Abstract:Federated learning is a kind of distributed machine learning technology to ensure that local data is not compromised when training with big data for machine learning models.However,a series of attacks shows that the adversary can steal private information from machine learning model parameters even if local data is inaccessible.Thus,many privacy threats must be mitigated,since they can arise from the intermediate model parameters transmitted between participants and the aggregator in the training phase or from the finally released aggregated model.Therefore,various privacy-preserving federated learning approaches have emerged,primarily based on cryptography and differential privacy technology.This paper surveys the privacy threats and adversary models that may appear when we publish local models and aggregated model of federated learning.Furthermore,we systematically summarize the related defense technologies and research advances.Finally,we also presents a prospect for the development trend of privacy-preserving federated learning.

Abstract:Based on the online measurements from December 25,2017 to January 16,2018,we investigated the chemical composition and sources of PM_2.5 in Yangjiang,a South China coastal city.PM_2.5 in Yangjiang mainly consisted of OM,NO^-₃,SO^2-₄,NH⁺₄ and EC,with proportions of 32.75%,25.59%,16.41%,12.37% and 4.82% in mass fraction,respectively.In two pollution events,the mass concentration of NO^-₃ increased to over 6 times of that during clean period,much higher than the increments of other components.The mass fractions of NO^-₃ in two polluted events increased to 29.38% and 30.81%,over twice of that in clean period.Source apportionment of PM_2.5 were performed by Positive Matrix Factorization,which pointed out that secondary formation was the main source/process (51.41%) for PM_2.5 in winter in Yangjiang.The secondary formation from NO_x,contributing 27.18% to the total PM_2.5 mass concentration,was the dominating secondary source.Vehicle exhausts were the biggest primary source (15.11%).In the pollution events,the contribution of the secondary formation from NO_x sharply increased from 11.85% to 33.15% and 36.96%,which should be considered as the dominating source of PM_2.5 pollution in winter of Yangjiang.This study revealed that,as a coastal city,Yangjiang exhibited similar characteristics to those big cities and mega-cities in terms of the PM_2.5 sources,i.e.,dominated by the secondary formation.It suggested that more effort for pollution control should be put into the reduction of nitrates as well as their precursor NO_x,meanwhile a persistent control of vehicles in Yangjiang city is recommended.

Abstract:Here,a CNN-Attention-BP model is proposed for precipitation forecast based on analysis of the characteristics of precipitation statistical prediction models,then empirical analyses are made on the summer rainfall in Changchun,Baicheng and Yanji stations of Jilin province for the period of 1961-2020.First,a Convolution Neural Network (CNN) is used to study the characteristics of precipitation,air pressure,wind speed,air temperature and relative humidity.Second,the attention mechanism is used to determine the weight of meteorological factors for precipitation forecast.Then a BP neural network is applied to predict the precipitation probability.And the performance of the proposed CNN-Attention-BP model is evaluated by accuracy,cross-entropy loss function and F1-score,which is compared with that of the support vector machine,multi-layer perceptron and convolution neural network model.The results show that the CNN-Attention-BP model is characterized by autonomous learning and paying more attention to significant information,as well as the improved forecasting performance in summer precipitation occurrence for Jilin province.Meanwhile,the proposed model would perform better with more balanced sample and precipitation frequency closer to 0.5,when accuracy would reach up to 88.4％.Compared with single models,the CNN-Attention-BP improves the forecast accuracy by 17 percentage points.

Abstract:China strives to achieve carbon emission peaking by 2030,and the exhaust from vehicles on road are an important source of greenhouse gas (GHG) emissions.Due to the upgrading of vehicle emission limits and fuel standards,it is particularly necessary to study the GHG emission factors of currently in-use vehicles.In this study,14 in-use light- and heavy-duty vehicles in China were tested for CO₂ (carbon dioxide) Emission Factors (EFs) under the cold start procedure of WLTC (Worldwide harmonized Light-duty Test Cycle) and C-WTVC (China-World Transient Vehicle Cycle) operating conditions respectively using chassis dynamometer,and the corresponding fuel consumptions as well as influencing factors were studied.The results show that the CO₂ EFs of vehicles are affected by displacement,hot／cold start,fuel and driving condition.Vehicles have the highest fuel consumption under cold start conditions on urban roads,resulting in higher CO₂ EFs,which are 26.6％-199.7％ and 8.3％-35.5％ higher than those under full working cold start and hot start conditions on urban roads,respectively.Meanwhile,the higher fuel consumption of high-displacement heavy-duty diesel trucks in urban traffic condition leads to a significant increase in CO₂ EFs,so banning high-displacement heavy-duty diesel trucks from entering urban areas can effectively control CO₂ emissions.Using LPG (Liquefied Petroleum Gas) as a fuel substitute can reduce CO₂ emissions from vehicles.LPG buses and taxis would reduce CO₂ EFs by 37.2％ and 12.1％ on urban roads,and by 51.8％ and 20.3％ on highway roads.The current WLTC conditions,which are more in line with actual road conditions,still underestimate the CO₂ EFs and fuel consumption of light-duty vehicles on Chinese roads by 31％-46％ and 17.7％-26.8％,therefore to obtain more accurate vehicle emission data,we must accelerate the localization of vehicle test cycles.

