An efficient asynchronous federated training method for edge computing

Abstract

The present invention discloses an efficient asynchronous federated training method for edge computing, including: for the long-tail problem of interaction, a timed inquiry mechanism is designed to control the interaction duration from the perspective of time constraints and inquiry supplements, and within the time constraints, the parameter server receives the first part of the aggregated subset; for the problem of aggregate heterogeneity, a screening and supplement strategy is designed to continue screening the truncated clients that meet the iteration round limit in the active inquiry phase from the perspective of data combination, and within the communication constraints, the parameter server receives the second part of the aggregated subset; under the interaction of the timed inquiry mechanism and the screening and supplement strategy, through the two-stage client reception, the parameter server receives the complete aggregated subset for global model aggregation update, and synchronizes the updated model to the client. The present invention can improve the interaction and convergence efficiency of global model update under the federated learning framework.

Description

Technical Field

[0001] The present invention relates to the field of computer science and technology, and in particular to an efficient asynchronous federated training method for edge computing. Background Art

[0002] With the increasing awareness of data security and privacy protection, the combination of edge computing service sinking and federated learning data federation has effectively overcome challenges such as service overload, resource constraints, data silos, and privacy security, further promoting the development of deep learning and attracting the attention of many relevant researchers. However, the phenomena of system heterogeneity, device heterogeneity, and data heterogeneity in the edge computing environment make the federated learning framework face challenges such as unstable device performance, network environment, and distribution differences between local data sets, which poses a threat to the efficiency of federated learning from different angles:

[0003] (1) The long tail of interaction, that is, due to the instability of device performance and network environment, slow clients appear in the single-round aggregation subset, resulting in unstable interaction time between the parameter server and the client. First, edge devices are not dedicated devices for model training. Due to limited device resources, it is impossible to continuously guarantee the high performance of local model training, resulting in dynamic changes in local update time; on the other hand, edge devices are geographically distributed and are mostly in an open network environment. Competition for network resources causes the network quality of each communication link to change dynamically, causing the communication time to change. Since local update time and communication time are the main components of interaction time, the federated learning framework in the edge environment cannot guarantee the stability of interaction time.

[0004] (2) Aggregation heterogeneity, that is, the difference in the correlation between the size of different aggregate subsets and the quality of global model updates caused by factors such as data heterogeneity and device heterogeneity. Different clients have differences in data quality, device performance, and network environment. Therefore, the client subsets passively received by the parameter server are prone to non-optimal combinations of local models, and cannot guarantee the positive correlation between the quality of the aggregate subset model and the global model update, especially when the data volume varies greatly and the label distribution is uneven. The heterogeneity of the aggregate subset causes the update direction of the global model to fluctuate, reducing the convergence speed of the global model.

[0005] It can be seen that the long tail of interactions and the heterogeneous aggregation problems under federated learning increase the collaborative interaction time and model convergence time in the global aggregation model training process, resulting in a significant reduction in the model training efficiency under the federated learning framework. Summary of the invention

[0006] In response to the above problems, the purpose of the present invention is to provide an efficient asynchronous federated training method for edge computing, which is used to flexibly regulate each client participating in federated learning to ensure the efficiency of collaborative training of the global model under the federated learning framework.

[0007] To solve the above technical problems, the embodiments of the present invention provide the following solutions:

[0008] An efficient asynchronous federated training method for edge computing includes the following steps:

[0009] To address the long-tail interaction problem, a timed query mechanism is designed to control the interaction duration from the perspective of time constraints and query supplements. Within the time constraints, the parameter server receives the first part of the aggregated subset.

[0010] Aiming at the problem of heterogeneous aggregation, a screening and supplementation strategy is designed to continue screening the truncated clients that meet the iteration round limit in the active inquiry phase from the perspective of data combination. Within the communication constraints, the parameter server receives the second part of the aggregate subset.

[0011] Under the interaction of the timed query mechanism and the screening and supplementation strategy, through two-stage client reception, the parameter server receives the complete aggregated subset for global model aggregation update and synchronizes the updated model to the client. No information exchange is performed between clients.

