Recently, semantic segmentation has been widely applied in various realistic scenarios. Many semantic segmentation backbone networks use various forms of dense connection to improve the efficiency of gradient propagation in the network. They achieve excellent segmentation accuracy but lack inference speed. Therefore, we propose a backbone network SCDNet with a dual path structure and higher speed and accuracy. Firstly, we propose a split connection structure, which is a streamlined lightweight backbone with a parallel structure to increase inference speed. Secondly, we introduce a flexible dilated convolution using different dilation rates so that the network can have richer receptive fields to perceive objects. Then, we propose a three-level hierarchical module to effectively balance the feature maps with multiple resolutions. Finally, a refined flexible and lightweight decoder is utilized. Our work achieves a trade-off of accuracy and speed on the Cityscapes and Camvid datasets. Specifically, we obtain a 36% improvement in FPS and a 0.7% improvement in mIoU on the Cityscapes test set.
Privacy policy statements are an essential approach to self-regulation by website operators in the area of personal privacy protection. However, these policies are often lengthy and difficult to understand, with users appearing to actually read the privacy policy in only a few cases. To address these obstacles, we propose a framework, Privacy Policy Analysis Framework for Automatic Annotation and User Interaction (PPAI) that stores, classifies, and categorizes queries on natural language privacy policies. At the core of PPAI is a privacy-centric language model that consists of a smaller fine-grained dataset of privacy policies and a new hierarchy of neural network classifiers that take into account privacy practices with high-level aspects and finegrained details. Our experimental results show that the eight readability metrics of the dataset exhibit a strong correlation. Furthermore, PPAI’s neural network classifier achieves an accuracy of 0.78 in the multi-classification task. The robustness experiments reached higher accuracy than the baseline and remained robust even with a small amount of labeled data.
Visual sensors on autonomous vehicles are vulnerable to adverse weather, which seriously reduces the performance of semantic segmentation task and threats people’s safety. Therefore, segmentation models need to carry out extensive training on a large amount of adverse weather data which is difficult and expensive to acquire to improve their robustness. To solve this problem, researchers have proposed domain generalization methods that do not need target domain (such as adverse weather) data to adapt during training. However, most of them focus on the synthetic-to-real problem which is caused by the difference in texture between real and virtual images. To address these challenges, we analyze the formation mechanism behind adverse weather, extract two kinds of weather cues and establish the relationship between them and adverse weather. On this basis, we propose CSCR, a domain randomization framework for simulating adverse weather. Specifically, CSCR includes the Common Cue Randomization (CCR) module which simulates adverse illumination style and the Specific Cue Randomization (SCR) module which randomizes weather cues that occur in specific adverse weather. We conduct extensive experiments from fair weather to fog, night, rain and snow on driving datasets. Compared with the source model, CSCR increases by more than 9% points in mIoU on average which even exceeds some domain adaptation methods while still keeping the memory of fair weather. The CSCR framework can be easily applied to existing segmentation models and significantly improve their generalization ability.
Medical Recommendation Service (MRS) can provide simple diagnostics and recommend a suitable doctor for patients when they input their disease symptoms by using AI and big data. However, people may be unwilling to share medical data due to data privacy issues and a lack of incentives, which makes it hard for MRS to continue to develop. Until now, there are no efficient solutions. To solve this, we propose a blockchain-based anonymous reward mechanism for medical recommendation (BARM) in smart healthcare. The proposed scheme uses accumulator, commitment, and Signatures of Knowledge (SoK) to provide anonymous rewards for users to mobilize more users to share data as much as possible. And we design a smart contract that can provide submit, query, and revoke functions to realize anonymous authentication of patients. The formal security analysis shows that the proposed scheme meets the required security requirements. Detailed performance analysis results show that the computation cost and communication cost of the scheme is feasible, which can be applied to the medical recommendation scenario of anonymous protection and promote the development of smart healthcare.
The fMRI functional imaging technology has become an important means to detect brain diseases. It mainly diagnoses whether there are lesions in the patient’s brain region by monitoring the changes of the blood oxygen level (BOLD) in the brain region. Existing brain disease classification techniques are mainly achieved through network construction, feature extraction and classification model construction. But in the existing methods, almost no one uses the time attribute. In this paper, we propose a brain disease diagnosis framework with temporal attribute and feature fusion, use temporal and BOLD signal’s changes in different time periods to construct a brain connectivity network. We extract frequent sequences and sequences’ connection rules in the subjects’ brain connection network, and finally use pseudo-binary matrix to fuse the two features and finish classification model construction. Experiment turns out that we achieve better accuracy than other schemes.