Abstract:With the rapid urbanization in China,the evaluation of ecological benefits of urban renewal based on sustainable development has been receiving increasing attention.Based on the practice of urban renewal in China,we constructed a Structural Equation Model (SEM) of citizen satisfaction,which uses six indicators including air quality,heat island effect,vegetation greenery,resource conservation,urban transportation and urban open space to measure the ecological benefits of urban renewal,then carried out a questionnaire to validate the SEM in evaluating citizens' satisfaction with the ecological benefits of Xi'an happiness forest belt urban renewal project.The study shows that most of the citizens' expectation can be met as the achievement rates of ecological expectation and overall expectation of citizens for Xi'an happiness forest belt urban renewal project were 94.3％ and 97.4％,respectively.Citizen expectation has a positive impact on citizen perception of ecological benefits,and the path coefficient is 0.381;citizen perception of ecological benefits has a significant positive impact on citizen satisfaction,and the path coefficient is 0.903;and citizen satisfaction has a significant positive impact on citizen trust,and the path coefficient is 0.955.It can be concluded that the proposed SEM of citizen satisfaction can quantitatively evaluate the ecological benefit of urban renewal.Furthermore,the quality of ecological services should be focused in future urban renewal to improve residents' ecological environment experience,and efforts should be made to promote the coordinated consideration of ecological,economic,social and cultural benefits of urban renewal,as well as the collaborative incorporation of multiple stakeholders such as government agencies,local residents and property development firms in urban renewal.

Abstract:In order to improve the management level and processing efficiency of government affairs,we propose a multi-algorithm-combined model to evaluate the replies to public consultations in government service.First,we define the five aspects of evaluation including the reply length,similarity,completeness,interpretability and timeliness,and evaluate the text from four perspectives of content,format,reasonableness,and timeliness.Second,we analyze the types of replies by regression analysis.Then,grade the replies by clustering algorithms of K-means,DBSCAN,Meanshift,and HC clustering.Comparison shows that K-means clustering outperforms the other three algorithms in clustering performance,thus it is combined with regression analysis to evaluate the replies.Finally,the replies to public consultations are graded into 6 categories.The proposed model integrates machine learning including data mining and data analysis into "smart government affairs",and provides a quantitative analysis tool to evaluate the performance of government affairs management.

Abstract:Findings confirm that poor people may be more vulnerable to climate change.The huge socioeconomic costs of droughts make themselves a crucial target for impact assessments of climate change scenarios.Based on outputs from fifteen CMIP6 climate models under the four latest SSP-RCP scenarios (SSP1-2.6,SSP2-4.5,SSP3-7.0,and SSP5-8.5),drought characteristics (frequency,intensity and duration) at 1.5 and 2 ℃ warming levels were analyzed over the poverty-stricken areas of China.Results show that:(1) Relative to the baseline period (1995-2014),the annual mean temperature in poverty-stricken areas will increase by 1.1 and 1.8 ℃ at 1.5 and 2 ℃ global warming,respectively,which is faster than the global warming rate.The annual precipitation will also increase significantly,especially in north-west of poverty-stricken areas;(2) Relative to the baseline period,drought intensity will increase from slightly dryness to moderate dryness at the global warming of 1.5 ℃.However,most regions will see a decrease trend in both drought frequency (74％ of whole region) and duration (61％ of whole region).(3) At the global warming level of 2 ℃,drought intensity will still remain a moderate level,while drought frequency may keep decreasing.In some areas,the frequency is possible to decrease by 0.6 times annually.The drought duration tends to increase in the south and decrease in the north,which is expected to decrease by 1.3 months in some places (55％ of whole region).(4) When facing the additional global warming of 0.5 ℃,there are spatial similarities in the change of drought characteristics.The frequency and duration will both increase in southern region and decrease in northern region,while the intensity will increase in most regions.Our study suggests that limiting anthropogenic warming to 1.5 ℃,as aspired by the Paris Climate Agreement,may have benefits for future drought risk alleviation over the poverty-stricken areas of China.

Abstract:Sliding Mode Variable Structure Control (SMC) has been widely used in typical nonlinear multi-robot control systems due to its quick response,strong adaptability,and easy engineering implementation.In this article,we introduce the SMC's development and principle,as well as its application in the field of multi-robot cooperative formation.In view of the SMC's defects such as chattering,which still hampered its application,we made an in-depth survey on researches of fusion of SMC with advanced intelligent control approaches such as neural networks,fuzzy logic,robust adaptive,and comparatively analyzed these new formation control strategies in aspects of response time,robustness against disturbances,and steady-state performance.Finally,the future research directions in SMC are prospected.