[0012] Preferably, the timing inquiry mechanism is designed for the interactive long tail problem, including:

[0013] For the timing constraint phase, the parameter server faces all clients in the entire interaction round and waits for any reachable client through the timing constraint (based on the premise that local data has been pre-processed, and the processing of dirty data and Byzantine devices is not within the scope of this invention). Within the time constraint, the parameter server receives the first client C that has arrived in advance. pre The aggregate subset composed of is the first part of the entire aggregate subset;

[0014] During the active inquiry phase, the parameter server is transformed from a traditional passive receiver to an active inquirer, and continues to send requests to other clients with supplementary inquiry information. When the client's local model meets the query requirements of the parameter server, it can choose to truncate the current local update status and directly upload the local model parameters.

[0015] Preferably, the client C is reached in advance pre It means that all iterative rounds of local updates are completed in advance during the timing constraint phase of the timing query mechanism, and the local model parameters are delivered to the client of the parameter server in a timely manner.

[0016] Preferably, a minimum iteration round threshold is imposed as a constraint in the active inquiry phase, and the iteration round is used as a basic criterion for screening the model quality to determine whether to truncate the current local update state of the client to obtain the qualification to upload supplements to the parameter server.

[0017] Preferably, under the action of the timing query mechanism, after receiving the aggregated subset in the timing constraint phase, the parameter server actively queries through the parameter server to obtain the candidate subset that can continue to be supplemented to the parameter server in the current interaction round, that is, the truncated client model that meets the iteration round limit.

[0018] Preferably, the screening supplement strategy is designed to address the problem of polymerization heterogeneity, including:

[0019] Within the communication constraints, from the perspective of data combination, the truncated clients that meet the iteration round limit in the active query phase are further screened according to the screening index, and the aggregated subset received by the parameter server is further expanded, that is, the truncated supplementary client C trunc and communication supplement client C gap The second aggregate subset composed of the two aggregate subsets is the second part of the entire aggregate subset.

[0020] Preferably, the truncated supplementary client C trunc Refers to clients whose local updates are not fully completed but are truncated for upload and supplementation in the current round due to meeting the iteration round limit and screening supplement conditions;

[0021] Communication Supplement Client C gap It refers to the client that has completed the fixed iteration round requirements in the timed query phase, but is still on the communication link due to insufficient waiting time in the timed query phase. Through the subsequent screening and supplement phase, it gets another upload opportunity to the parameter server during the waiting time.

[0022] Preferably, under the interaction of the timing query mechanism and the screening supplement strategy, through two-stage client reception, the parameter server receives a complete aggregated subset, including:

[0023] Based on the timed query, the client of the aggregated subset is converted from one-stage reception to two-stage reception and supplementation. Through the two-stage client reception, the parameter server receives the complete aggregated subset;

[0024] A topology architecture centered on a parameter server is adopted. A global aggregate model is obtained through multiple rounds of communication interactions between the aggregate model update of the parameter server and the local model of the client, and the updated model is synchronized to the client. There is no communication interaction between the clients.

[0025] Preferably, the parameter server sends query information, which includes a minimum iteration round threshold and a screening indicator for screening the supplementary strategy.

[0026] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0027] The present invention proposes an efficient asynchronous federated training method for edge computing. Under the action of the timed inquiry mechanism, on the one hand, abnormal clients such as slow and disconnected clients are directly controlled through time constraints to avoid delays of varying degrees in the interaction time, thereby ensuring the interaction efficiency of learning the global aggregation model; on the other hand, the parameter server is actively queried to request slow clients that meet the requirements to participate in the current round of aggregation updates, thereby improving the initiative of the parameter server and increasing the client's participation; under the action of the screening and supplementation strategy, the secondary supplementation combination of client subsets can be flexibly performed according to the local update status of the client, ensuring the global model aggregation requirements on the number of client subsets and model quality, thereby improving the performance improvement and convergence efficiency of the global aggregation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 It is a flow chart of an efficient asynchronous federated training method for edge computing provided by an embodiment of the present invention;

[0030] Figure 2 is a schematic diagram of a timing inquiry mechanism provided by an embodiment of the present invention;

[0031] Figure 3 is a schematic diagram of a screening and supplementation strategy provided by an embodiment of the present invention;

[0032] Figure 4 This is a diagram describing the overall framework of an efficient asynchronous federated training method for edge computing provided by an embodiment of the present invention.

