With the continuous advancement of adversarial attack algorithms,the security risks that deep neural networks face are increasingly severe.Optical phenomena frequently occur in real-world scenarios,and the robustness against optical adversarial attacks directly reflects the safety of deep neural networks.Nevertheless,current research on optical adversarial attacks commonly encounters challenges such as optical perturbation distortion and optimization instability.To solve this problem,this paper proposes a novel optical attack method named AdvFlare,to help explore the effect of flare perturbations on the safety of deep neural networks.AdvFlare constructs a parameterized flare simulation model,which models the multiple attributes of the flare pattern,such as shape and color,with great simulation,on the basis of which this paper addresses the problems of adversarial perturbation distortion and convergence difficulties through strategies such as parameter space constraints,random initialization,and stepwise optimization.Experimental results indicate that AdvFlare can induce misclassification in deep neural networks with a significantly higher success rate compared to existing methods,while also offering a superior visual perturbation quality and stability.Furthermore,it is discovered that adversarial training using AdvFlare can markedly enhance the robustness of deep neural networks,in both the digital and physical world,providing valuable insights for improving model robustness in public transportation contexts.

In order to solve the inherent defects of a traditional network framework such as limited coverage and low system capacity,by considering the security problems of cooperative communication with untrusted user relay and rate segmentation,the weighted sum of the average confidentiality rate of remote users and the maximum reachable rate of public information is constructed.This paper combines Rate-Splitting Multiple Access(RSMA) technology with cooperative relaying technology for the first time,and considers the risk of public information leakage caused by the characteristics of common message broadcasting.Due to the coupling of variables and nonlinear constraints in this problem,it is non-convex and can not be solved directly.Therefore,by considering the constraints of the power budget of the base station and the relay equipment,the precoding matrix,public information segmentation and deal-to-device transmission power are jointly optimized.The Successive Convex Approximation(SCA) method is adopted to introduce slack variables and linearize the non-convex constraints.The non-convex problem is converted into a convex problem that is easy to solve,and an iterative optimization algorithm for full-duplex cooperative rate segmentation is designed so as to avoid the waste of time and resources.Simulation results prove the superiority of the proposed scheme.Compared with the existing half-duplex rate split multiple access scheme,it has a better convergence and improves the real-time hardware configuration.At the same time,compared with the non-cooperative rate split multiple access scheme,it can achieve a higher secure transmission rate and provide practical security for remote users.

With the emergence of a large number of services in wireless communication scenarios,wireless communication is affected by various interference factors and the electromagnetic environment has become more and more complex.To solve this problem and improve the reliability of a communication system,we study the burst pulse interference environment.First,we establish a composite Poisson distribution based mixed channel of the additive white gaussian noise channel and burst erasure channel for burst pulse interference environment.Then,we study nonbinary polar codes with the multiplicative repetition spread spectrum scheme for burst pulse interference environment and propose a spread spectrum sequence selection method based on the Euclidean distance.Simulation results show that the block error rate performance and the error floor performance of the nonbinary polar codes outperform than those of binary polar codes.In addition,the proposed nonbinary polar codes with the multiplicative repetition spread spectrum scheme are better than those with the Hadamard spread spectrum scheme.

The landing task is characterized by time pressure and a complex operational process,making it crucial to accurately assess pilots' mental fatigue to enhance landing safety.To address the issue of evaluating pilots' mental fatigue during landing tasks,a series of simulated landing experiments of varying difficulties are conducted,with EEG signals from nine participants over a nine-day period collected.A cross-subject mental fatigue assessment model based on SSENet is developed for the landing scenario.To capture spatial information coupling and channel feature information in the EEG signals,an SEConv module is designed within the model,targeting the spatial characteristics of the EEG signals and cross-subject training methods.The results show significant differences in the level of mental fatigue among participants during the various difficulty levels of landing tasks(p<0.001).The model achieves a maximum classification accuracy of 95.55% during five-fold cross-validation,with an average classification accuracy of 93.00%.Ablation experiments verify the effectiveness of each module,with a classification accuracy improvement of approximately 4% compared to the classical EEG signal training model EEGNet.The SSENet demonstrates promising results in the cross-subject mental fatigue assessment task,offering new strategies for enhancing the research on landing safety.