Abstract:Brain tumor is one of the most fatal cancers in the world.Its image classification has become the hot spot due to the diverse characteristics of brain tumors.In recent years,Deep Neural Networks (DNN) are commonly used for medical image classification,but the problem of gradient vanishing and over fitting will appear with the increase of depth,while the Residual Network (ResNet) can solve this problem by introducing identity mapping.Therefore,this paper proposes an MDM-ResNet approach for brain tumor classification,which is composed of multi-size convolution kernel module,dual-channel pooling layer and multi-depth fusion residual block.The experiment was carried out on Figshare dataset,the image was preprocessed by data enhancement operation,and the performance of network was evaluated based on five-fold cross validation.The experimental results prove that the MDM-ResNet approach can effectively classify meningioma,glioma and pituitary tumor.

Abstract:Moving obstacles,unlike stationary ones,cannot be located or avoided by traditional navigation technologies.To address this,a collision avoidance navigation path planning strategy for blind people based on deep learning is proposed.First,a speech recognition system is used to collect speech signal and sort out the speech feature parameters,which is then analyzed to obtain the speech sequence input thus recognize the destination.Second,an obstacle detection model is constructed to detect the edge features as well as moving directions and velocities of the obstacles on the path to destination.Then a convolutional neural network of deep learning is exploited to plan the optimal path with collision avoidance.Finally,experiments are conducted and the results show that the radial velocity of the moving obstacles detected by this model is consistent with actual conditions,specifically,when the actual speed is 33.6 cm／s,the detected speed error is in the range of 0.2-0.4 cm／s,and the accuracy of obstacle avoidance reaches 96.5％ when the test time lasts 50 min.It can be concluded that the proposed strategy can realize the optimal path planning and navigation with collision avoidance for the blind people.

Abstract:In recent years,the transmission line faults caused by birds in Guangdong power grid have been increasing gradually,which have become one of the main hidden dangers for power grid security.How to reduce the bird damage has become a new topic of transmission line operation and maintenance.For the widely distributed transmission lines,it is difficult to effectively prevent bird damage by driving birds approach.Here,based on the operation and maintenance data related with bird damage on Guangdong power grid from 2015 to 2019,we analyzed the geographical characteristics,pole tower structures and seasons,then established a model to analyze the transmission line faults caused by bird damage.First,the influence of geographical characteristics,pole tower structure and season on bird caused fault is analyzed.Then,a Mask R-CNN neural network is trained to extract the geographical characteristics around the pole tower.Finally,a bird damage fault model based on Bayesian linear regression is established,and the accuracy of the model is evaluated by the correlation coefficient of R².The experimental results show that the model has high accuracy and reliability.

Abstract:Multi-beam satellite communication system can provide mobile users with global communication services.However,the on-board resource utilization is low due to the limited resources as well as the long distance between satellites and the earth,and the resulted time delay in service transmission.To address this,we propose a new downlink resource allocation algorithm for a multi-beam satellite communication system with beam interference,which is designed under two constraints of strict service delay and limited maximum interference power.Meanwhile,the influence of inter-beam interference on the system capacity is considered,and the Lagrange duality theory and sub-gradient descent method are used to obtain the optimal resource allocation strategy in order to maximize the service capacity of spot beam.The simulation results show that under the same noise power spectral density channel,the proposed algorithm meets the time delay requirement for each spot beam,has good fairness,improves the system in total transmission rate,and reduces the difference between the spot beam service demand and the allocated communication resources,therefore can provide maximized communication service to users;while under the condition of different noise power spectral densities,the proposed algorithm still maintains good fairness.

Abstract:The fixed-time containment control is investigated for multi-agent systems with inherent nonlinear dynamics.Assuming that not all followers can directly receive information from the leaders and the communication topology between the followers is directed,a distributed control law is designed to solve the fixed-time containment control problem.By using algebraic graph theory,matrix theory and fixed-time stability theory,the conditions on the communication topology are derived for realization of fixed-time containment control.Finally,a simulation example is given to verify the correctness of the theory.

Abstract:The demand for higher temporal and spatial resolution of atmospheric detection has brought about rapid increase of atmospheric hyperspectral remote sensing data;however,traditional methods have low efficiency for hyperspectral data processing.Here,we summarize the application examples of using GPU to accelerate the processing of hyperspectral remote sensing data,among which we focus on the parallel computing of the Fourier analysis of hyperspectral data based on the CPU-GPU heterogeneous mode.Then the CPU-GPU heterogeneous mode based parallel computing is implemented and then compared with traditional CPU-based computing.The results show that the data processing speed is increased by about 90 times while ensuring the data accuracy as well,thus verify the effectiveness of GPU to accelerate the processing of atmospheric remote sensing hyperspectral data.