[0033] As shown in the figure, in order to clearly implement the structure of the embodiment of the present invention, specific structures and devices are marked in the figure, but this is only for illustrative purposes and is not intended to limit the present invention to the specific structure, device and environment. According to specific needs, ordinary technicians in this field can adjust or modify these devices and environments, and the adjustments or modifications made are still included in the scope of protection of the present invention. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] The embodiment of the present invention provides an efficient asynchronous federated training method for edge computing, such as Figure 1 As shown, the method comprises the following steps:

[0036] Step 1: To address the long-tail interaction problem, a timed query mechanism is designed to control the interaction duration from the perspective of time constraints and query supplements. Within the time constraints, the parameter server receives the first part of the aggregated subset.

[0037] In this step, a timed query mechanism is designed to address the long-tail interaction problem, including:

[0038] For the timing constraint phase, the parameter server faces all clients in the entire interaction round and waits for any reachable client through the timing constraint (based on the premise that local data has been pre-processed, and the processing of dirty data and Byzantine devices is not within the scope of this invention). Within the time constraint, the parameter server receives the first client C that has arrived in advance. pre The aggregate subset composed of is the first part of the entire aggregate subset;

[0039] During the active inquiry phase, the parameter server is transformed from a traditional passive receiver to an active inquirer, and continues to send requests to other clients with supplementary inquiry information. When the client's local model meets the query requirements of the parameter server, it can choose to truncate the current local update status and directly upload the local model parameters.

[0040] Among them, the client C is reached in advance pre It means that all iterative rounds of local updates are completed in advance during the timing constraint phase of the timing query mechanism, and the local model parameters are delivered to the client of the parameter server in a timely manner.

[0041] In the active inquiry phase, a minimum iteration round threshold is imposed as a constraint, and the iteration round is used as a basic criterion for screening model quality to determine whether to truncate the client's current local update status to obtain the qualification to upload supplements to the parameter server.

[0042] Under the action of the timing query mechanism, after receiving the aggregated subset in the timing constraint phase, the parameter server actively queries the parameter server to obtain the candidate subset that can continue to be supplemented to the parameter server in the current interaction round, that is, the truncated client model that meets the iteration round limit.

[0043] In the embodiment of the present invention, in federated learning, the interaction duration of each interaction round directly affects the learning efficiency of the global aggregation model, mainly including the client local update duration and the model parameter transmission duration.

[0044] For example, the timing query mechanism diagram is as follows: Figure 2 shown.

[0045] For the timing constraint phase, in the R-1th round of interaction, some clients (the model in the device is in the Updatebegin state) participated in the aggregate update of the R-1th round. In the Rth round, the local model needs to restart the update based on the received global model parameters, that is, e=0. However, some clients (the model in the device is in the Partial begin state) did not participate in the aggregate update of the R-1th round. The local model will continue to update in the Rth round based on the updated state, that is, e>0. Figure 2 As shown in (a) in the interactive round R, the client updates the local model at a variable training speed due to the performance of its own device, which reflects the impact of unstable device performance on the local update duration, such as Figure 2 As shown in (b). Figure 2 Between (b) and (c), due to network influence, only some clients successfully upload the local model parameters to the parameter server after completing the local update. This reflects the impact of low network quality and unstable network environment on the model parameter transmission time of some clients.

[0046] For the active inquiry phase, Figure 2 As can be seen in (c), after the parameter server receives certain clients at a fixed time, it does not immediately perform a global aggregate update. The present invention allows the parameter server as the inquiring party to continue to send requests to the remaining clients with supplementary query information, so that the remaining clients can truncate the current update state and upload the current model parameters to further obtain a local model with certain reference information. When the parameter server sends the supplementary query information, consider setting a minimum iteration round threshold E for the local iteration round ctrl , each client according to E ctrl It determines whether the current update state can be truncated by itself, thereby obtaining a candidate subset that can continue to be supplemented to the parameter server in the current interaction round, that is, a truncated client model that meets the iteration round limit.

[0047] Step 2: Aiming at the problem of heterogeneous aggregation, a screening and supplementation strategy is designed to continue screening the truncated clients that meet the iteration round limit in the active inquiry phase from the perspective of data combination. Within the communication constraints, the parameter server receives the second part of the aggregated subset.

[0048] In this step, the screening and supplementation strategies are designed to address the problem of heterogeneous polymerization, including:

[0049] Within the communication constraints, from the perspective of data combination, the truncated clients that meet the iteration round limit in the active query phase are further screened according to the screening index, and the aggregated subset received by the parameter server is further expanded, that is, the truncated supplementary client C trunc and communication supplement client C gap The second aggregate subset composed of the two aggregate subsets is the second part of the entire aggregate subset.