Due to the small size of Unmanned Aerial Vehicles(UAVs),complex airspace background,and easy confusion with sky objects such as birds,the existing object detection models lack sufficient accuracy.Although increasing the model size can improve the detection accuracy to a certain extent,it also reduces the inferring speed and significantly increases the number of parameters and computational complexity of the model.In addition,the lack of datasets which are suitable for small UAV object detection makes it challenging to provide adequate support for designing effective models.To address the aforementioned deficiencies,this paper first constructs a dataset from existing open-source datasets using a target-area-compression based small object sample enhancement method,which can be utilized in small UAV object detection tasks.Then,we design a lightweight and high-accuracy network model called YOLO-LADC,based on the YOLOv8.This model incorporates a novel downsampling convolution structure that reduces the number of model parameters and computations while enhancing the detection accuracy.Moreover,we add a small object detection branch to the neck network of the YOLO-LADC to achieve the YOLO-LADCS,which is better suited for small UAV object detection tasks.Comparative experiments show that the YOLO-LADCS is able to improve the average accuracy of a small object by 1.1% with a 14% reduction in the number of parameters compared to the YOLOv8n(a lightweight version of the YOLOv8).

Facial animation technology aims to dynamically drive static facial images using source data such as audio or video to produce realistic animation effects.The development of deep learning technology has greatly promoted the progress of facial animation technology.This deep learning technology can learn and capture facial features and movement patterns,achieving realistic and personalized facial animation through an automated driving process.Currently,there are numerous research achievements in the field of facial animation based on deep learning.However,existing reviews focus mostly on specific technologies or single-modality driving sources.This paper systematically reviews the facial animation driving technology based on deep learning,summarizing the research status according to the process of audio and video driving facial animation.First,it introduces the common process of extracting facial features from input source data.Second,it deeply analyzes the key technologies of feature extraction and animate generation,and compares the advantages and disadvantages of different deep learning network architectures in each step.Finally,it summarizes the animation generation methods under different architectures and compares their similarities and differences.In addition,this paper also lists the commonly used datasets and evaluation metrics for facial animation,summarizes the existing challenges in the field,further elaborates on the development trends of future work,and makes some prospects,aiming to provide researchers with a more comprehensive perspective on the application of deep learning in the field of facial animation.

To address the issues of excessive loss of vascular features at the encoding end,poor segmentation ability of vascular regions in lesion areas,and insufficient extraction of global contextual information in existing algorithms,this paper proposes a dynamic feature weighting and graph reasoning Transformer retinal vessel segmentation algorithm.First,an adaptive weighting encoding end is designed to alleviate the problem of vascular loss caused by continuous convolution and downsampling,thus enhancing vascular texture features.Second,a graph reasoning Transformer module is constructed to simultaneously extract pixel-level vascular features and relationships between nodes,thereby effectively capturing global and local information in image data.The final construction of the dynamic feature enhancement module at the decoder side and the encoder-decoder base effectively improves the ability to segment vascular lesions.Experimental results on DRIVE,CHASE-DB1 and STARE datasets show that the proposed algorithm exhibits a superior segmentation performance and generalization ability with only 0.91M model parameters,with accuracies of 97.01%,97.37%,and 97.42%,sensitivities of 82.51%,84.47%,and 81.21%,and AUC-ROC of 98.74%,98.83%,and 98.94%,respectively,showing a certain clinical application value in the diagnosis of ophthalmic diseases.

In the field of Synthetic Aperture Radar(SAR) Automatic Target Recognition(ATR),the availability of high-quality training datasets is often severely limited.Existing SAR image generation methods based on Generative Adversarial Networks(GANs) suffer from training instability and low-quality outputs.To address these challenges,we propose the Attentional Residual Wasserstein Conditional Generative Adversarial Network(ARWCGAN) for generating high-quality multi-category SAR images.ARWCGAN features attentional residual layers to enhance SAR image feature extraction,thus improving the detail and texture of the generated images.It also utilizes a combined WGAN-GP(Wasserstein Generative Adversarial Network with Gradient Penalty) loss function and classification loss function to improve the training stability and generated image diversity.We conducted generation experiments on the MSTAR dataset and evaluated the generated images from three perspectives:qualitative visual inspection,quantitative quality assessment,and contribution to the ATR model performance.Experimental results demonstrate that ARWCGAN is capable of generating high-quality images,significantly enhancing the recognition accuracy of ATR models.

