A Trusted Offloading System and Method for Federated Learning Tasks in an Edge-Cloud Collaborative Environment

Abstract

The present invention relates to a method and system for trusted offloading of federated learning tasks in an edge-cloud collaborative environment. The present invention optimizes the trusted bilateral collaborative offloading strategy to improve the quality of service for edge federated learning task offloading. In response to the bilateral intelligent DDoS attacks on the offloading link and edge cache, the present invention establishes an offloading link channel attack perception model and an edge cache attack perception model, and formalizes the quality of service assurance model for task offloading distribution delay and energy constraint under bilateral intelligent DDoS attacks. To model the evolutionary behavior of collaborative offloading when the task offloading links of multiple intelligent devices and the caches of multiple edge nodes are under bilateral intelligent DDoS attacks, the present invention designs an evolutionary game model to capture the dynamics of the task distribution strategy of intelligent devices and the model data caching strategy of edge nodes, and obtains an adaptive trusted offloading strategy under energy constraint and cache constraint by analyzing the evolutionary stable strategy. Under bilateral intelligent DDoS attacks, by deploying IDSs at bilateral nodes to share trusted state space information, estimating the trusted state and collaborative game utility, the present invention proposes a bilateral collaborative offloading optimization algorithm based on evolutionary game.

Description

Technical Field

[0001] The present invention belongs to the technical field of the Internet of Things, and more specifically relates to a trustworthy offloading system and method for federated learning tasks in an edge-cloud collaborative environment. Background Art

[0002] In edge federated computing, a task usually consists of a local model and CPU cycles for completing the task. To accelerate the execution of federated learning tasks and reduce the energy consumption of end devices, intelligent devices offload the federated learning model to edge nodes. Meanwhile, to meet the latency-sensitive requirements of task distribution, edge caches are used to cache the offloaded tasks on the edge node side and wait for the edge node to process them. The edge federated computing architecture using cache-assisted computing task offloading can efficiently handle latency-sensitive and computationally intensive tasks. Edge federated learning task offloading is divided into two stages. One is the distribution of model data in the federated learning task, and the other is that the edge cache stores the model data and uses CPU cycles for model training. In most cases, intelligent device federated learning task offloading only considers its distribution ability and rarely jointly considers the storage and CPU processing capabilities of edge caches. Further, existing research assumes that there are abundant resources for the distribution and caching of edge federated learning tasks to ensure the task offloading process, but in actual systems, the offloading process of federated learning tasks becomes unreliable due to limited computing power. In particular, when the task distribution link and cache are simultaneously under dual DDoS attacks, the quality of service of intelligent device federated learning task offloading is severely degraded. Due to the distributed nature of intelligent device federated learning task offloading, the computing task offloading process is subject to two main threats: offloading link attacks and edge cache attacks. An abnormal edge federated model aggregator or malicious intelligent device may launch DDoS attacks during the entire task offloading process. The former consumes link energy by sending interference signals and causes task distribution blockages, eventually resulting in link transmission interruptions. The latter injects invalid data into the edge cache, contaminating the edge cache or causing cache overflow. However, existing research ignores the collaborative attack characteristics of intelligent DDoS attackers, and few studies consider DDoS attackers launching DDoS attacks on both the intelligent device side and the edge node side simultaneously. The present invention considers collaborative strategies for task offloading by jointly analyzing both sides of the nodes under bilateral intelligent DDoS attacks.

