Fl-engine system in federated learning platform

Abstract

The application discloses a Fl-engine system in a joint learning platform, which comprises a joint learning platform, a local server and an internet of things access end, wherein the joint learning platform is internally provided with an intelligent ecological circle module, a joint learning plan module, a joint learning engine module and a platform support module; the joint learning plan module and the joint learning engine module are electrically connected; and the joint learning engine module and the platform support module are electrically connected. The Fl-engine system in the joint learning platform provided by the application is provided with the joint learning engine module, and incorporates a joint machine learning algorithm, a joint deep learning algorithm, a horizontal joint algorithm and a vertical joint algorithm in the joint learning engine module; through a self-adaptive mechanism, a k-means clustering strategy, a hierarchical clustering strategy, a SOM clustering strategy or a FCM clustering strategy matched with the joint learning engine module is selected from an aggregation strategy module, so that aggregation calculation is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of federated learning, in particular to a Fl-engine system in a federated learning platform. BACKGROUND

[0002] Machine learning algorithms have been widely used in data analysis, data mining, etc., and have achieved good results. At the same time, machine learning algorithms are also a kind of algorithm with strong specificity for specific applications or specific data, that is, once the training is completed using the given training data and is uniquely determined, the model will be difficult to have the generalization ability for different data or different applications. Therefore, in order to make the model have the self-adaptive ability for different data or applications, for example, by adjusting some controllable parameters in the model to automatically adjust its model structure, it is a very effective solution to expand the generalization ability of the model.

[0003] Through the joint learning of machines, we can provide various data training, the first kind is to do data set enhancement training, this kind of scene is mainly aimed at each party participating in the joint has a certain number of samples, training together is equivalent to data set enhancement for each party, this kind actually is equivalent to horizontal joint. The second kind can do feature aggregation, which is actually equivalent to vertical joint. The third kind of scene is to do some data labeling, one of the participating parties has features and the other has labeled data, both can learn the relationship between them without leaking privacy, so as to achieve the effect of labeling. The last kind is the most extreme case of independence of model and data. That is, one party only has a model and the other party only has data. Through joint learning, the model of one party can consume the data of the other party, and the value sharing is achieved on the premise of protecting the privacy of data and model.

[0004] The engine module in the federated learning platform is crucial. As a central node, the engine is deployed in the cloud and is responsible for uniformly arranging various joint training tasks and agents, playing a central coordinating role. SUMMARY

[0005] The purpose of the present application is to provide a Fl-engine system in a federated learning platform. The federated learning engine module is based on a federated learning framework, and incorporates federated machine learning algorithms, federated deep learning algorithms, horizontal joint algorithms and vertical joint algorithms. Through an adaptive mechanism, the k-means clustering strategy, hierarchical clustering strategy, SOM clustering strategy or FCM clustering strategy matched with the federated learning engine module is selected from the aggregation strategy module, aggregation calculation is realized, and a distributed exception handling module is provided to adopt distributed processing, improve work efficiency, and take into account various exception handling. In addition, the privacy installation protocol encrypts user privacy, has high security performance, can guarantee the integrity of federated training, and solves the problems in the above background technology.

[0006] In order to achieve the above object, the present application provides the following technical scheme: the Fl-engine system in the joint learning platform, including the joint learning platform, the local server and the Internet of Things access end, the intelligent ecological circle module, the joint learning plan module, the joint learning engine module and the platform support module are arranged in the joint learning platform, the joint learning plan module and the joint learning engine module are electrically connected, the joint learning engine module and the platform support module are electrically connected, the Internet of Things system, the local training configuration module, the local training agent module and the local resource management cooperation module are arranged in the local server, the Internet of Things system is electrically connected with the local training configuration module and the local training agent module respectively, the local training configuration module is electrically connected with the joint learning plan module, the local training agent module is electrically connected with the joint learning engine module, and the local resource management cooperation module is electrically connected with the platform support module, the machine equipment, the edge box and the camera are connected with the Internet of Things system respectively.

[0007] Preferably, the joint learning engine module is provided with an ML / DL algorithm module, an aggregation strategy module, a distributed exception processing module, a privacy installation protocol and an adaptive mechanism, and the joint learning engine module is also provided with a local model receiving node, the local model receiving node is in communication connection with the distributed exception processing module, the privacy installation protocol and the adaptive mechanism, and the adaptive mechanism is in corresponding connection with the ML / DL algorithm module and the aggregation strategy module.