The distributed nature of distributed learning renders the learning process susceptible to model poisoning attacks. Most existing countermeasures are designed based on a presumed attack model, and can only perform under the presumed attack model. However, in reality a distributed learning system typically does not have the luxury of knowing the attack model it is going to be actually facing in its operation when the learning system is deployed, thus constituting a zero-day vulnerability of the system that has been largely overlooked so far. In this paper, we study the attack-model-agnostic defense mechanisms for distributed learning, which are capable of countering a wide-spectrum of model poisoning attacks without relying on assumptions of the specific attack model, and hence alleviating the zero-day vulnerability of the system. Extensive experiments are performed to verify the effectiveness of the proposed defense.
Analyzing the key factors affecting the delay of hierarchical federated learning and reducing the generation of delay is an important issue to be addressed. In this paper, we analyze the hierarchical federated learning network in the case of simultaneous access of mobile devices and model the arrival process and service process of data streams satisfying Poisson distribution. This paper analyzes the delay bound of the hierarchical federated learning network under a round of global updates using stochastic network calculus. We model a more realistic service model by considering the service rate variation of edge servers due to channel fading and other factors when analyzing the delay bound of the wireless access network. Finally, we analyze the parameters affecting the end-to-end delay performance of the hierarchical federated learning network in numerical analysis. The factors that affect the latency are the number of mobile nodes, the number of edge nodes, and the arrival rate of the data flow.
Nowadays, with the significant increasement of the deep learning training (DLT) task workload in GPU clusters, the number and the scale of GPU clusters grow rapidly. A crucial question is how to efficiently schedule DLT tasks with limited cluster resources. Existing GPU schedulers do not fully consider the connection between users and clusters, and few methods optimize the GPU allocation of DLT tasks. In this study, we propose a scheduling framework for GPU clusters, which improves performance and reduces energy consumption of clusters. We first analyze the relationship between the characteristics of performance and energy consumption and the task configurations for DLT tasks. Then, we propose a prediction method to predict the completion time and energy consumption of DLT tasks. To make better use of cluster resources, based on the prediction model, we propose GAS, which adopts the GPU Allocation Strategy by specifying the parallelism for DLT tasks. Compared to FIFO and SJF schedulers, GAS reduces the makespan by 19.6%-19.8%, reduces the average queueing time by 84.4%-93.9% and reduces the energy consumption by 22.2%22.5%. For users, GAS also reduces the cost of users by 21.3%21.6%. The large-scale simulation experiment further illustrates the effectiveness and scalability of GAS.
Deep flow inspection (DFI) has been increasingly important for network management and security protection. Network flows have rich correlations due to multiparty communications such as video conferences and DDoS attacks, and furthermore, rich features are vital for artificial intelligence-driven DFI models and algorithms. However, existing approaches do not correlate network flows by just providing simple counters for key-valued network flow records due to the limitations of the measurement models.We present Mercury, a broad and deep flow inspection system, which continuously collects a wide range of features for large scales of correlated network flows and supports expressive query interfaces for diverse network measurement tasks. Mercury organizes rich and dynamic features of the network traffic based on a streaming graph model, which defines rich states for each network flow to the edge of the graph and incrementally updates these states in high-performance pipelines. Mercury raises significant challenges for storing and updating states in the streaming graph model due to the fast arrival rate and the in-memory storage limits. We present a nested two-dimensional data store, called cuckoo matrix, to provide a two-dimensional index for the streaming graph model. The cuckoo matrix optimizes the state-update throughput and reduces the memory cost through the multiple dimensions of the cuckoo tables. We speed up the ingestion process based on a threaded lock-free butter and update the feature statistics based on a unified streaming framework.Extensive evaluation with real-world data sets demonstrates that Mercury achieves a significantly better trade-off between throughput and memory consumption than state-of-the-art methods. Mercury supports both traditional machine learning methods and accurate graph neural networks for traffic classification through extensive and efficient graph query APIs.
A Convolutional Neural Network (CNN) is a class of artificial neural networks that have shown advantages in visual imagery analysis. Implementing a high-performance CNN with limited computation and memory resources is desirable for commercial use and green computation purpose. Tremendous FPGA-based accelerators are designed to achieve this goal. However, existing accelerators are mostly designed for high-performance FPGAs and it is difficult to achieve the desired performance when deploying them to resource-limited ones. Furthermore, in different application scenarios, different CNN and FPGA boards are usually adopted. Existing accelerators require complex configurations to adapt to a different scenario where the CNN input size or FPGA resources are changed. To deal with these problems, we propose a reusable convolutional accelerator for CNN. We utilize the particle swarm optimization (PSO) method to mathematically model the FPGA resources and find the optimal parameters to realize reusability. In the experiment, we implemented four CNN networks on two FPGA boards. The experiment result shows that the proposed accelerator can achieve 30$\sim$40 GOP/s even with resourcelimited FPGA.