[0003] Existing related research has proposed some optimization methods for the secure offloading of tasks. T. Dbouk et al. proposed a multi-objective resource optimization model for malware and other intrusions, providing intelligent decisions based on context and statistical data from edge cloud devices. This method does not consider the impact of bilateral DDoS collaborative attacks on intelligent offloading decisions ("A Novel Ad-Hoc Mobile Edge Cloud Offering Security Services Through Intelligent Resource-Aware Offloading," in IEEE Transactions on Network and Service Management, vol. 16, no. 4, pp. 1665-1680, Dec. 2019). Since multi-UAV systems have limited computing and communication resources, to enhance the system's operating capabilities, S. Islam et al. proposed an intelligent distributed task offloading mechanism based on federated deep reinforcement learning. However, due to backdoor attacks interfering with the normal operation of the system, the system performance rapidly degrades, and a lightweight defense mechanism was further proposed. Although this method considers backdoor attacks, it does not provide a corresponding offloading decision mechanism for federated system offloading links and cache attacks ("A Triggerless Backdoor Attack and Defense Mechanism for Intelligent Task Offloading in Multi-UAV Systems," in IEEE Internet of Things Journal, doi: 10.1109 / JIOT.2022.3172936). Due to the challenges posed by adversarial attacks and heterogeneous mobile environments introducing uncertainties in the supply and demand of network resources for task offloading decisions, B. Cho et al. addressed this challenge by selecting edge nodes based on online learning algorithms to reduce the complexity of task offloading decisions. However, this method does not design a security mechanism for bilateral collaborative scenarios ("Learning-Based Decentralized Offloading Decision Making in an Adversarial Environment," in IEEE Transactions on Vehicular Technology, vol. 70, no. 11, pp. 11308-11323, Nov. 2021).To minimize the transmission and computational resource consumption of offloading tasks, C. Feng et al. proposed a two-stage offloading decision mechanism, but did not consider the security mechanism of task offloading ("Two-Stage Task Offloading Optimization With Large Deviation Delay Analysis in IoT Networks," in IEEE Transactions on Communications, vol. 70, no. 3, pp. 1834-1847, March 2022). X. He et al. proposed a privacy protection and cost-efficient task offloading mechanism for attackers to launch inference attacks based on the characteristics of offloading tasks and steal users' private data. However, the privacy protection mechanism is prone to affecting the performance of task offloading decisions. This method does not start from the task offloading decision itself, embed secure offloading actions without changing the offloading performance, and optimize task offloading decisions ("Peace: Privacy-Preserving and Cost-Efficient Task Offloading for Mobile-Edge Computing," in IEEE Transactions on Wireless Communications, vol. 19, no. 3, pp. 1814-1824, March 2020).

[0004] The existing research solutions also have the following deficiencies:

[0005] 1. The proposed solutions only consider DDoS attacks on the smart device side or the edge node side. When facing bilateral intelligent DDoS attacks, the reliability of the federated learning task offloading process cannot be guaranteed. Therefore, the proposed solutions are defective in ensuring the federated learning task offloading process and do not propose corresponding protection methods for bilateral collaborative intelligent DDoS attacks.

[0006] 2. To minimize the task distribution and computational resource costs, the proposed solutions consider the two-stage decision-making mode of task offloading and the privacy protection of offloading data, but do not embed a secure offloading decision mechanism from the perspective of bilateral collaborative offloading and cannot improve the performance of the security guarantee mechanism.

[0007] 3. The proposed bilateral collaborative offloading solutions on the end device side and the edge node side do not consider deploying IDS (intrusion detection system) on bilateral nodes and optimizing the federated learning task offloading strategy by sharing trusted state information, which makes the existing collaborative offloading solutions unable to achieve efficient collaborative guarantee. Summary of the Invention

[0008] In view of the deficiencies of the existing technologies and the security reinforcement requirements of task offloading strategies, the present invention proposes a bilateral attack and defense system and method in an edge-cloud collaborative environment based on evolutionary game, optimizes a credible bilateral collaborative offloading strategy, and improves the quality of service for edge federated learning task offloading. To achieve reliable offloading of federated learning tasks under bilateral DDoS attacks, a method and system for trustworthy offloading of federated learning tasks in an edge-cloud collaborative environment are provided, including the following steps:

[0009] (1) Under bilateral intelligent DDoS attacks, intelligent devices use the offloading link attack perception module to observe their credible state space Γ s ={H0, H1}, and edge nodes use the cache attack perception module to observe their credible state space Γ e ={D0, D1} to collaboratively adjust their task offloading strategies, calculate the evolutionary game utility under different credible state spaces, so that the game participants who collaborate bilaterally reach an evolutionary strategy equilibrium of task offloading after a period of collaborative interaction. At the equilibrium state, the strategy action taken by the un-attacked intelligent device is κ, and the strategy action taken by the attacked intelligent device is The attacked edge nodes do not need to allocate caches either, and the un-attacked edge nodes allocate caches to store the federated learning task models;

[0010] (2) According to the optimal intelligent device offloading task distribution actions and edge node offloading task cache actions obtained in step (1).