[0008] Preferably, the ML / DL algorithm module is provided with a joint machine learning algorithm, a joint deep learning algorithm, a horizontal joint algorithm and a vertical joint algorithm, the joint machine learning algorithm, the joint deep learning algorithm, the horizontal joint algorithm and the vertical joint algorithm are arranged in parallel, and are in communication connection with the adaptive mechanism.

[0009] Preferably, the aggregation strategy module includes a k-means clustering strategy, a hierarchical clustering strategy, a SOM clustering strategy and an FCM clustering strategy, the k-means clustering strategy, the hierarchical clustering strategy, the SOM clustering strategy and the FCM clustering strategy are arranged in parallel, and are in communication connection with the adaptive mechanism.

[0010] Preferably, the distributed exception processing module processing exception includes the following steps:

[0011] S13301: The distributed exception processing module adopts a distributed structure, an exception detection logic is arranged in each flow water level in the local model receiving node, the type of exception generated by the flow water level in the local model receiving node in the current clock cycle is detected, and the related information of the highest priority exception generated by the level in the cycle is transmitted to the centralized exception arbitration logic;

[0012] S13302: Add a mirror register to the pipeline architecture register of the local model receiving node;

[0013] S13303: When processing the response exception, save the architecture register to the mirror register corresponding thereto, load the exception vector, and execute the exception service subroutine;

[0014] S13304: When the exception returns, restore the value of the mirror register to the architecture register corresponding thereto, the processor completes an exception response, and switches to the normal program flow.

[0015] Preferably, the local model receiving node comprises a fetch stage, a decoding stage, an execution stage, a memory access stage, and a write-back stage, which are connected in sequence, and each of the fetch stage, the decoding stage, the execution stage, the memory access stage, and the write-back stage is provided with an exception detection logic of a distributed exception processing module.

[0016] Preferably, the privacy installation protocol comprises an acquisition module, a determination module, and a replacement module, which are connected in sequence, and the acquisition module and the replacement module are connected with the local model receiving node.

[0017] Preferably, the method for privacy encryption of the privacy installation protocol comprises the following steps:

[0018] S13401: A pre-trained model is set in the acquisition module, and a loss function is designed according to a task, and the loss function adopts a combination of cross entropy and crf;

[0019] S13402: The acquisition module performs model training to obtain a trained model;

[0020] S13403: The trained model is used for privacy data recognition of a text, and the determination module determines the privacy data used;

[0021] S13404: The replacement module replaces the privacy data of the user by using a set symbol.

[0022] Preferably, the adaptive mechanism comprises a model recognition module, a key information extraction module, an intelligent matching module, and a feedback module, which are connected in sequence, and the intelligent matching module is connected with an ML / DL algorithm module and an aggregation strategy module.

[0023] Preferably, the working process of the adaptive mechanism comprises the following steps:

[0024] S13501: The model recognition module recognizes the model category and quantity received by the local model receiving node;

[0025] S13502: The key information extraction module extracts key information from the model category and quantity, and delivers the key information to the intelligent matching module;

[0026] S13503: The intelligent matching module selects a suitable algorithm from the ML / DL algorithm module according to the key information, and determines the corresponding aggregation strategy in the aggregation strategy module according to the selected algorithm.

[0027] Compared with the prior art, the beneficial effects of the present application are:

[0028] 1、The Fl-engine system in the joint learning platform proposed by the present application, the joint learning engine module is based on a joint learning framework, and is provided with a joint learning engine module, and incorporates a joint machine learning algorithm, a joint deep learning algorithm, a horizontal joint algorithm and a vertical joint algorithm in the joint learning engine module, selects a k-means clustering strategy, a hierarchical clustering strategy, a SOM clustering strategy or an FCM clustering strategy matched therewith from the aggregation strategy module through an adaptive mechanism, realizes aggregation calculation, and can perform aggregation calculation on different data models.

[0029] 2、The Fl-engine system in the joint learning platform proposed by the present application, the joint learning engine module is provided with a distributed exception handling module, and the distributed exception handling module adopts distributed processing, improves work efficiency, and takes into account various exception handling.