The use of collected data is a valuable source for analysis that benefits both medical research and practice. Information privacy is considered a significant challenge that hinders using such information for research purposes. In terms of research, releasing patients’ information for research purposes may lead to privacy breaches for patients in various cases. Individual patients may not wish to be identifiable when using information about their health for research. This work proposes a utility-aware data anonymization model for sharing patients’ health information for research purposes in a privacy-preserving manner. The proposed model is interactive and involves a number of operations that are performed on patients’ information before releasing it for research purposes according to certain requirements specified by the data user (researcher).
Vehicles participating in routing operations in vehicular ad hoc networks (VANETs) are required to periodically broadcast beacon messages. However, frequently broadcasting the vehicular information can pose a serious threat upon users’ location privacy. While the exploitation of collective anonymity scheme on pseudonym change is an effective approach to protecting the location privacy of a target vehicle in VANETs, the exposure impact of the collective anonymity on privacy risk of a target vehicle has not been investigated yet. In this paper, we focus on improving anonymity level for a target vehicle in a collective anonymity scenario while achieving a high level of privacy exposure risk management. To achieve this, we first propose a pseudonym change scheme based on an assessment of the privacy exposure risk, aiming to manage and mitigate the privacy exposure risk of the target vehicle brought by exposed neighboring vehicles. Then we formulate an analytical model according to our proposed scheme in order to quantify the anonymity level of the target vehicle simultaneously updating pseudonyms with its neighbors. Simulation results show that our proposed scheme succeeds in promoting the size of anonymity set of the target vehicle with the growth in the number of its exposed neighboring vehicles. Meanwhile, with different setting parameters, our scheme retains effective performance compared to the random pseudonym change scheme.
With the development of Internet and AI related technologies, task-based dialogue system has become a research hot-spot. This paper uses the common characteristics of dialogue management and reinforcement learning in task-based dialogue system to introduce deep reinforcement learning into the process of building dialogue management model, and proposes a dialogue management model based on A3C algorithm to complete the learning task of dialogue strategy. The model is based on Actor-Critical architecture, which can dynamically improve the learning result of dialogue strategies through internal evaluation mechanism, and improve the prediction accuracy of dialogue management model. At the same time, the model is trained in a multi-threaded way, which can execute multiple sub-nets at the same time, and can effectively improve the training efficiency of the model.
Network Digital Twin is a model that emulates the behavior of a real communication network and is commonly used for prototyping, testing, and network optimization. Utilizing the inherent graph structure in the communication network, Graph Neural Networks (GNNs) have been successful in modeling digital twins of small-scale network. However, they face the challenge of scaling up the GNN model to make predictions on large-scale networks. In this paper, we build a scalable network digital twin capable of predicting the per-path mean delay from traffic and routing configuration on larger topology networks. In this paper, we investigate the effectiveness of our proposed solutions when modeled using different neural network architectures and combined through various ensemble methods. We explore the performance of GNN models implemented using traditional Artificial Neural Networks and the novel Spiking Neural Network (SNN) techniques. We study the performance of different ensemble weighing techniques such as Averaging, Grid Search, and a Reinforcement Learning(RL)-based weighing technique. Through our experiments, we intend to understand the challenges and advantages of each architecture and ensemble technique in building a scalable network digital twin using GNN.
In recent years, more and more students are willing to choose courses on online learning platforms, such as MOOC. However, the online learning platform cannot capture the learning status of students, and cannot give real-time feedback when students’ attention is out of learning. The lack of feedback mechanisms mainly leads to a staggering dropout rate of MOOC. Therefore, engagement detection is very important for online learning platforms. Nowadays, many researchers have focused on engagement detection which are based on pictures. This method of selecting pictures is through random selection or repeated selection after short interval of time. However, both of these methods may lead to the problem of inaccurate samples or heavy identification workload. In this paper, we propose to introduce Key Frames that can more accurately represent the engagement level of the online learning personal video in the sample selection stage. Experimental results show that our proposed method has better results.