[0011] (3) On the premise of meeting the constraint conditions, the optimal offloading task distribution strategy and cache strategy obtained in step (2) are used to minimize the offloading costs of all intelligent devices. The total cost of all devices on the intelligent device side is denoted as:

[0012]

[0013]

[0014]

[0015]

[0016]

[0017]

[0018] Among them, represents the set of intelligent devices, represents the set of edge federated nodes. Constraint (a) means that under bilateral intelligent DDoS attacks, the distribution time and the caching time of the offloading tasks do not exceed the maximum completion time and D m represents the data size of the federated learning task model. In the offloading link awareness module, the transmission rate at which the intelligent device m offloads tasks to the edge node during time slot t is:

[0019]

[0020]

[0021] where represents the channel bandwidth, and the intelligent device offloading task transmission decision represents that the intelligent device transmits the offloading task to the edge node; otherwise, it represents non - transmission. represents the transmission power used by the intelligent device to offload tasks to the edge node. represents the channel gain of the intelligent device offloading tasks to the edge node. σ 2 represents Gaussian noise. J mn represents the intensity of the interference attack on the link channel by the intelligent DDoS attacker during the process of the mobile intelligent device offloading tasks to the edge node. F0 represents that the intelligent device offloading link is not attacked, and F1 represents that the intelligent device offloading link is attacked. Thus, the average offloading rate within time T s is:

[0022]

[0023] The intelligent device pre - defines a threshold F th to determine whether its offloading link is under interference attack. The judgment conditions are as follows:

[0024] H0: F i ≥F th , then the intelligent device offloading link is not attacked. H1: F i <F th , then the intelligent device offloading link is attacked.

[0025] In the edge cache attack awareness module, the waiting time of the offloaded task in the cache is: where D w represents the task waiting to be processed in the cache. When the intelligent DDoS attacker launches a cache attack on the edge node, for the N s offloading tasks of the mobile intelligent device, the cache reception rate of the edge node is expressed as:

[0026]

[0027] where α∈(0,1) is a pre - defined parameter, β τ(t) represents the number of tasks to be processed in the edge node cache at time slot t. Δμ(t) represents the growth rate of the cached task number, represents the decreasing rate of the cached tasks. For the offloaded task τ, the edge caching decision is represented as a binary variable indicating that the mobile intelligent device offloads the task τ to the edge node and caches it, otherwise indicating that the edge node does not cache the offloaded task τ. The average task offloading and caching rate of the edge node within T s time is as follows: The edge node predefines a threshold D th to determine whether the cache is under a smart DDoS attack. When D0:D i < D th , then the edge node cache is not attacked. When D1:D i ≥ D th , then the edge node cache is attacked.

[0028] In constraint (b), represents that the transmission power constraint cannot exceed the maximum value (c) represents that the size of the cached data cannot exceed the cache capacity of the edge node Constraint (d) means that the intelligent device only selects one offloading and distribution strategy at a certain time slot. Constraint (e) means that the edge node only selects one caching strategy. By minimizing the total offloading task cost of all intelligent devices to ensure the quality of task offloading service, the total cost of all devices on the intelligent device side is denoted as:

[0029] Among them, and is the offloading task distribution time and energy consumption factor, The intelligent device selects different weights according to the DDoS attack situation. For example: The bilateral smart DDoS attacker attacks both the offloading task distribution link and the edge cache simultaneously, resulting in an increase in the latency cost. The intelligent device selects a large weight to reduce latency. When the energy of the intelligent device is almost exhausted, it selects a large weight to reduce energy consumption. represents the energy consumption of the device.