[0030] 3、The Fl-engine system in the joint learning platform proposed by the present application, the joint learning engine module is also provided with a privacy installation protocol for user privacy encryption, high safety performance, and can guarantee the integrity of joint training. DETAILED DESCRIPTION

[0031] Figure 1 It is a whole structure diagram of the present application;

[0032] Figure 2 It is a structure diagram of the joint learning engine module of the present application;

[0033] Figure 3 It is a connection relationship diagram of the joint learning engine module of the present application;

[0034] Figure 4 It is a work flow diagram of the distributed exception handling module of the present application;

[0035] Figure 5 It is a work flow diagram of the privacy installation protocol of the present application;

[0036] Figure 6 It is a work flow diagram of the adaptive mechanism of the present application;

[0037] Figure 7 Structure diagram of embodiment two of the present application.

[0038] In the figure: 1, joint learning platform; 11, intelligent ecological circle module; 12, joint learning plan module; 13, joint learning engine module; 131, ML / DL algorithm module; 1311, joint machine learning algorithm; 1312, joint deep learning algorithm; 1313, horizontal joint algorithm; 1314, vertical joint algorithm; 132, aggregation strategy module; 1321, k-means clustering strategy; 1322, hierarchical clustering strategy; 1323, SOM (Self Organizing Maps) clustering strategy; 1324, FCM clustering strategy (Fuzzy C-means); 133, distributed exception processing module; 134, privacy installation protocol; 1341, acquisition module; 1342, determination module; 1343, replacement module; 135, adaptive mechanism; 1351, model identification module; 1352, key information extraction module; 1353, intelligent matching module; 1354, feedback module; 136, local model receiving node; 1361, instruction fetching stage; 1362, decoding stage; 1363, execution stage; 1364, memory access stage; 1365, write back stage; 14, platform support module; 2, local server; 21, Internet of Things system; 22, local training configuration module; 23, local training agent module; 24, local resource management collaboration module; 3, Internet of Things access end; 31, machine device; 32, edge box; 33, camera. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0040] Embodiment one:

[0041] Please refer to Figure 1The Fl-engine system in the joint learning platform comprises a joint learning platform 1, a local server 2 and an Internet of Things access end 3. The joint learning platform 1 is internally provided with an intelligent ecological circle module 11, a joint learning plan module 12, a joint learning engine module 13 and a platform support module 14. The joint learning plan module 12 and the joint learning engine module 13 are electrically connected. The joint learning engine module 13 and the platform support module 14 are electrically connected. The local server 2 is internally provided with an Internet of Things system 21, a local training configuration module 22, a local training agent module 23 and a local resource management cooperation module 24. The Internet of Things system 21 is electrically connected with the local training configuration module 22 and the local training agent module 23 respectively. The local training configuration module 22 is electrically connected with the joint learning plan module 12. The local training agent module 23 is electrically connected with the joint learning engine module 13. The local resource management cooperation module 24 is electrically connected with the platform support module 14. The Internet of Things access end 3 comprises a machine device 31, an edge box 32 and a camera 33. The machine device 31, the edge box 32 and the camera 33 are connected with the Internet of Things system 21 respectively.

[0042] Please refer to Figure 2 The joint learning engine module 13 is internally provided with an ML (Maching Learning) / DL (Deep Learning) algorithm module 131, an aggregation strategy module 132, a distributed exception processing module 133, a privacy installation protocol 134 and an adaptive mechanism 135. The joint learning engine module 13 is also internally provided with a local model receiving node 136. The local model receiving node 136 is communicatively connected with the distributed exception processing module 133, the privacy installation protocol 134 and the adaptive mechanism 135. The adaptive mechanism 135 is correspondingly connected with the ML / DL algorithm module 131 and the aggregation strategy module 132. The ML / DL algorithm module 131 provides a basic algorithm module. The aggregation strategy module 132 is used for realizing an aggregation calculation strategy. The adaptive mechanism 135 is used for supply-demand matching. The algorithm in the ML / DL algorithm module 131 is matched with the aggregation strategy in the aggregation strategy module 132. The distributed exception processing module 133 gives a processing strategy in time when a training node appears an exception or is interrupted. The privacy installation protocol 134 is used for an encryption algorithm.