Aiming at the problem of low convergence efficiency and high algorithmic complexity of the path planning model of the Unmanned Aerial Vehicle(UAV) due to multiple factors such as wind conditions and obstacles in a multi-constraint unknown environment,we propose a path planning strategy based on progressive reinforcement learning(Progressive Deep Reinforcement Q-learning Network,PR-DQN).The algorithm considers the class-teaching training and learning method,and by constructing feature-differentiated scenarios and dynamically adjusting the UAV training scenarios during the model training process,it solves the learning difficulties caused by the model facing the complex task too early,avoids the model falling into the local optimum,and improves the model learning efficiency.In addition,the algorithm comprehensively considers the impact of multiple constraints such as wind conditions,obstacles and energy consumption on the flight trajectory of the UAV in the unknown environment,and constrains the path selection of the UAV in flight by constructing the energy consumption,collision factor and multi-constraint reward function,which ensures that the UAV completes the path planning task as long as the safety and energy consumption are allowed.Experimental results show that the average planning success rate of the scheme proposed in the paper is approximately 5.4% higher than that of similar algorithms,and the average training overhead is lower than that of similar algorithms by approximately 11.7%,which makes the PR-DQN algorithm highly promising for application in an unknown environment where multiple types,multiple numbers of obstacles,and multivariate energy consumption coexist.

To address two key issues in existing multivariate long-term series forecasting models—namely,the inability to capture long-term dependencies using single-period scale time-domain information,and the difficulty in capturing effective multivariate dependencies—a multivariate long-term series forecasting model based on multi-scale time-frequency domain learning is proposed by utilizing multilayer perceptrons.The model first employs the Fourier transform to adaptively identify different periods of the sequence as multiple scales.Then,for each scale,the sequence is decomposed to conduct two-stage learning in both the time and frequency domains,capturing local and global temporal dependencies.Subsequently,based on the results of correlation analysis among the variables,the model adaptively constructs the variable dependencies within the multivariate time series.Finally,different aggregation methods are applied to the decomposed components of the sequence across different scales to achieve complementary integration of multi-scale information.Experiments on seven real-world datasets demonstrate that this model achieves an optimal or suboptimal performance in over 90% of tests.Compared to the linear model DLinear based on sequence decomposition,the proposed model achieves an average reduction of 11% and a maximum reduction of 49.22% in MSE,as well as an average reduction of 10% and a maximum reduction of 33.03% in MAE.Furthermore,the model enhances the forecasting accuracy while also demonstrating an advanced operational efficiency.

Pansharpening is to fuse the panchromatic(PAN) and multispectral(MS) images to produce High-Resolution Multispectral(HRMS) images,which is helpful for applications in ground object identification and land monitoring in the field of remote sensing.However,existing pansharpening methods based on multi-resolution analysis often overlook the relationship between the image gradients,leading to inaccuracies in extracting the detailed features from the source image and causing spatial distortion in the fusion results.To address these issues,this paper proposes a novel pansharpening method based on gradient recursion to optimize the injection coefficient.This method first analyzes the gradient relationship between the source and the fusion image.It constructs a recursive model between the ideal HRMS image and the source images at full-scale.Then,a gradient regression algorithm is designed to solve the injection coefficients iteratively.Finally,the injection coefficients are employed to refine the details obtained through a multi-resolution analysis,and the optimized details are injected into the MS image to generate the optimal HRMS image.The method is tested through simulation and real experiments on three datasets,including Pléiades,IKONOS,and WorldView-3.Compared to the second-best performing method,the ERGAS values increase by 3.59%,4.46%,and 2.18% in the simulation experimental results,respectively,and the QNR values also increase by 3.83% and 1.92% in real experiments on the Pléiades and IKONOS datasets.However,the QNR value achieves a second-best performance on the WorldView-3 dataset.In ablation experiments,compared to a gradient-free pansharpening method,the ERGAS values increase by 11.33%,14.08%,and 1.95% respectively.The HRMS images generated by our method effectively integrate the spectral information from MS images with the spatial information from PAN images,thereby significantly enhancing HRMS's spectral and spatial resolution,with the computational efficiency relatively fast.

