A joint method for d2d pairing and device scheduling for federated edge learning

Abstract

The application discloses a kind of D2D pairing and equipment scheduling of federal edge learning-oriented joint method, belong to the equipment participation and data sharing technical field under federal edge learning environment.This application proposes a D2D communication federal edge learning architecture, mainly concerns the association mode and entity scheduling strategy of D2D device pair, focuses on the influence of communication packet loss error rate and the data heterogeneous characteristics of mobile device on the convergence performance of model, aims at improving the problem of device access quantity limit caused by limited base station spectrum resources.This method introduces equipment association scheduling strategy, enhances the flexibility of federal edge learning system to fluctuating communication conditions;By introducing fairness constraints, ensure that each device has the opportunity to participate in training, so that the data learned by the global model is more comprehensive;By introducing D2D communication to increase the access amount of base station, improve the performance of edge federal learning;Under the condition of meeting the limit of edge base station receiving capacity, the global loss is minimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of device participation and data sharing in a federal edge learning environment, in particular to a joint method for D2D pairing and device scheduling for federal edge learning. BACKGROUND

[0002] Thanks to the rapid development of 5G / IoT technology, the number of Internet of Things (IoT) devices is growing exponentially. With more and more IoT devices accessing the Internet of Things, a large amount of data is generated at the edge of the wireless network. In order to utilize these valuable data resources, traditional machine learning uploads them to a central server for training. However, this process can lead to the leakage of device privacy data and cause adverse effects of data misuse, so a federal edge learning system deployed at the edge of the network is proposed to overcome the above problems, which allows edge devices participating in model training to collaboratively train a global model without exposing local data. Unfortunately, the deployment of federal edge learning currently faces two practical challenges: limited spectrum resources and data heterogeneity. Specifically, due to the constraints of concurrent processing capacity and spectrum resources, edge base stations cannot simultaneously support the transmission services of a large number of IoT devices, and can only allow a small number of devices to access in a single round, thereby greatly limiting the extent to which devices participate in global model training and negatively affecting model convergence performance. In addition, the local data sets of large-scale distributed devices usually have statistical heterogeneity and high redundancy, which can easily lead to slow model convergence or even model drift, especially when only a limited number of devices participate in each training round. Existing research considers increasing the number of devices participating in training in each round through Device-to-Device (D2D) communication to improve model convergence performance while alleviating the access constraints of base stations, but due to the influence of unreliable communication environment, there are potential data packet loss and device availability constraints during model transmission, and the unbalanced characteristics of local data sets between devices also affect the D2D association and scheduling strategy of devices in each training round. Therefore, it is necessary to design a safe and reliable D2D association and scheduling scheme that guarantees learning performance in a federal edge learning system under complex communication environment.

[0003] Therefore, how to solve the above-mentioned problems has become a difficult problem that needs to be solved urgently. SUMMARY

[0004] The present application aims to provide a joint method for D2D pairing and device scheduling for federal edge learning, which aims to improve the problem of limited device access caused by limited spectrum resources of base stations, increase the number of devices participating in each training round through the establishment and access of D2D device pairs, thereby reducing the influence of model bias caused by data heterogeneity and improving the robustness and convergence speed of the model.

[0005] To achieve the above object, the present application provides the following technical solutions:

[0006] A joint method for D2D pairing and device scheduling in a federated edge learning,

[0007] The joint method for D2D pairing and device scheduling in a federated edge learning has the following steps:

[0008] S1: Use a federated edge learning system for federated learning, which is a global model formed by cooperative training of an edge base station with limited wireless resources and a plurality of devices, the device set is C={1,2,…,C}, C=|C|, the base station can simultaneously establish a connection with N devices; the device set directly communicating with the base station in the tth round is S t , two said device groups form a D2D device pair set P t , wherein each said D2D device pair contains a device B i,t communicating with the base station and a device D i,t communicating with the B i,t , for i∈S t ∪P t , scheduling strategy a i,t ∈[0,1], a i,t =1 indicates that entity i participates in training at training round t, otherwise it does not participate;