[0043] Please refer to Figure 3 The ML / DL algorithm module 131 is internally provided with a joint machine learning algorithm 1311, a joint deep learning algorithm 1312, a horizontal joint algorithm 1313 and a vertical joint algorithm 1314. The joint machine learning algorithm 1311, the joint deep learning algorithm 1312, the horizontal joint algorithm 1313 and the vertical joint algorithm 1314 are arranged in parallel and are communicatively connected with the adaptive mechanism 135.

[0044] The aggregation strategy module 132 includes a k-means clustering strategy 1321 (K-means is an unsupervised algorithm used in the process of clustering problems), a hierarchical clustering strategy 1322, a SOM clustering strategy 1323, and an FCM clustering strategy 1324, which are arranged in parallel and are in communication connection with the adaptive mechanism 135.

[0045] Referring to Figure 4 The distributed exception processing module 133 processes exceptions, including the following steps:

[0046] S13301: The distributed exception processing module 133 adopts a distributed structure, and sets an exception detection logic in each pipeline stage in the local model receiving node 136 to detect the type of exception generated by the pipeline stage in the local model receiving node 136 in the current clock cycle, and deliver the related information of the highest priority exception generated by the stage in the cycle to the centralized exception arbitration logic;

[0047] S13302: An image register is added to the architecture register of the pipeline system mechanism of the local model receiving node 136;

[0048] S13303: When processing a response exception, save the architecture register to the image register corresponding thereto, load the exception vector, and execute the exception service subroutine;

[0049] S13304: When the exception returns, restore the value of the image register to the architecture register corresponding thereto, and the processor completes an exception response and switches to the normal program flow.

[0050] The local model receiving node 136 includes a fetch stage 1361, a decode stage 1362, an execution stage 1363, a memory access stage 1364, and a write back stage 1365, which are connected in sequence, and each of the fetch stage 1361, the decode stage 1362, the execution stage 1363, the memory access stage 1364, and the write back stage 1365 is provided with an exception detection logic of the distributed exception processing module 133.

[0051] Referring to Figure 5 The privacy installation protocol 134 includes an acquisition module 1341, a determination module 1342, and a replacement module 1343, which are connected in sequence, the acquisition module 1341 and the replacement module 1343 are connected with the local model receiving node 136, and the privacy installation protocol 134 includes the following steps:

[0052] S13401: Set the pre-training model in the acquisition module 1341, and design the loss function according to the task. The loss function adopts the combination of cross entropy and crf;

[0053] S13402: The acquisition module 1341 performs model training to obtain a trained model;

[0054] S13403: Use the trained model to identify private data in the text, and determine the private data used by the determination module 1342;

[0055] S13404: The replacement module 1343 replaces the user's private data with a set symbol.

[0056] Please refer to Figure 6 , the adaptive mechanism 135 includes a model identification module 1351, a key information extraction module 1352, an intelligent matching module 1353 and a feedback module 1354, the model identification module 1351, the key information extraction module 1352, the intelligent matching module 1353 and the feedback module 1354 are connected in turn, the intelligent matching module 1353 is connected with the ML / DL algorithm module 131 and the aggregation strategy module 132, and the working process of the adaptive mechanism 135 includes the following steps:

[0057] S13501: The model identification module 1351 identifies the model category and quantity received by the local model receiving node 136;

[0058] S13502: The key information extraction module 1352 extracts key information from the model category and quantity, and delivers the key information to the intelligent matching module 1353;

[0059] S13503: The intelligent matching module 1353 selects appropriate algorithms from the ML / DL algorithm module 131 according to the key information, and determines the corresponding aggregation strategy in the aggregation strategy module 132 according to the selected algorithm.