The transposed convolution is widely used in DL(Deep Learning) tasks,but in an FSRCNN-s(Fast Super-Resolution Convolutional Neural Network-small) network,the transposed convolution has become the primary performance bottleneck during the inference stage,so that designing efficient transposed convolution hardware accelerators is essential.Based on the TDC(Transforming Deconvolution to Convolution) algorithm,the transposed convolution software inference process with a stride of 2 is transformed into a 4-way parallel direct convolution hardware implementation.The correctness of the algorithm and the hardware accelerators are validated for imperfect mapping scenarios.After designing the transposed convolution accelerator,the FSRCNN-s ×2 network is selected for end-to-end deployment.The trade-off between transposed convolution inference accuracy and speed is addressed through hardware-software co-design and an INT8(Integer 8-bit) quantization scheduling strategy.Experimental results demonstrate that the designed transposed convolution hardware accelerator incurs an accuracy loss of less than 0.5dB and that the inference speed is reduced to 17ms compared to the CPU baseline.Compared to other transposed convolution accelerators,the designed integer inference accelerator significantly reduces DSP(Digital Signal Processor) resource utilization and improves the DSP efficiency to 0.200 GOPS(Giga Operations Per Second)/DSP,offering a reference for the design of low-bit-width integer inference transposed convolution accelerators.

The memristor compute-in-memory(CIM) architecture,as a new technique that integrates storage and computing,can effectively address the problems of a limited data transmission rate,frequent data migration,increased transmission power consumption and delay caused by the separation of storage and computation in traditional von Neumann architecture data error correction,thereby improving the satellite-borne electronic system reliability and stability.However,existing CIM error correction techniques can only correct the single-bit data error and fail to handle continuous multi-bit error detection and correction.Thus,this paper proposes a memristor CIM-based BCH multi-bit error correction method.First,we convert the traditional encoding and decoding operations such as modulo,multiply add and forward search into matrix operations to simplify the calculation process and reduce resource overhead.Second,we construct finite field multiply-accumulate and multiply calculation units separately,and based on the operational requirements and data characteristics of each stage of the BCH algorithm,parallel processing is utilized to adaptively select the corresponding computing cores to further improve the operational efficiency.Finally,the proposed method is verified on the Calculator and MNSIM simulation platforms of Cadence.Experimental results show that the proposed method achieves efficient and stable multi-bit error correction,that the data throughput is 8.8 MHz,that the operating power consumption is less than 40mW,and that the area overhead is 3×105 um² in 65 nm.Specifically,compared to FPGA and IMPLY architectures,the computational efficiency has increased by 7× and 400×,respectively.

High reliability train positioning technology is the foundation of the train operation control system.When the train reaches a scene where the tunnel is severely obstructed,the Beidou Satellite Navigation System(BDS) will fail,leading to the accumulation of positioning errors and seriously affecting the positioning performance of the train.In order to improve the positioning performance of trains in severely obstructed tunnel scenes,this paper proposes an adaptive positioning method for trains in severely obstructed tunnel fields with Beidou failure.First,the positioning error of SINS in the scenario of Beidou failure is analyzed,and the dead reckoning DR algorithm is used to solve the problem of SINS self-calibration failure.Second,an adaptive estimation and compensation method for positioning errors in severely obstructed tunnel scenes is prsented.Based on the analysis of positioning errors,the state equation and observation equation for tunnel train combination positioning are constructed.Finally,the Unscented Kalman Filter algorithm is designed to optimize the estimation of positioning errors,achieving calibration and compensation of the continuous position information on the train,and improving the performance of tunnel positioning when the Beidou satellite signal fails.The proposed method is effectively validated through train positioning simulation experiments and actual measurement data of the Yinxi high-speed railway tunnel route.The results show that compared with the comparative methods,the proposed method can effectively improve the positioning performance of trains in severe occlusion scenarios of Beidou failure tunnels,and has a better stability.