[0030] (4) The optimization module is the trusted cooperation offloading guarantee model G based on evolutionary game, denoted as:

[0031] G = (Σ, Γ, Λ, ρ, I)

[0032] Among them, Σ: represents the intelligent device and the edge node, which is the set of game participants. Γ: Γ = Γ s × Γ e , where Γ s={H0, H1} represents the trusted state space of the intelligent device. Γ e ={D0, D1} represents the trusted state space of the edge node. Λ: Λ = Λ z ×Λ x , represents the strategic action space of the intelligent device and the edge node. Among them represents the strategic action space of the intelligent device, κ represents the intelligent device offloading task, represents that the intelligent device does not offload the task. represents the strategic action space of the edge node, ε represents the task offloaded and cached, represents the task not offloaded and cached. When the offloading task of the intelligent device is not attacked or attacked, that is, in the trusted state spaces H0D0, H1D1, the intelligent device and the edge node cooperate to take four different strategic actions ρ: Γ → [0, 1], Λ → [0, 1] represents the probability distribution of the intelligent device and the edge node in the trusted state space and the strategic action space. Among them, ρ = (ρ κ , ρ ε ), where ρ κ represents the proportion distribution of intelligent devices adopting the offloading task strategy κ, ρ ε represents the proportion distribution of edge nodes adopting the caching strategy. I: I(Γ, Λ) represents the utility function of the game participants in the trusted state space and the strategic action space.

[0033] (5) For the trusted collaborative offloading guarantee model G based on evolutionary game, the intelligent device, as a game participant, repeatedly executes the game process and selects its own offloading strategy in the strategic action space Λ z to obtain the maximum utility of offloading task distribution. The repeated game process of the intelligent device using the offloading strategy is described by the dynamic replication equation. In the trusted state spaces H0D0, H1D1, the dynamic replication equation of the strategic action space adopted by the intelligent device is:

[0034]

[0035] Among them, η is the end-edge collaboration factor. In the trusted state spaces H0D0, H1D1, the average utility of the strategic action κ adopted by the intelligent device is I κ (H0D0, H1D1) = μ0I(κ, H0D0) + (1 - μ0)I(κ, H1D1). Among them, μ0 represents the probability of the trusted state space being H0D0, which is obtained from the observation information exchanged by the bilateral deployed IDS. 1 - μ0 represents the probability of the trusted state space being H1D1. In the trusted state space H0D0, the utility of the intelligent device adopting the strategic action κ is:

[0036]

[0037] Among them, T s represents the task offloading time of the intelligent device. Among them, C a represents the caching cost. represents the revenue obtained from offloading tasks. Here, the obtained cache is used as the revenue. The revenue obtained by the cooperation between the intelligent device and the edge node for offloading tasks is Among them, represents the probability of the edge node that selects the caching strategy action.

[0038] I(κ, H1D1) represents the utility of the intelligent device taking the strategy action κ in the trusted state space H1D1, and

[0039]

[0040] Among them, p(ε, H1D1) = ρ ε P ca , g ca represents the amount of cached data, represents the total amount of offloaded data. represents the number of intelligent devices that select to offload tasks.

[0041] represents the average utility of the strategy action taken by the intelligent device in the trusted state spaces H0D0 and H1D1, and Among them represents the utility of the intelligent device taking the strategy action in the trusted state space H0D0, and

[0042]

[0043] Among them, C a represents the caching cost.

[0044] represents the utility of the intelligent device taking the strategy action in the trusted state space H1D1, and

[0045]

[0046] Among them,

[0047] (6) For the trusted collaborative offloading guarantee model G based on evolutionary game, as a game participant on the edge side, the edge node selects its caching policy action in the policy action space Λ by repeatedly executing its game process to obtain the maximum utility of caching usage. The repeated game process of the edge node using the caching policy is described by the dynamic replication equation. Under the trusted state spaces H0D0 and H1D1, the dynamic replication equation of the policy action space adopted by the intelligent device is as follows: x Among them, under the trusted state spaces H0D0 and H1D1, I

[0048] where I ε (H0D0, H1D1) is the average utility of the policy action ε taken by the edge node, and I ε (H0D0, H1D1) = μ0I(ε, H0D0) + (1 - μ0)I(ε, H1D1). I(ε, H0D0) represents the utility when the edge node takes the policy action ε under the trusted state space H0D0, and

[0049]

[0050] where and where C a represents the caching cost. represents the revenue obtained from offloading tasks. Here, the obtained cache is used as the revenue. The revenue obtained by the intelligent device and the edge node through collaborative offloading tasks is where represents the probability of the edge node choosing the caching policy action.

