The concept,architecture,methods and applications of federated visualization were introduced.The federated visualization framework is capable of encrypting and training a visual model that reflect the characteristics of the entire data for specific tasks and scenarios.The federated visualization framework is an extension and application of federated learning,which emphasized using mutual benefit and win-win federal cooperation to visually analyze multi-source data under the premise of ensuring data privacy,towards breaking down data barriers in various fields and industries and realizing the sharing of data and knowledge.
In recent years,federated learning has been proposed and received widespread attention to overcome data isolated island challenge.Federated learning related researches were adopted in areas such as financial field,healthcare domain and smart city related application.Federated learning concept was introduced into three different layers.The first layer introduced the definition,architecture,classification of federated learning and compared the federated learning with traditional distributed learning.The second layer presented comparison and analysis of federated learning algorithms from machine learning and deep learning aspects.The third layer separated federated learning optimization algorithms into three aspects to optimize federated learning algorithm through reducing communication cost,selecting proper clients and different aggregation method.Finally,the current research status and three main challenges on communication,heterogeneity of system and data to be solved were concluded,and the future prospects in federated learning domain were proposed.
To this end, many laws and regulations on privacy protection have been introduced, and the phenomenon of data-island has become a major bottleneck hindering the development of big data and artificial intelligence technology.Federated learning has received widespread attention to break this phenomenon.Started with the historical development of federated learning, the definition, and architecture and classification of federated learning, the advantages of federated learning in privacy protection domainwere introduced.At the same time, various attack methods and their classification aboutfederated learning were introduced in detail.The classification of various encryption algorithms in federated learning were summarized.In conclusion, the contribution of federated learning in privacy protection and security mechanism were summarized and the new challenges in these domains were proposed.
To improve the communication efficiency in FL (federated learning), for the scenario with heterogeneous edge user's computing capacity and channel state, a class of time division multiple access (TDMA) based user scheduling policies were proposed for FL.The proposed policies aim to minimize the system delay in each round of model training subject to a given sample size constraint required for computing in each round.In addition, the convergence rate of the proposed scheduling algorithms was analyzed from a theoretical perspective to study the tradeoff between the convergence performance and the total system delay.The selection of the optimal batch size was further analyzed.Simulation results show that the convergence rate of the proposed algorithm is at least 30% higher than all the considered benchmarks.
To cope with the impact of different device computing capabilities and non-independent uniformly distributed data on federated learning performance, and to efficiently schedule terminal devices to complete model aggregation, a method of node selection based on deep reinforcement learning was proposed.It considered training quality and efficiency of heterogeneous terminal devices, and filtrate malicious nodes to guarantee higher model accuracy and shorter training delay of federated learning.Firstly, according to characteristics of model distributed training in federated learning, a node selection system model based on deep reinforcement learning was constructed.Secondly, considering such factors as device training delay, model transmission delay and accuracy, an optimization model of accuracy for node selection was proposed.Finally, the problem model was constructed as a Markov decision process and a node selection algorithm based on distributed proximal strategy optimization was designed to obtain a reasonable set of devices before each training iteration to complete model aggregation.Simulation results demonstrate that the proposed method significantly improves the accuracy and training speed of federated learning, and its convergence and robustness are also well.
Although there is a great value hidden in the massive data, it can also easily expose user privacy.Aiming at efficiently and securely sharing data from multiple parties and avoiding leakage of user private information, the development of related research and technologies on the non-aggregated data sharing field was introduced.Firstly, secure multi-party computing and its technologies were briefly described, including homomorphic encryption, oblivious transfer, secret sharing, etc.Secondly, the federated learning architecture was analyzed from the aspects of source data nodes and transmission optimization.Finally, the existing non-aggregated data sharing frameworks were listed and compared.In addition, the challenges and future potential research directions were summarized, such as complex multi-party scenarios, the balance between optimization and cost, as well as related security risks.
Federated learning has rapidly become a research hotspot in the field of security machine learning in recent years because it can train the global optimal model collaboratively without the need for multiple data source aggregation.Firstly, the federated learning framework, algorithm principle and classification were summarized.Then, the main threats and challenges it faced, were analysed indepth the comparative analysis of typical research programs in the three directions of communication efficiency, privacy and security, trust and incentive mechanism was focused on, and their advantages and disadvantages were pointed out.Finally, Combined with application of edge computing, blockchain, 5G and other emerging technologies to federated learning, its future development prospects and research hotspots was prospected.
Aiming at the needs of edge computing for data privacy, the correctness of calculation results and the auditability of data processing, a privacy protection method for edge computing based on blockchain and federated learning was proposed, which can realize collaborative training with multiple devices at the edge of the network without a trusted environment and special hardware facilities.The blockchain was used to endow the edge computing with features such as tamper-proof and resistance to single-point-of-failure attacks, and the gradient verification and incentive mechanism were incorporated into the consensus protocol to encourage more local devices to honestly contribute computing power and data to the federated learning.For the potential privacy leakage problems caused by sharing model parameters, an adaptive differential privacy mechanism was designed to protect parameter privacy while reducing the impact of noise on the model accuracy, and moments accountant was used to accurately track the privacy loss during the training process.Experimental results show that the proposed method can resist 30% of poisoning attacks, and can achieve privacy protection with high model accuracy, and is suitable for edge computing scenarios that require high level of security and accuracy.
Aiming at the problem that the difference of node data distribution has adverse effect on the performance of federated learning algorithm, a node selection algorithm based on label quantity information was proposed.An optimization objective based on the label quantity information of nodes was designed, considering the optimization problem of selecting the nodes with balanced label distribution under a certain time consumption limit.According to the correlation between the aggregated label distribution of selected nodes and the convergence of the global model, the upper bound of the weight divergence of the global model was reduced to improve the convergence stability of the algorithm.Simulation results shows that the new algorithm had higher convergence efficiency than the existing node selection algorithm.
Federated learning is a new distributed machine learning technology, where training tasks are deployed on user side and training model parameters are sent to the server side.In the whole process, participants do not need to share their own data directly, which greatly avoids privacy issues.However, the trust relationship between mobile users in the learning model has not been established in advance, there is hidden safety when users perform cooperative train with each other.In view of the above problems, a federated learning scheme for mobile network based on reputation evaluation mechanism and blockchain was proposed.The scheme allowed the server side to use subjective logic models to evaluate the reputation of participating mobile users and provided them with credible reputation opinions sharing environment and dynamic access strategy interface based on the technique of smart contract of blockchain.Theoretical and experimental analys is results show that the scheme can enable the server side to select reliable users for training.And it can achieve more fair and effective reputation calculations, which improves the accuracy of federated learning.
Volumn Content