[0009] The D2D device communication and the direct communication of the device to the base station both use orthogonal frequencies to transmit in quasi-static Rayleigh fading; each said device encapsulates model parameters into a single data packet for transmission; the base station verifies the integrity of the data packet by cyclic redundancy check, and if it is shown to be damaged or lost, the packet is discarded;

[0010] Before each round of federated learning, the packet loss rate of all devices directly communicating with the base station and the communication packet loss rate of all feasible D2D associated device pairs are calculated respectively, and an availability constraint index is introduced to represent whether the local model scheduling strategy of the device can be successfully uploaded;

[0011] S2: Considering the negative impact of device data heterogeneity on model training, a virtual queue is introduced for each device during each round of device selection to achieve fairness constraints;

[0012] S3: The data quality, packet loss rate, and virtual queue are comprehensively considered as a measurement standard as the basis for device pairing and selection, and the optimization problem is modeled as a maximum weight independent edge set problem with constraints;

[0013] S4: D2D device pairing and selection is performed in each round of training by the two-part graph maximum weight matching algorithm, the selected D2D device pairs and the direct communication device upload the local updated model to the base station for global aggregation, and the new global model is broadcast to the selected devices for the next round of training until the model converges.

[0014] Preferably, in S1, the packet loss rate of all devices communicating directly with the base station and the packet loss rate of all possible D2D associated device pairs are as follows:

[0015] The packet loss rate of device i communicating directly with the base station is represented as:

[0016]

[0017] where P i,B , h i,B is the transmission power and channel gain of device i communicating directly with the base station, W B is the allocated communication bandwidth, N0 is the noise power spectral density, and τ is a threshold parameter specific to the system.

[0018] The packet loss rate of D2D device communication is:

[0019]

[0020] where P i,D , h i,D is the transmission power and channel gain of device i for D2D communication, W D is the dedicated bandwidth for D2D communication, and τ' is a threshold parameter related to the reverse coding gain of D2D communication.

[0021] Define the model transmission indicator C i,B from device i to the base station, C i,B ∈{0,1}, when C i,B =1 indicates that the local model of device i is uploaded successfully, otherwise C t =0, which can be represented as:

[0022]

[0023] For D2D device pair i∈P t , introduce an indicator C i,D ∈{0,1}, when C i,D =1 indicates that the D2D model transmission is successful, otherwise C i,D =0, which can be represented as:

[0024]

[0025] Preferably, in S2, considering that the data heterogeneity of devices will have a negative impact on the efficiency of model training, the fairness constraint is introduced:

[0026]

[0027] where T represents the total training rounds, c i ∈[0,1) represents the proportion value of the minimum number of communication rounds required to select device i compared to the total training rounds; in order to solve the above fairness constraint, a virtual queue D i is introduced for device i∈C

[0028] D i (t)=[D i (t-1)+c i -a i,t-1 ] + .

[0029] where D i (t) represents the queue length at the beginning of the tth round, [x] + =max{x,0}, and the global device is represented as D(t)=[D1(t),...,D C (t)], D(1)=0.

[0030] Preferably, in S3, by optimizing the D2D association and scheduling strategy, an optimization problem is formulated to minimize the global loss, including the following steps:

[0031] F1: the federated learning learning rate satisfies Given the device scheduling matrix a t and the optimal global federated learning model w * , the convergence upper bound is described as:

[0032]

[0033] where, in order to meet the above global training requirements, the local loss function of device training satisfies ζ1 and ζ2 are non-negative constants, the global loss function L(w) is quadratic and continuously differentiable, and satisfies L and μ are non-negative constants of Lipschitz continuity and strong convexity, respectively, and

[0034] F2: in order to minimize the convergence upper bound, only the interval is needed to be minimized. Rewrite A as Rewrite B as Further, the optimization problem is represented as:

[0035]

[0036] Where T is the set of total training rounds, and N is the maximum number of devices that the base station can access at any given time;

[0037] F3: Model the optimization problem as a graph theory problem, where devices are considered vertices and potential matchings between devices are considered edges. Therefore, the optimization problem can be simplified to a constrained maximum weighted independent edge set problem, which maximizes the total weight by finding a set of edges that do not share vertices.