[0060] Working process: data acquisition is performed through the Internet of Things access end 3, the Internet of Things system 21 in the local server 2 processes the data and transmits them to the local training configuration module 22 and the local training agent module 23, the local training configuration module 22 and the local training agent module 23 are connected with the federated learning engine module 13 respectively, the federated learning engine module 13 is arranged in the cloud and is responsible for unified arrangement and management of various federated training tasks and agents, plays a central coordinating role, performs a series of aggregation operations on the models uploaded by each node locally, different aggregation strategies of different algorithms are different, the federated learning engine module 13 performs unified processing, the federated machine learning algorithm 1311, the federated deep learning algorithm 1312, the horizontal federated algorithm 1313 and the vertical federated algorithm 1314 perform algorithm calculation on different models, the adaptive mechanism 135 automatically selects an aggregation strategy, before the aggregation operation, the distributed exception processing module 133 adopts distributed processing to improve work efficiency and take into account various exception processing, in addition, the privacy installation protocol 134 encrypts user privacy, in the embodiment, the adaptive mechanism 135 selects the federated machine learning algorithm 1311 and matches the k-means clustering strategy 1321, and the specific steps are as follows:

[0061] Step 1: obtaining a controllable parameter for controlling the calculation time of the machine learning algorithm, and establishing a quantization model library of the calculation time of the machine learning algorithm according to the actual calculation time of the machine learning algorithm at each specific numerical value of the controllable parameter,

[0062] The processing process of the k-means algorithm is as follows: first, randomly select k objects, each object initially represents the average value or center of a cluster; for each remaining object, assign it to the nearest cluster according to its distance from each cluster center; then recalculate the average value of each cluster, and this process is repeated until the criterion function converges. Usually, the squared error criterion is used, which is defined as follows:

[0063]

[0064] Here E is the sum of the squared errors of all objects in the database, p is a point in space, and mi is the average value of cluster Ci. The objective function makes the generated clusters as compact and independent as possible, and the distance metric used is the Euclidean distance, but other distance metrics can also be used.

[0065] Step 2: according to the complexity of the input data in each time window, the structure of the machine learning algorithm is adjusted in a coarse-grained manner, the complexity range of the algorithm model is given, the input data is quantitatively described according to the machine learning algorithm, the specific numerical value of the controllable parameter is determined in the quantization model library combined with the given time limit, and the specific numerical value is applied to the machine learning algorithm.

[0066] Embodiment Two:

[0067] Please refer to Figure 7 , the Fl-engine system in the joint learning platform, including a joint learning platform 1, a local server 2 and an Internet of Things access end 3, the joint learning platform 1 is provided with an intelligent ecological circle module 11, a joint learning plan module 12, a joint learning engine module 13 and a platform support module 14, the joint learning plan module 12 and the joint learning engine module 13 are electrically connected, the joint learning engine module 13 and the platform support module 14 are electrically connected, the local server 2 is provided with an Internet of Things system 21, a local training configuration module 22, a local training agent module 23 and a local resource management cooperation module 24, the Internet of Things system 21 is respectively electrically connected with the local training configuration module 22 and the local training agent module 23, the local training configuration module 22 is electrically connected with the joint learning plan module 12, the local training agent module 23 is electrically connected with the joint learning engine module 13, the local resource management cooperation module 24 is electrically connected with the platform support module 14, the Internet of Things access end 3 includes a machine device 31, an edge box 32 and a camera 33, the machine device 31, the edge box 32 and the camera 33 are respectively connected with the Internet of Things system 21.

[0068] The joint learning engine module 13 is provided with an ML / DL algorithm module 131, an aggregation strategy module 132, a distributed exception processing module 133, a privacy installation protocol 134 and an adaptive mechanism 135, the joint learning engine module 13 is also provided with a local model receiving node 136, the local model receiving node 136 is communicatively connected with the distributed exception processing module 133, the privacy installation protocol 134 and the adaptive mechanism 135, the adaptive mechanism 135 is correspondingly connected with the ML / DL algorithm module 131 and the aggregation strategy module 132, the ML / DL algorithm module 131 provides a basic algorithm module, the aggregation strategy module 132 is used for realizing an aggregation calculation strategy, the adaptive mechanism 135 is used for supply and demand matching, matching the algorithm in the ML / DL algorithm module 131 with the aggregation strategy in the aggregation strategy module 132, the distributed exception processing module 133 gives a processing strategy in time when a training node appears an exception or interruption, the privacy installation protocol 134 is used for an encryption algorithm.

[0069] The ML / DL algorithm module 131 is provided with a joint machine learning algorithm 1311, a joint deep learning algorithm 1312, a horizontal joint algorithm 1313 and a vertical joint algorithm 1314, the joint machine learning algorithm 1311, the joint deep learning algorithm 1312, the horizontal joint algorithm 1313 and the vertical joint algorithm 1314 are arranged in parallel, and are all communicatively connected with the adaptive mechanism 135.