While the wide application of face recognition technology brings great convenience to people,this technology also faces the risk of identity and other private information being leaked.Some current attacks against face recognition can obtain the privacy information on the original image by reconstructing the face image.To prevent such attacks,this paper proposes a privacy-preserving face recognition model against adversarial sample perturbation(PPFR-ASP).Specifically,our scheme employs collaborative inference in the frequency domain.First,the face image is transformed into frequency domain features,which are divided into high-frequency and low-frequency components.Then we add adversarial sample perturbation on the frequency domain features.Furthermore,a target image is prepared for each original image,and the same frequency domain feature transformation is applied to the target image.The frequency component channels of the target image are used to mask the corresponding components in the original image so as to cause the attacker to reconstruct the image as the target image,thereby concealing the true identity of the original image.Finally,extensive experiments are conducted on multiple datasets,demonstrating that the privacy protection performance of this solution for facial image data exceeds that of comparative approaches.The query accuracy and computational overhead of this solution are comparable to those of the unprotected ArcFace scheme.

To address the limitations of current privacy protection schemes,particularly the insufficient research on data perturbation and the inadequate integration between privacy measurement and privacy protection,we propose a Dynamic Balanced Privacy Model for Data Disturbance(DBPM-DD).First,based on users' privacy preferences and data quality,we design a dynamic measurement mechanism for real-time evaluation of data privacy to precisely measure the amount of private information contained in the data.Second,we propose a data reconstruction mechanism based on probability partitioning and accordingly provide a privacy measurement method for perturbed data,achieving multi-paradigm adaptation of private data.Finally,we introduce a noise scale adaptive adjustment method that adaptively adjusts the noise intensity based on feedback from privacy measurement results,ensuring user privacy while maximizing data utility,thereby achieving a dynamic balance between privacy protection and data utility.Experimental results show that under different noise scales,data sizes,and attack intensities,the model can effectively enhance the degree of privacy protection while maintaining high data utility,providing consistent and effective privacy guarantee under various conditions,and outperforming other privacy protection models.This research provides a new theoretical support for privacy protection in data perturbation technologies,possessing a significant practical application value and a wide range of applicability.

In order to address the potential issues of privacy leakage,a single point of failure,and poisoning attacks in traditional federated learning application center servers,a fair federated learning framework based on the alliance chain is proposed.Through the mutual selection of leader nodes and consensus committee nodes in each round,secure aggregation and updating of data are achieved,ensuring the decentralized and distributed characteristics of the system.Meanwhile,by leveraging the immutability of the blockchain and its resilience to single-point attacks,a client-level data quality assessment method is designed within the consensus mechanism to provide necessary quantitative metrics for multi-party training,ensure the transparency and traceability of evaluation results,and optimize the node selection process,thereby ensuring the prioritization of high-quality clients.To improve the fairness of node selection,an improved algorithm based on the Shapley value is proposed that incorporates the historical behavioral performance of clients to make contribution evaluation more flexible and accurate,thus reducing the proportion of low-quality nodes in contribution evaluation and mitigating the negative impact of low-quality data on model training.Experimental results show that the scheme significantly enhances the fairness of leader node elections and the accuracy of client marginal contribution assessments,while maintaining the model prediction accuracy.Through a dynamic node reward mechanism,the long-term fairness of the system is ensured,effectively addressing fairness issues in the alliance chain.

Network coding technology can effectively improve the network throughput.However,traditional network coding involves a high complexity in both encoding and decoding and it is difficult to adapt to the influence of dynamic factors such as environmental noise,which leads easily to decoding distortion.In recent years,researchers have introduced neural networks to optimize the network coding process,but in high-resolution image transmission,the existing neural network coding schemes have an insufficient ability to capture high-dimensional spatial information,resulting in large communication and computation overhead.To solve this problem,this paper proposes a joint source deep learning Network coding scheme that uses a two-dimensional Convolutional Neural Network(CNN) to parameter-design the encoder and decoder of each network node,which captures deep spatial structure information and reduces the computational complexity of network nodes.At the source node,the convolution layer operation is used to reduce the dimension of the transmission data and improve the data transmission rate;At the intermediate node,the data from the two sources are received and compressed by CNN coding for single channel transmission;At the destination node,the received data is decoded using a CNN to increase the dimension and restore the original image.Experimental results show that under different channel bandwidth occupancy ratios and channel noise levels,the proposed scheme shows an excellent decoding performance in peak signal-to-noise ratio and structural similarity.