[0038] Preferably, in S3, all devices participating in the training use standard gradient descent for local training; by introducing a scheduling strategy and considering packet loss rate, the update rule for the global model w can be described as:

[0039]

[0040] Among them, K i,B K i,D and K j Representing device B respectively i,t D i,t And the size of j's local dataset, w i,t w j,t Let i and j represent the local model parameters of the D2D device pair i and the direct communication device j from round t, respectively.

[0041] Preferably, in S4, the steps for solving the above maximum weighted independent edge set problem are as follows:

[0042] Step 1): Transform device i∈C into node u i To construct a general graph G(U) from ∈U, define the vertex weights of device i:

[0043] V i∈St =βK i (1-q i,B )+(1-β)D i (t)

[0044] Where β∈[0,1]; without loss of generality, the edge weight of the D2D device for i is defined as:

[0045] V i∈Pt =β(1-q) i,B (K) i,B +K i,D (1-q i,D ) 2 )+(1-β)(D i,B (t)+D i,D (t))

[0046] Among them, D i,B (t) and Di,D (t) respectively represents the virtual queue of the device B i,t and D i,t ;

[0047] Step 2): Perform symmetric operation on the nodes contained in U to obtain a new node set, denoted as V, and construct a bipartite graph G(U, V), for u i ∈U,v j ∈V, establish an edge e(i, j), realize the maximum matching M using the bipartite graph maximum weight matching algorithm, and sort the edge weight of M in descending order, and take the top-N elements to meet the base station access device quantity constraint.

[0048] Compared with the prior art, the beneficial effects of the present application are:

[0049] (1) The method considers the problem of data packet loss in the model transmission process caused by poor channel conditions and exhausted computing power, introduces a device association scheduling strategy, and enhances the flexibility of the federated edge learning system to fluctuating communication conditions.

[0050] (2) The method takes into account the negative impact of the data heterogeneity of the mobile device local data set on the efficiency performance of model training, and by introducing fairness constraints, ensures that each device has the opportunity to participate in training, making the data learned by the global model more comprehensive.

[0051] (3) The method jointly considers the limited number of base station access in the edge environment, aims to increase the number of base station access by introducing D2D communication to improve the performance of edge federated learning. Secondly, under the condition of meeting the edge base station receiving capacity limit, the global loss is minimized, that is, through D2D pairing and device scheduling, superior learning performance is achieved in complex communication environment. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is the system model framework diagram of the present application;

[0053] Figure 2 is the overall flowchart of the present application; DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0055] As Figure 1As shown, the device D2D association scheduling method for federated edge learning provided by the application first proposes a federated edge learning architecture based on D2D communication to enhance the federated learning mechanism, and through the association scheduling of D2D device pairs by considering factors such as transmission packet loss rate and data heterogeneity before starting each round of training, the reliability of the overall device participating in training is maximized.

[0056] Embodiment one

[0057] As Figure 2 shown, the D2D pairing and device scheduling joint method for federated edge learning provided by the application, the federated edge learning system includes an edge base station with limited wireless resources and a global model formed by a large number of device collaborative training, the device set is C={1,2,...,C}, C=|C|, the base station can at most simultaneously establish connection with N devices, S t and P t respectively represent the device set directly communicating with the base station in the tth round and the D2D device pair set composed of two devices, wherein each D2D device pair contains a device B i,t communicating with the base station and a device D i,t communicating with B i,t , for i∈S t ∪P t , the scheduling strategy a i,t ∈[0,1], a i,t =1 indicates that entity i participates in training at training round t, otherwise it does not participate. The specific steps are as follows:

[0058] Step one, before each round of federated learning, calculate the packet loss rate of all devices directly communicating with the base station and the communication packet loss rate of all feasible D2D associated device pairs, and introduce an availability constraint index to represent whether the local model scheduling strategy of the device can be uploaded successfully.

[0059] Step two, considering the negative impact of device data heterogeneity on model training, a virtual queue is introduced for each device during each round of device selection to realize fairness constraint.

[0060] Step three, consider data quality, packet loss rate, and virtual queue as a comprehensive measure, as the basis for device pairing and selection, and model the optimization problem as a maximum weight independent edge set problem with constraints.