[0070] The aggregation strategy module 132 includes a k-means clustering strategy 1321, a hierarchical clustering strategy 1322, a SOM clustering strategy 1323, and an FCM clustering strategy 1324, which are arranged in parallel and are in communication connection with the adaptive mechanism 135.

[0071] The distributed exception processing module 133 processes the exception, including the following steps:

[0072] S13301: The distributed exception processing module 133 adopts a distributed structure, and sets an exception detection logic in each flow stage in the local model receiving node 136 to detect the type of exception generated by the flow stage in the local model receiving node 136 in the current clock cycle, and transmit the related information of the highest priority exception generated by the stage in the cycle to the centralized exception arbitration logic;

[0073] S13302: An image register is added to the architecture register of the flow stage system mechanism of the local model receiving node 136;

[0074] S13303: When processing the response exception, save the architecture register to the image register corresponding thereto, load the exception vector, and execute the exception service subroutine;

[0075] S13304: When the exception returns, restore the value of the image register to the architecture register corresponding thereto, and the processor completes a response to the exception and switches to the normal program flow.

[0076] The local model receiving node 136 includes a fetch stage 1361, a decoding stage 1362, an execution stage 1363, a memory access stage 1364, and a write-back stage 1365, which are connected in sequence. The fetch stage 1361, the decoding stage 1362, the execution stage 1363, the memory access stage 1364, and the write-back stage 1365 are each provided with an exception detection logic of the distributed exception processing module 133.

[0077] The privacy installation protocol 134 includes an acquisition module 1341, a determination module 1342, and a replacement module 1343, which are connected in sequence. The acquisition module 1341 and the replacement module 1343 are connected with the local model receiving node 136. The privacy installation protocol 134 performs a privacy encryption method, including the following steps:

[0078] S13401: A pre-trained model is set in the acquisition module 1341, and a loss function is designed according to a task. The loss function adopts a combination of cross-entropy and crf.

[0079] S13402: The obtaining module 1341 performs model training to obtain a trained model;

[0080] S13403: The trained model is used for privacy data recognition of the text, and the determining module 1342 determines the privacy data used;

[0081] S13404: The replacement module 1343 replaces the user's privacy data with a set symbol.

[0082] The adaptive mechanism 135 includes a model identification module 1351, a key information extraction module 1352, an intelligent matching module 1353, and a feedback module 1354, which are connected in sequence. The intelligent matching module 1353 is connected with the ML / DL algorithm module 131 and the aggregation strategy module 132. The working process of the adaptive mechanism 135 includes the following steps:

[0083] S13501: The model identification module 1351 identifies the model category and quantity received by the local model receiving node 136;

[0084] S13502: The key information extraction module 1352 extracts key information from the model category and quantity, and delivers the key information to the intelligent matching module 1353;

[0085] S13503: The intelligent matching module 1353 selects a suitable algorithm from the ML / DL algorithm module 131 according to the key information, and determines the corresponding aggregation strategy in the aggregation strategy module 132 according to the selected algorithm.

[0086] Working process: data acquisition is performed through the Internet of Things access end 3, the Internet of Things system 21 in the local server 2 processes the data and transmits them to the local training configuration module 22 and the local training agent module 23, the local training configuration module 22 and the local training agent module 23 are connected with the federated learning engine module 13 respectively, the federated learning engine module 13 is arranged in the cloud and is responsible for unified arrangement and management of various federated training tasks and agents, plays a central coordination role, performs a series of aggregation operations on the models uploaded by each node locally, different aggregation strategies of different algorithms are different, the federated learning engine module 13 performs unified processing, the federated machine learning algorithm 1311, the federated deep learning algorithm 1312, the horizontal federated algorithm 1313 and the vertical federated algorithm 1314 perform algorithm calculation on different models, the adaptive mechanism 135 automatically selects an aggregation strategy, before the aggregation operation, the distributed exception processing module 133 adopts distributed processing to improve work efficiency and take into account various exception processing, in addition, the privacy installation protocol 134 encrypts user privacy, in the embodiment, the adaptive mechanism 135 selects the federated deep learning algorithm 1312 and matches the SOM clustering strategy 1323, and the specific steps are as follows:

[0087] Step 1: the federated deep learning algorithm 1312 obtains a set of input data, wherein the set of input data includes original data arranged into a plurality of data clusters;

[0088] Step 2: the federated deep learning algorithm 1312 includes an input layer, an output layer and a plurality of hidden layers, the hierarchical clustering strategy 1322 uses a deep learning algorithm to perform statistical clustering on the original data, thereby generating statistical clusters, the SOM clustering strategy 1323 includes an input layer and an output layer, the input layer corresponds to a high-dimensional input vector, the output layer is composed of a series of ordered nodes organized in a two-dimensional grid, the input nodes and the output nodes are connected through a weight vector, in the learning process, the output layer unit with the shortest distance is found, that is, the winning unit, the weight values of the adjacent regions are updated, so that the output nodes maintain the topological characteristics of the input vector, and the algorithm flow is as follows:

[0089] (1) network initialization, initial value is assigned to the weight of each node of the output layer;

[0090] (2) an input vector is randomly selected from the input sample, and a weight vector with the minimum distance from the input vector is found;

[0091] (3) the winning unit is defined, and the weights in the adjacent region of the winning unit are adjusted to make them approach the input vector;

[0092] (4) a new sample is provided, and training is performed;

[0093] (5) the neighborhood radius is contracted, the learning rate is reduced, and the above steps are repeated until the value is less than the allowed value, and the clustering result is output.

[0094] Step 3: obtaining a label from each statistical cluster, wherein each label is a single data label contained in the cluster, and the set of input data is evaluated based on the label to derive associated data about one or more objects.

[0095] To sum up: the Fl-engine system in the joint learning platform provided by the application, the joint learning engine module 13 is based on a joint learning framework, and incorporates a joint machine learning algorithm 1311, a joint deep learning algorithm 1312, a horizontal joint algorithm 1313 and a vertical joint algorithm 1314, selects a k-means clustering strategy 1321, a hierarchical clustering strategy 1322, a SOM clustering strategy 1323 or a FCM clustering strategy 1324 that matches from the aggregation strategy module 132 through an adaptive mechanism 135, realizes aggregation calculation, and is provided with a distributed exception processing module 133 that adopts distributed processing, improves work efficiency, and takes into account various exception processing, in addition, the privacy installation protocol 134 encrypts the privacy of the user, has high safety performance, and can guarantee the integrity of joint training.

[0096] The above is only the preferred embodiment of the application, but the protection scope of the application is not limited to this, any person skilled in the art can replace or change the technical solution and the inventive concept of the application within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A Fl-engine system in a joint learning platform, comprising a joint learning platform (1), a local server (2) and an Internet of Things access end (3), characterized in that: The joint learning platform (1) is provided with an intelligent ecological circle module (11), a joint learning plan module (12), a joint learning engine module (13) and a platform support module (14), the joint learning plan module (12) and the joint learning engine module (13) are electrically connected, the joint learning engine module (13) and the platform support module (14) are electrically connected, the local server (2) is provided with an Internet of Things system (21), a local training configuration module (22), a local training agent module (23) and a local resource management cooperation module (24), the Internet of Things system (21) is electrically connected with the local training configuration module (22) and the local training agent module (23) respectively, the local training configuration module (22) is electrically connected with the joint learning plan module (12), the local training agent module (23) is electrically connected with the joint learning engine module (13), and the local resource management cooperation module (24) is electrically connected with the platform support module (14), the Internet of Things access end (3) includes a machine device (31), an edge box (32) and a camera (33), and the machine device (31), the edge box (32) and the camera (33) are connected with the Internet of Things system (21) respectively; The joint learning engine module (13) is arranged in the cloud, is responsible for unified arrangement and management of each joint training task and Agent, plays a central coordination role, and performs a series of aggregation operations on the models uploaded by each node locally, and the aggregation strategies of different algorithms are different; The joint learning engine module (13) is provided with an ML / DL algorithm module (131), an aggregation strategy module (132), a distributed exception processing module (133), a privacy installation protocol (134) and a self-adaptive mechanism (135), the joint learning engine module (13) is further provided with a local model receiving node (136), the local model receiving node (136) is in communication connection with the distributed exception processing module (133), the privacy installation protocol (134) and the self-adaptive mechanism (135) respectively, the self-adaptive mechanism (135) is in communication connection with the ML / DL algorithm module (131) and the aggregation strategy module (132) respectively; The ML / DL algorithm module (131) provides a basic algorithm module, the aggregation strategy module (132) is used for realizing an aggregation calculation strategy, the self-adaptive mechanism (135) is used for supply and demand matching, the algorithm in the ML / DL algorithm module (131) is matched with the aggregation strategy in the aggregation strategy module (132), the distributed exception processing module (133) is used for giving a processing strategy in time when an exception or interruption occurs in a training node, and the privacy installation protocol (134) is used for an encryption algorithm.