[0051] I(ε, H1D1) represents the utility when the edge node takes the policy action ε under the trusted state space H1D1, and

[0052]

[0053] where v of represents the amount of data successfully distributed, represents the total amount of data distributed. represents the number of intelligent devices that choose to offload tasks.

[0054] represents the average utility of the policy action taken by the intelligent device under the trusted state spaces H0D0 and H1D1, and where represents the policy action taken by the edge node under the trusted state space H0D0 the utility, and

[0055]

[0056] wherein,

[0057] represents the utility of the edge node taking a strategic action in the trusted state space H1D1, and the utility, and

[0058]

[0059] wherein,

[0060] (7) For the trusted collaborative offloading guarantee model G based on evolutionary game, in the evolutionary game of bilateral collaborative strategy selection on the intelligent device side and the edge node side, under the bilateral intelligent DDoS attack, the game equilibrium (ESS) point is (ρ κ , ρ ε ) = (0, 0), which are the energy and cache thresholds when obtaining the maximum benefit. The energy threshold condition of the intelligent device is The cache threshold condition of the edge node is

[0061] The positive and progressive effect of the present invention is that the existing methods ignore the simultaneous attack characteristics of intelligent DDoS attackers on the intelligent device side and the edge node side, and cannot effectively detect and defend against attacks. The present invention considers the collaborative strategy of task offloading by jointly analyzing bilateral nodes based on evolutionary game under bilateral intelligent DDoS attacks, and uses the offloading link attack perception module and the edge cache attack perception module to effectively defend against DDoS attacks from the intelligent device side and the edge node side, improving the service quality of federated learning task offloading in the end-edge collaborative environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is the federated learning task offloading process under bilateral DDoS attacks provided by an embodiment of the present invention.

[0063] Figure 2 is the collaborative offloading strategy optimization process based on evolutionary game provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0064] The following provides a preferred embodiment of the present invention with reference to the accompanying drawings to detail the technical solution of the present invention.

[0065] Figure 1 is the federated learning task offloading process under bilateral DDoS attacks provided by an embodiment of the present invention. Figure 2This is the flow chart of the collaborative offloading strategy optimization based on evolutionary game provided by the present invention.

[0066] In view of the defects of the prior art and the security reinforcement requirements of the task offloading strategy, the present invention provides a bilateral DDoS attack defense system and method in the edge-cloud collaborative environment based on evolutionary game, optimizes the trusted bilateral collaborative offloading strategy, and improves the service quality of edge federated learning task offloading. The content of the present invention is as follows:

[0067] (1) For the bilateral intelligent DDoS attacks on the offloading link and the edge cache, an offloading link channel attack perception model and an edge cache attack perception model are established, and a service quality assurance model for task offloading distribution delay and energy constraint is formalized under the bilateral intelligent DDoS attack.

[0068] (2) To model the evolutionary behavior of collaborative offloading when the task offloading links of multiple intelligent devices and the caches of multiple edge nodes are under bilateral intelligent DDoS attacks, the present invention designs an evolutionary game model to capture the dynamics of the task distribution strategy of intelligent devices and the model data caching strategy of edge nodes, and obtains an adaptive trusted offloading strategy under energy constraint and cache constraint by analyzing the evolutionary stable strategy.

[0069] (3) Under the bilateral intelligent DDoS attack, by deploying IDSs at bilateral nodes to share the trusted state space information, estimating the trusted state and the collaborative game utility, a bilateral collaborative offloading optimization algorithm based on evolutionary game is proposed.