[0050] Among them, the truncated supplementary client C trunc Refers to clients whose local updates are not fully completed but are truncated for upload and supplementation in the current round due to meeting the iteration round limit and screening supplement conditions;

[0051] Communication Supplement Client C gap It refers to the client that has completed the fixed iteration round requirements in the timed query phase, but is still on the communication link due to insufficient waiting time in the timed query phase. Through the subsequent screening and supplement phase, it gets another upload opportunity to the parameter server during the waiting time.

[0052] Exemplary, a schematic diagram of the screening supplement strategy is shown in Figure 3 shown.

[0053] After the timing constraint, some clients have uploaded their model parameters to the parameter server, and the remaining clients are Figure 3 After the parameter server sends active query information to the remaining clients in Step 1, each remaining client starts to execute the iteration round limit process in Step 2. Each client will be based on the minimum iteration round threshold E ctrl Determine whether to truncate the current update state, thereby obtaining a candidate subset with better model quality, such as Figure 3 As shown in (b) in the figure, for the aggregated subset in (b), we continue to filter the clients from the perspective of data combination and combined with the filtering and supplementing strategy, thus obtaining Figure 3 After the parameter server actively inquires and screens the supplement strategy, the remaining clients are screened in turn from the model quality and data quality to obtain the final supplementable aggregate subset. In addition, during the supplement upload process, the parameter server obtains two different types of supplement clients, namely, truncated supplement clients Ctrunc and communication supplement client C gap .

[0054] Step 3: Under the interaction of the timed query mechanism and the screening and supplementation strategy, through two-stage client reception, the parameter server receives the complete aggregated subset for global model aggregation update and synchronizes the updated model to the client. No information exchange is performed between clients.

[0055] In this step, under the interaction of the timing query mechanism and the screening supplement strategy, through two stages of client reception, the parameter server receives a complete aggregated subset, including:

[0056] Based on the timed query, the client of the aggregated subset is converted from one-stage reception to two-stage reception and supplementation. Through the two-stage client reception, the parameter server receives the complete aggregated subset;

[0057] A topology architecture centered on a parameter server is adopted. A global aggregate model is obtained through multiple rounds of communication interactions between the aggregate model update of the parameter server and the local model of the client, and the updated model is synchronized to the client. There is no communication interaction between the clients.

[0058] Furthermore, in the above steps 1, 2 and 3, the parameter server sends query information, and the query information includes a minimum iteration round threshold and a screening indicator for screening the supplement strategy.

[0059] In the embodiment of the present invention, given a client set C, a timing constraint duration t1, and a supplementary constraint duration t2, the parameter server receives an aggregated subset S1=C within a limited interaction duration T=t1+t2. pre and aggregate subset S2 = C trunc ∪C gap In the process of updating the global model, the global model learning efficiency (Global model update efficiency, GME) under the federated learning framework is optimized by constraining the interaction time T and optimizing the quality of the aggregated subset Q. The formula is as follows:

[0060] max(GME(T,Q))

[0061] Under the interaction of the timed query mechanism and the screening and supplementation strategy, the two aggregate subsets received by the parameter server together constitute the complete aggregate subset that can be used to aggregate and update the global model in the interaction round R, which can be described as the following formula:

[0062]

[0063] in, represents the model parameters of the kth client of the received aggregate subset in the Rth interaction round, Represents C trunc The client set composed of k C pre and C gap The client set, E ctrl represents the lowest threshold of the iteration round limit, e represents the actual iteration round in which the client performs local updates in the interaction round R, and E represents the fixed iteration round of local updates.

[0064] Furthermore, the parameter server performs global model aggregation updates based on the above aggregation subsets, and the formula is as follows:

[0065]

[0066] in, represents the global model parameters after the aggregation update of the Rth interaction round, m represents the number of clients in the aggregated subset, n represents the total number of samples contained in the aggregated subset, and n k Represents the number of samples contained in the local data set of client k.

[0067] An exemplary overall framework description of an efficient asynchronous federated training method for edge computing is shown in the figure below: Figure 4 shown.

[0068] Step 1. Initialize: Before starting federated learning, the parameter server will initialize the local model of all clients through Step 1, and each client is divided into three types of clients due to its own performance differences (such as fast, medium and slow clients). Figure 4 In each interaction round, each client will randomly switch between three clients with different speeds.