2. The Fl-engine system in a federated learning platform of claim 1, wherein: The ML / DL algorithm module (131) is internally provided with a joint machine learning algorithm (1311), a joint deep learning algorithm (1312), a horizontal joint algorithm (1313) and a vertical joint algorithm (1314), which are arranged in parallel and are in communication connection with the adaptive mechanism (135).

3. The Fl-engine system in a federated learning platform of claim 1, wherein: The aggregation strategy module (132) includes a k-means clustering strategy (1321), a hierarchical clustering strategy (1322), a SOM clustering strategy (1323) and an FCM clustering strategy (1324), which are arranged in parallel and are in communication connection with the adaptive mechanism (135).

4. The Fl-engine system in a federated learning platform of claim 1, wherein: The distributed exception processing module (133) processes the exception, including the following steps: S13301: The distributed exception processing module (133) adopts a distributed structure, and sets an exception detection logic in each flow stage in the local model receiving node (136) to detect the type of exception generated by the flow stage in the local model receiving node (136) in the current clock cycle, and transmit the related information of the highest priority exception generated by the stage in the cycle to the centralized exception arbitration logic; S13302: An image register is added to the flow stage architecture register of the local model receiving node (136); S13303: When responding to an exception, save the architecture register to the corresponding image register, load the exception vector, and execute the exception service subroutine; S13304: When the exception returns, restore the value of the image register to the corresponding architecture register, and the processor completes an exception response and switches to the normal program flow.

5. The Fl-engine system in a federated learning platform of claim 4, wherein: The local model receiving node (136) includes a fetch stage (1361), a decode stage (1362), an execution stage (1363), a memory stage (1364) and a write back stage (1365), which are connected in turn, and each of the fetch stage (1361), the decode stage (1362), the execution stage (1363), the memory stage (1364) and the write back stage (1365) is provided with an exception detection logic of the distributed exception processing module (133).

6. The Fl-engine system in a federated learning platform of claim 1, wherein: The privacy installation protocol (134) includes an acquisition module (1341), a determination module (1342) and a replacement module (1343), which are connected in turn, and the acquisition module (1341) and the replacement module (1343) are connected with the local model receiving node (136).

7. The Fl-engine system in a federated learning platform of claim 6, wherein: The method for privacy encryption of the privacy installation protocol (134) includes the following steps: S13401: Set the pre-training model in the acquisition module (1341), and design the loss function according to the task. The loss function adopts the combination of cross-entropy and crf; S13402: The acquisition module (1341) performs model training to obtain a trained model; S13403: Use the trained model to identify private data in the text, and determine the private data used by the determining module (1342); S13404: The replacement module (1343) replaces the user's private data with a set symbol.

8. The Fl-engine system in a federated learning platform of claim 1, wherein: The adaptive mechanism (135) includes a model identification module (1351), a key information extraction module (1352), an intelligent matching module (1353), and a feedback module (1354), which are connected in sequence. The intelligent matching module (1353) is connected with the ML / DL algorithm module (131) and the aggregation strategy module (132).

9. The Fl-engine system in a federated learning platform of claim 8, wherein: The working process of the adaptive mechanism (135) includes the following steps: S13501: The model identification module (1351) identifies the model category and quantity received by the local model receiving node (136); S13502: The key information extraction module (1352) extracts key information from the model category and quantity, and delivers the key information to the intelligent matching module (1353); S13503: The intelligent matching module (1353) selects appropriate algorithms from the ML / DL algorithm module (131) according to the key information, and determines the corresponding aggregation strategy in the aggregation strategy module (132) according to the selected algorithm.

Images