[0061] Step four, D2D device pairing and selection are performed through the maximum weight matching algorithm of bipartite graph, and the selected D2D device pairs and directly communicating devices upload the local update model to the base station for global aggregation, and the new global model is broadcast to the selected devices for the next round of training until the model converges.

[0062] In step one of the embodiment, both D2D communication and device-to-base station direct communication adopt orthogonal frequency transmission in quasi-static Rayleigh fading, and there is no interference between devices. Each device encapsulates model parameters into a single data packet for transmission, and the base station verifies the integrity of the data packet through cyclic redundancy check. If it shows damage or loss, the message is discarded. The packet loss rate of device i communicating directly with the base station is represented as:

[0063]

[0064] where P i,B , h i,B are the transmission power and channel gain of device i communicating directly with the base station, W B is the allocated communication bandwidth, N0 is the noise power spectral density, and τ is a threshold parameter specific to the system.

[0065] Similarly, the packet loss rate of D2D device communication is:

[0066]

[0067] where P i,D , h i,D are the transmission power and channel gain of device i for D2D communication, W D is the dedicated bandwidth for D2D communication, and τ' is a threshold parameter related to the reverse coding gain of D2D communication.

[0068] Further, define the device-to-base station model transmission indicator C i,B ∈{0,1}, where C i,B =1 indicates that the local model upload of device i is successful, otherwise C i,B =0. This indicator can be represented as:

[0069]

[0070] Similarly, for D2D device pair i∈P t , introduce the indicator C i,D ∈{0,1}, where C i,D =1 indicates that the D2D model transmission is successful, otherwise C i,D =0. This indicator can be represented as:

[0071]

[0072] In step two of the embodiment, considering that the data heterogeneity of devices will have a negative impact on the efficiency of model training, a fairness constraint is introduced

[0073]

[0074] where T denotes the total training rounds, c i ∈ [0, 1) denotes the ratio value of the minimum number of communication rounds required to select device i compared to the total training rounds. The normalized product of the local dataset size and the upload success rate is used as a measure to describe the importance of the device, so where v i denotes the weight of device i. In order to solve the above fairness constraint, a virtual queue D i is introduced for device i ∈ C

[0075] D i (t) = [D i (t - 1) + c i -a i,t-1 ] +

[0076] where D i (t) denotes the queue length at the beginning of the t-th round, [x] + = max{x, 0}, and the global device is denoted as D(t) = [D1(t),..., D C (t)], D(1) = 0.

[0077] In step three of the embodiment, by optimizing the D2D association scheduling strategy, an optimization problem is formulated to minimize the global loss. The steps include:

[0078] Step 1-1): Assuming that the federal learning learning rate satisfies Given the device scheduling matrix a t and the optimal global federal learning model w * , the convergence upper bound can be described as

[0079]

[0080] where, in order to meet the above global training requirements, the local loss function of device training satisfies ζ1 and ζ2 are non-negative constants, and the global loss function L(w) is quadratic and continuously differentiable, satisfying L and μ are non-negative constants of Lipschitz continuity and strong convexity, respectively, and

[0081] Step 1-2): In order to minimize the above convergence upper bound, it is only necessary to minimize the interval

[0082] Rewrite A as Rewrite B as Further, the optimization problem can be expressed as:

[0083]

[0084] where T is the set of total training rounds, and N is the maximum limit of the number of devices that a base station can access at any given time.

[0085] Step 1-3): Model the problem as a graph theory problem. Where devices are considered as vertices, and potential matches between devices are considered as edges. Thus the optimization problem can be simplified as a maximum weighted independent edge set problem with constraints, by finding a set of edges that do not share vertices, maximizing the total weight.

[0086] In step three of this embodiment, during the training process, all participating devices use the standard gradient descent method for local training. By introducing a scheduling strategy, the update rule of the global model w can be described as

[0087]

[0088] where K i,B , K i,D and K j represent the size of the local data set of devices B i,t , D i,t and j respectively, w i,t , w j,t represent the local model parameters of D2D device pair i and direct communication device j from the t-th round respectively.