[0069] Step 2. Local update: In the timed query mechanism, each client updates the model at a different local training speed. It can be divided into two categories according to different update states. One category indicates that the client has completed all iterations of local updates (such as Figure 4 (c) in the figure), and this type of client can upload local model parameters to the parameter server; the other type indicates that the client is still in the local update process (such as Figure 4 As shown in (d) in the figure, further updates are needed.

[0070] Step 3.Uploadparameter: After the waiting time is over, the parameter server will receive the first part of the aggregated subset (such as Figure 4As shown in (e), under the time constraint of the timing query mechanism, some clients will be constrained due to the long local training time or long communication time.

[0071] Step 4. Setthreshold: Set additional query information, such as the minimum iteration round threshold E ctrl and screening indicators proposed by the screening supplement strategy.

[0072] Step 5. Inquiry: The parameter server is transformed from a passive receiver to an active inquiring party, sending inquiry signals to the remaining clients, looking for clients that can temporarily pause the local training process and directly upload the current model parameters.

[0073] Step 6.Epoch filter: By iterating the round limit, the candidate supplement subset with supplement qualification is obtained, that is, Figure 4 The client subset shown in (g).

[0074] Step 7. Data evaluation: On this basis, continue to filter the clients in (g) by combining the screening and supplementary strategies, and obtain Figure 4 The client subset shown in (h).

[0075] Step 8.Supplement: Within the communication constraints, the parameter server receives the second aggregate subset consisting of the truncated supplementary clients and the communication supplementary clients, such as Figure 4 As shown in (j), it is combined with the first part of the aggregation subset to become the entire aggregation subset participating in the aggregation update in this interaction round.

[0076] In summary, the present invention provides an efficient asynchronous federated training method for edge computing, whose core ideas mainly include two aspects: a timed inquiry mechanism and a screening and supplement strategy. For the long-tail problem of interaction, the time constraint and inquiry supplement of the timed inquiry mechanism are used to effectively constrain the duration of collaborative interaction on the premise of expanding the scale of the aggregated subset. For the problem of aggregate heterogeneity, on the basis of timed inquiries, a screening and supplement strategy is proposed based on the idea of ​​optimizing the data combination quality in the aggregated subset, and the clients that can be supplemented in the timed inquiry process are further screened to optimize the combination quality of the aggregated subset in terms of model quality and data quality. Therefore, the method of the present invention can effectively improve the interaction efficiency and convergence efficiency of global model updates under the federated learning framework.

[0077] It should be noted that the references to "one embodiment", "embodiment", "exemplary embodiment", "some embodiments" and the like in the specification indicate that the embodiments described may include specific features, structures or characteristics, but not every embodiment may include the specific features, structures or characteristics. In addition, when a specific feature, structure or characteristic is described in conjunction with an embodiment, it should be within the knowledge of a person skilled in the art to implement such feature, structure or characteristic in conjunction with other embodiments (whether or not explicitly described).

[0078] In general, a term can be understood, at least in part, from its use in context. For example, depending, at least in part, on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending, at least in part, on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0079] As used herein, the term "nominal / nominal" refers to an expected or target value for a characteristic or parameter of a component or process operation set during the design phase of a production or manufacturing process, as well as a range of values ​​above and / or below the expected value. The range of values ​​may be due to slight variations in the manufacturing process or tolerances. As used herein, the term "approximately" indicates a value of a given quantity that may vary based on a particular technology node associated with the subject semiconductor device. Based on a particular technology node, the term "approximately" may indicate a value of a given quantity that varies, for example, within 5%-15% of the value (e.g., ±5%, ±10%, or ±15% of the value).

[0080] It will be understood that the meaning of “on,” “over,” and “above” in this disclosure should be interpreted in the broadest manner, so that “on” means not only “directly on” something, but also includes the meaning of being “on” something with intervening features or layers therebetween, and “on” or “over” means not only “on” or “above” something, but also includes the meaning of being “on” or “above” something with no intervening features or layers therebetween.

[0081] Additionally, spatially relative terms such as "under," "beneath," "lower," "above," "upper," and the like may be used herein for descriptive convenience to describe the relationship of one element or feature to another element or features, as shown in the accompanying drawings. Spatially relative terms are intended to encompass different orientations of the device in use or operation in addition to the orientation depicted in the accompanying drawings. The device may be oriented in other ways, and the spatially relative descriptors used herein may be similarly interpreted accordingly.