[0070] The following is the implementation process:

[0071] The edge federated learning offloading task L of the intelligent device in the present invention m ={D m ,C m}, where D m represents the size of the federated learning task model data, and C m represents the CPU processing cycle. The edge federated learning task offloading process is divided into two collaborative stages. The first stage is that the intelligent device distributes the task to the edge node through the wireless link, and is interfered and attacked by malicious nodes during the distribution process; the second stage is that the edge node receives the model data into the edge cache and waits for the CPU to process it, and is attacked by malicious nodes during the caching process. The collaborative attack on the federated learning task offloading process is as Figure 1 shown.

[0072] The edge-side collaborative environment described in the present invention is an edge federated learning network composed of M intelligent devices and N edge nodes. The node on the intelligent device side collaborates to sense whether the offloading link is attacked. Let F0 denote that the offloading link of the intelligent device is not attacked, and F1 denote that the offloading link of the intelligent device is attacked. The transmission rate at which the intelligent device m offloads tasks to the edge node at time slot t is:

[0073]

[0074]

[0075] where, represents the channel bandwidth, and the transmission decision of the intelligent device offloading task represents that the intelligent device transmits the offloading task to the edge node, otherwise it means not to transmit. represents the transmission power used by the intelligent device to offload the task to the edge node. represents the channel gain of the intelligent device offloading the task to the edge node. σ 2 represents Gaussian noise. J mn represents the intensity of the interference attack by the intelligent DDoS attacker on the link channel during the process of the mobile intelligent device offloading the task to the edge node.

[0076] Thus, the average offloading rate within time T s is obtained as:

[0077]

[0078] The intelligent device predefines a threshold F th to determine whether its offloading link is under interference attack, and the judgment conditions are as follows:

[0079] H0: F i ≥ F th , then the offloading link of the intelligent device is not attacked. H1: F i <F th , then the offloading link of the intelligent device is attacked.

[0080] Under the edge-side collaborative environment, multiple intelligent devices participating in federated learning offload tasks to the edge node. The edge node first caches the offloaded tasks into the cache queue. The DDoS attacker injects a large amount of invalid data into the cache queue, prolonging the waiting time for the normal task to be transferred to the CPU for processing. At this time, the intelligent device can select an adjacent intelligent device to store the task instead of offloading the task to the edge node. This mechanism can efficiently utilize the cache resources and prevent the edge node cache from overflowing. For the offloaded task τ, the edge cache decision is represented as a binary variable represents that the mobile intelligent device offloads the task τ to the edge node and caches it, otherwise Indicates that the edge node does not cache the offloading task τ. When the intelligent DDoS attacker launches a cache attack on the edge node, for the N s offloading tasks of the mobile intelligent device, the cache reception rate of the edge node is expressed as:

[0081]

[0082] where α ∈ (0, 1) is a predefined parameter, and β τ (t) represents the number of tasks to be processed in the edge node cache at time slot t. Δμ(t) represents the growth rate of the cached task number, represents the decrement rate of the cached task. The average task offloading cache rate of the edge node within T s time is:

[0083]

[0084] The edge node predefines a threshold D th to determine whether the cache is under an intelligent DDoS attack. When D0:D i < D th , the edge node cache is not attacked. When D1:D i ≥ D th , the edge node cache is attacked. 3. Federated Learning Task Offloading Assurance Model

[0085] Each intelligent device offloads tasks to the edge node at a transmission rate , and the offloading decision of the intelligent device is From this, the offloading decision constraint of the intelligent device is obtained as:

[0086]

[0087] This constraint ensures that tasks are not distributed or are distributed to one edge node. The time taken for task L m to be distributed to the edge node is:

[0088]

[0089] The corresponding energy consumption is:

[0090] When the task model offloaded by the intelligent device is relatively large, the edge node first caches this task model to wait for CPU processing. The waiting time of the offloaded task in the cache is:

[0091]

[0092] where D wRepresents the tasks waiting to be processed in the cache. Finally, the task offloading cost of the smart device under the DDoS attack is obtained according to the task offloading distribution delay and the cache waiting delay as follows:

[0093]

[0094] Wherein, And Is the offloading task distribution time and the energy consumption factor, The smart device selects different weights according to the DDoS attack situation. For example: The bilateral smart DDoS attacker attacks both the distribution link of the offloading task and the edge cache at the same time, resulting in an increase in the delay cost. The smart device selects a large weight To reduce the delay. When the energy of the smart device is almost exhausted, a large weight is selected To reduce the energy consumption.