With the rapid development of big data technology, how to mine useful association rules from a large amount of time series data is an important task in the field of water resources. Mining potential association rules between time series modalities has become a hot topic, which will provide effective help for water governance to predict future hydrological trends. The traditional method of mining hydrological data has two drawbacks: Firstly, the traditional method focuses more on the similarity of modalities in the process of modal mining, so it is easy to mine similar modalities and instances in large-scale data, thus generating redundancy; secondly, the traditional method ignores asynchronous rules in the mining of temporal association rules, and this will lose many valuable rules. In this paper, we propose a three-step pruning method for modal mining, which improves on the MASS algorithm. We propose a constraint-based association rule mining algorithm that achieves asynchronous rule mining by introducing a minimum overlap threshold and a maximum time interval. Finally, we conducted experiments on the hydrological dataset. Compared with other methods, the experimental results show that the method proposed in this paper is effective.
Digital Twin (DT) technology, which extract test scenarios from real-world driving data, is considered as an effective solution for ICVs (Intelligent and connect vehicle) on-road test. However, due to the time-varying feature of environment, validity of current ICV oriented DT test scheme is hardly to verify. As a traditional test issue, In-chamber test is considered as an effective approach for vehicles’ regulation mandatory tests, such as EMC test. To promote standardization progress of ICV pre-market test, in this paper, an in-chamber C-V2X application test scheme is proposed, while corresponding test procedure is given. Moreover, two key technologies, test case design and GNSS control, are discussed. A cross validation progress is done to verify the validity of proposed scheme, and the effectiveness of test parameter setting, which is extracted from on-road acquisition data.
Hospital emergency departments (EDs) are crucial medical facilities providing emergency healthcare. Understanding and measuring the patient flow in EDs plays a key role in maximizing the utility of scarce hospital resources. Existing approaches either require manual reporting by staff or collect sensitive data about patients via camera or hospital registration system, causing potential privacy breaches. Against this background, we propose a data-driven modeling strategy of ED crowdedness leveraging multi-source urban crowdsensing data, which automatically provides fine-grained and timely information about the crowdedness of EDs. Specifically, our model can not only accurately extract the emergency visit demand from noisy human mobility data with minimum expert knowledge using active learning and co-training techniques, but also estimate ED crowdedness by modeling the emergency service process by integrating three queueing models for general practice, internal medicine, and surgical clinics, respectively. We evaluate our method leveraging large-scale taxi and hire vehicle trajectory datasets and hospital information system datasets from the government open data portal. Results show that our approach effectively extracts the emergency visit demand and models the emergency service process to assess the crowdedness of EDs. We have deployed the system in Xiamen City by collaborating with the municipal government to provide services for citizens and health providers.
With the development of globalization, business collaboration and exchanges between companies become increasingly frequent. Data communication between two or more parties using certificate authentication is a frequently employed security assurance approach that creates a substantial quantity of certificate data traffic. In the process of utilizing conventional certificates, it is possible to leak sensitive information of the corporation. For example, certificate data traffic in the smart grid may expose important nodes and current security management level of a certain grid-related network. Therefore, a solution is required to evaluate the data of the grid-related network at the certificate level in a secure and efficient manner. This study provides a certificate-based enterprise privacy analysis and anomaly detection scheme. Using SSL/TLS and X.509 as standards, the scheme extracts valid certificate information by collecting certificate data traffic between enterprise networks and analyzing the certificate data. The scheme monitors and analyzes the flow of certificate data for the grid network, identifies suspicious anomalous behaviors and invalid certificates, and warns of risky behavior. By monitoring certificate data flow, it is feasible to improve the enterprise authentication framework and business communication method, prevent the leaking of enterprise privacy data.
The invention of personal computer together with its rapid deployment around the world are among the most important achievements that have changed lives of people in this planet. In modern society, people are becoming more dependent on computer devices. However, it’s difficult to bring computer hardware with us at any time and in any place, which is known as the desktop-go-with-person (DGWP) problem. Regarding it’s easy to find a nearby computer and bring tiny non-volatile memory devices with us, putting the entire desktop system software into pocket storage and loading it on any available hardware become an attractive approach. However, current computer software are tightly bound to hardware. Live USB systems are quite popular for trial run purpose; however, they are read-only system and limited to certain OSes. Also, the creating of live USB system from native installed desktop system is dependent to OS. In this paper, we proposed a new generic UEFI-based secure system-in-pocket-storage approach (SSPS) to address the DGWP problem by decoupling native installed system software from hardware, making software go with person and running it with any other computer hardware. The paper illustrated the SSPS overall block diagram and workflow, internal architecture, data structure and management of desktop system image, image loader and virtual disk driver, and security considerations during the design and implementation. Also, the paper provided system evaluation of SSPS approach as well as technology forecasts.