[0089] In step four of this embodiment, the steps to solve the above maximum weighted independent edge set problem can be summarized as follows:

[0090] Step 2-1): Convert device i∈C to node u i ∈U to construct a general graph G(U), define the vertex weight of device i as

[0091] V i∈St = βK i (1-q i,B )+(1-β)D i (t)

[0092] where β∈[0,1], K i represents the size of the local data set of device i. Without loss of generality, the edge weight of D2D device pair i is defined as

[0093] V i∈Pt = β(1-q i,B )(K i,B +K i,D (1-q i,D ) 2 )+(1-β)(D i,B (t)+D i,D (t))

[0094] wherein D i,B (t) and D i,D (t) represent the virtual queues of the devices B i,t and D i,t respectively.

[0095] Step 2-2): Perform symmetric operation on the nodes contained in U, obtain a new node set, denoted as V, construct a bipartite graph G(U, V), for u i ∈U, v j ∈V, establish an edge e(i, j), use the KM algorithm to realize the maximum matching M, and perform edge weight decreasing sorting on M, take the top-N elements to meet the base station access device quantity constraint.

[0096] Finally, it should be noted that: the above is only the preferred embodiments of the present application, and is not intended to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, it still can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A joint method for D2D pairing and device scheduling oriented towards federated edge learning, characterized in that: The steps of the joint D2D pairing and device scheduling method for federated edge learning are as follows: S1: Federated learning is performed using a federated edge learning system. This system is a global model formed by a wirelessly resource-limited edge base station and several devices collaboratively trained. The device set is C = {1, 2, ..., C'}, where C' = |C|. The edge base station can simultaneously establish connections with N devices. In round t, the set of devices directly communicating with the edge base station is S. t The two devices constitute a D2D device pair set P. t Each of the D2D devices includes device B that communicates with the edge base station. i,t And the aforementioned B i,t Communication device D i,t For i∈S t ∪P t Scheduling strategy a i,t ∈[0,1],a i,t =1 indicates that device i participated in training at training round t; otherwise, it did not participate. The device i is a pair of D2D devices. Both D2D device communication and direct communication from the device to the edge base station use orthogonal frequencies for transmission in quasi-static Rayleigh fading; each device encapsulates model parameters into a single data packet for transmission; the edge base station verifies the integrity of the data packet through cyclic redundancy check, and discards the packet if it is found to be corrupted or lost. Before each round of federated learning, the packet loss rate of all devices communicating directly with the edge base station and the packet loss rate of all feasible D2D device pairs are calculated respectively. An availability constraint index is introduced to indicate whether the local model scheduling strategy of the device can be successfully uploaded. S2: Considering the negative impact of heterogeneous device data on model training, a virtual queue is introduced for each device in each round of device selection to achieve fairness constraints; S3: Data quality, packet loss rate, and virtual queues are taken into account as a comprehensive evaluation standard and used as the basis for equipment pairing and selection. The optimization problem is modeled as a constrained maximum weight independent edge set problem. S4: In each round of training, D2D devices are paired and selected using the bipartite graph maximum weight matching algorithm. The selected D2D device pairs and direct communication devices upload the locally updated model to the edge base station for global aggregation, and broadcast the new global model to the selected devices for the next round of training until the model converges.

2. The joint method for D2D pairing and device scheduling for federated edge learning according to claim 1, characterized in that: In S1, the packet loss rates for all devices communicating directly with the edge base station and the packet loss rates for all feasible D2D device pairs are as follows: Packet loss rate q when device i communicates directly with the edge base station i,B Represented as: Among them, P i,B h i,B W represents the transmit power and channel gain for direct communication between device i and the edge base station. B Where N is the allocated communication bandwidth, N0 is the noise power spectral density, and τ is a system-specific threshold parameter. Indicates the channel gain h i,B The mathematical expectation; The packet loss rate in D2D device communication is: Among them, P i,D h i,D For device i, the transmit power and channel gain for D2D communication are set by W. D τ′ is the dedicated bandwidth for D2D communication, and it is a threshold parameter related to the inverse coding gain of D2D communication. Indicates the channel gain h i,D The mathematical expectation; Define the model transmission metric C from the device to the edge base station. i,B ∈{0,1}, when C i,B =1 indicates that the local model of device i was successfully uploaded; otherwise, C i,B =0, this indicator can be expressed as: For device i∈P t Introducing indicator C i,D ∈{0,1}, when C i,D =1 indicates successful D2D model transfer; otherwise, C i,D =0, this indicator can be expressed as:

