging of grains using micro-CT. For the accurate segmentation of kernels and chalky regions, we propose VSE-UNet, an improved VGG-UNet with an SE attention mechanism for enhanced feature learning. Through comprehensive training optimization strategies, including the Dice focal loss function and dropout technique, the model achieves robust and accurate segmentation of both kernels and chalky regions in continuous CT slices. To enable high-throughput 3D analysis, we developed a unified 3D detection framework integrating isosurface extraction, point cloud conversion, DBSCAN clustering, and Poisson reconstruction. This framework overcomes the limitations of single-grain analysis, enabling simultaneous multi-grain detection. Finally, 3D morphological indicators of chalkiness are calculated using triangular mesh techniques. Experimental results demonstrate significant improvements in both 2D segmentation (7.31% improvement in chalkiness IoU, 2.54% in mIoU, 2.80% in mPA) and 3D phenotypic measurements, with VSE-UNet achieving more accurate volume and dimensional measurements compared with the baseline. These improvements provide a reliable foundation for studying chalkiness formation and enable high-throughput phenotyping.