[0082] The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, but those skilled in the art can fully understand the present invention without the description of these details. In addition, in order to avoid unnecessary confusion about the essence of the present invention, well-known methods, processes, procedures, components and circuits are not described in detail.

[0083] A person skilled in the art will appreciate that all or part of the steps in the above-mentioned embodiment method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc.

[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An efficient asynchronous federated training method for edge computing, characterized in that: The following steps are involved: To address the long-tail interaction problem, a timed query mechanism is designed to control the interaction duration from the perspective of time constraints and query supplements. Within the time constraints, the parameter server receives the first part of the aggregated subset. To address the long-tail interaction problem, a timed query mechanism is designed, including: For the timing constraint phase, the parameter server faces all clients in the entire interaction round and waits for any reachable client through the timing constraint. Within the time constraint, the parameter server receives the first client C that has arrived in advance. pre The aggregate subset composed of is the first part of the entire aggregate subset; In the active inquiry phase, the parameter server is transformed from a passive receiver to an active inquirer, and continues to send requests to other clients with supplementary inquiry information. When the client's local model meets the query requirements of the parameter server, it can choose to truncate the current local update status and directly upload the local model parameters; Among them, the pre-reached client C pre It means that all iterations of local updates are completed in advance during the timing constraint phase of the timing query mechanism, and the local model parameters are delivered to the client of the parameter server in a timely manner; Aiming at the problem of heterogeneous aggregation, a screening and supplementation strategy is designed to continue screening the truncated clients that meet the iteration round limit in the active inquiry phase from the perspective of data combination. Within the communication constraints, the parameter server receives the second part of the aggregate subset. The above-mentioned screening supplementary strategy is designed to solve the problem of polymerization heterogeneity, including: Within the communication constraints, from the perspective of data combination, the truncated clients that meet the iteration round limit in the active query phase are further screened according to the screening index, and the aggregated subset received by the parameter server is further expanded, that is, the truncated supplementary client C trunc and communication supplement client C gap The second aggregate subset composed of the above is the second part of the entire aggregate subset; Among them, the truncated supplementary client C trunc Refers to clients whose local updates are not fully completed but are truncated for upload and supplementation in the current round due to meeting the iteration round limit and screening supplement conditions; Communication Supplement Client C gap It refers to the client that has completed the fixed iteration round requirements in the timed query phase, but is still on the communication link due to insufficient waiting time in the timed query phase. In this way, it can obtain another upload opportunity to the parameter server during the waiting time through the subsequent screening and supplementation phase; Under the interaction of the timed query mechanism and the screening and supplementation strategy, through two-stage client reception, the parameter server receives the complete aggregated subset for global model aggregation update and synchronizes the updated model to the client, without information exchange between clients.

2. The efficient asynchronous federated training method for edge computing according to claim 1, characterized in that: In the active inquiry phase, a minimum iteration round threshold is imposed as a constraint, and the iteration round is used as a basic criterion for screening model quality to determine whether to truncate the client's current local update status to obtain the qualification to upload supplements to the parameter server.

3. The efficient asynchronous federated training method for edge computing according to claim 1, characterized in that: Under the action of the timing query mechanism, after receiving the aggregated subset in the timing constraint phase, the parameter server actively queries through the parameter server to obtain the candidate subset that can continue to be supplemented to the parameter server in the current interaction round, that is, the truncated client model that meets the iteration round limit.

4. The efficient asynchronous federated training method for edge computing according to claim 1, characterized in that: Under the interaction of the timing query mechanism and the screening supplement strategy, through the two-stage client reception, the parameter server receives a complete aggregated subset, including: Based on the timed query, the client of the aggregated subset is converted from one-stage reception to two-stage reception and supplementation. Through the two-stage client reception, the parameter server receives the complete aggregated subset; A parameter server-centric topology is adopted. The global aggregate model is obtained through multiple rounds of communication interactions between the aggregate model update of the parameter server and the local model of the client, and the updated model is synchronized to the client. There is no communication interaction between the clients.

5. The efficient asynchronous federated training method for edge computing according to claim 1, characterized in that: The parameter server sends query information, which includes a minimum iteration round threshold and a screening indicator for a screening supplement strategy.

Images