3. A joint method for D2D pairing and device scheduling for federated edge learning according to claim 1, characterized in that: In S2, considering that the heterogeneous nature of device data can negatively impact the efficiency of model training, the aforementioned fairness constraint is introduced: Where T represents the total number of training rounds, c i ∈ [0,1) represents the ratio of the minimum number of communication rounds required to select device i to the total number of training rounds; to address the aforementioned fairness constraint, a virtual queue D is introduced for device i∈C. i ,Right now: D i (t)=[D i (t-1)+c i -a i,t-1 ] + Among them, D i (t) represents the queue length at the start of round t, [x] + =max{x,0}, global device representation is D(t)=[D1(t),…,D C' [(t)], D(1)=0.

4. A joint method for D2D pairing and device scheduling for federated edge learning according to claim 1, characterized in that: In S3, by optimizing D2D association and scheduling strategies, an optimization problem is formulated with the goal of minimizing global loss, including the following steps: F1: Federated learning rate meets Given a device scheduling matrix a t and the optimal global federated learning model w * The upper bound of convergence is described as follows: Among them, L(w t+1 L(w0) represents the loss function in round t+1, and L(w0) represents the initial loss function; to meet the global training requirements, the local loss function of the device training satisfies ζ1 and ζ2 are nonnegative constants. Let be the gradient of the loss function at round t. The global loss function L(w) is quadratically differentiable and satisfies the following condition: L and μ are nonnegative constants for Lipschitz continuity and strong convexity, respectively, and I is the identity matrix. K i,B Indicates device B i,t The size of the local dataset, K i,D Indicates device D i,t The size of the local dataset, q i,B q represents the packet loss rate during direct communication between device i and the edge base station; i,D This indicates the packet loss rate in D2D device communication; F2: To minimize the convergence upper bound, it is only necessary to minimize the interval. Rewrite A as Rewrite B as Furthermore, the optimization problem can be expressed as: Where T is the set of total training rounds, and N is the maximum number of devices that an edge base station can access at any given time; F3: Model the optimization problem as a graph theory problem, where devices are considered vertices and potential matchings between devices are considered edges. Therefore, the optimization problem can be simplified to a constrained maximum weighted independent edge set problem, which maximizes the total weight by finding a set of edges that do not share vertices.

5. A joint method for D2D pairing and device scheduling for federated edge learning according to claim 1, characterized in that: In S3, all devices participating in the training use standard gradient descent for local training. By introducing a scheduling strategy and considering packet loss rate, the update rule for the global model w can be described as: Among them, K i,B K i,D and K j Representing device B respectively i,t D i,t And the size of j's local dataset, w i,t w j,t Let C represent the local model parameters from device i and direct communication device j in round t, respectively. i,B Define the model transmission metrics from the device to the edge base station, C i,D This indicates the transmission index of the D2D model.

6. A joint method for D2D pairing and device scheduling for federated edge learning according to claim 1, characterized in that: In S4, the steps to solve the above maximum weight independent edge set problem are as follows: Step 1): Transform device i∈C into node u i To construct a general graph G(U), where U is the set of nodes, we define the vertex weights of device i: Where β∈[0,1]; without loss of generality, the edge weight of the D2D device for i is defined as: Among them, K i,B K i,D Representing device B respectively i,t D i,t The size of the local dataset, D i,B (t) and D i,D (t) represent equipment B respectively. i,t and D i,t virtual queue, q i,B and q i,D These represent the packet loss rate of device i in direct communication with the edge base station and the packet loss rate of D2B communication, respectively. Step 2): Perform symmetric operations on the nodes contained in U to obtain a new set of nodes, denoted as V. Construct a bipartite graph G(U,V). For u i ∈U,v j For each edge e(i,j) in V, the maximum matching M is achieved using the bipartite graph maximum weight matching algorithm. The edge weights of M are sorted in descending order. N is the maximum number of devices that the edge base station can access at any given time, in order to meet the constraint on the number of devices accessed by the edge base station.

Images