[0095] The present invention proposes a method for optimizing the task offloading distribution delay and energy consumption under the bilateral smart DDoS attack. By minimizing the total offloading task cost of all smart devices, the quality of service of the task offloading is guaranteed. The total delay cost of the smart device node m is expressed as follows:

[0096]

[0097] Thus, the total cost of all devices on the smart device side is expressed as follows:

[0098]

[0099] Under the smart DDoS attack, the problem of guaranteeing the quality of service subject to the task offloading distribution delay and energy constraints is expressed as follows:

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106] This problem minimizes the cost of all smart devices by finding the offloading task distribution strategy and the caching strategy. Constraint (a) means that under the bilateral smart DDoS attack, the distribution time and the caching time of the offloading task do not exceed the maximum completion time (b) represents the transmission power constraint. (c) represents that the size of the cached data cannot exceed the cache capacity of the edge node (d) represents that in a certain time slot, the intelligent device only selects one offloading distribution strategy. (e) represents that the edge node only selects one caching strategy. Through the analysis of the constraint conditions, it is concluded that this problem contains five optimization variables. If these variables are solved simultaneously, it becomes complicated. Among them, the decision-indicator variable of the offloading task and the decision-indicator strategy for the edge node to cache the offloading task are binary variables. Therefore, this optimization problem is non-convex. In addition, since multiple variables are combined in the objective function and it is difficult to obtain a solution in polynomial time, this problem is an NP-hard problem.

[0107] Since this problem is caused by a bilateral intelligent DDoS attacker and involves the optimization problem of the offloading task distribution strategy and the edge node caching strategy phase. To model the evolutionary behavior of the cooperative offloading strategy of intelligent devices and edge nodes when multiple intelligent devices and edge nodes are under intelligent DDoS attacks, the present invention uses evolutionary game to capture the dynamics of the task offloading strategy of intelligent devices and the caching strategy of edge nodes. In addition, a dynamic replication equation is established to analyze the evolutionary stable strategy, so as to obtain an adaptive and credible offloading strategy under offloading delay and energy constraints. Thus, the balance of task offloading under bilateral intelligent DDoS attacks is used to ensure the credible service quality of edge federated learning tasks. The credible cooperative offloading guarantee model based on evolutionary game can be represented by a five-tuple: G=(Σ,Γ,Λ,ρ,I), where

[0108] Σ: represents intelligent devices and edge nodes, which is the set of game participants.

[0109] Γ:Γ = Γ s ×Γ e , where Γ s ={H0,H1} represents the credible state space of intelligent devices. Γ e ={D0,D1} represents the credible state space of edge nodes.

[0110] Λ:Λ = Λ z ×Λ x , represents the strategic action space of intelligent devices and edge nodes. Among them represents the strategic action space of intelligent devices, κ represents the offloading task of intelligent devices, represents that the intelligent device does not offload the task. represents the strategic action space of edge nodes, ε represents the task of caching offloading, represents the task of not caching offloading.

[0111] ρ:Γ→[0,1],Λ→[0,1] represents the probability distribution of intelligent devices and edge nodes in the credible state space and the strategic action space. Among them, ρ=(ρκ , ρ ε ), where ρ κ represents the proportional distribution of intelligent devices adopting the offloading task strategy κ, and ρ ε represents the proportional distribution of edge nodes adopting the caching strategy.

[0112] I: I(Γ, Λ) represents the utility function of game participants in the credible state space and the strategic action space.

[0113] Under the bilateral intelligent DDoS attack, the present invention considers an evolutionary game model of collaborative task offloading distribution and caching. Nodes on the intelligent device side select their offloading strategies in the strategic action space Λ z , and nodes on the edge side select their caching strategy actions in the strategic action space Λ x to reduce the time delay and energy consumption when under the bilateral intelligent DDoS attack and ensure the quality of service of task offloading for legitimate intelligent devices. Since intelligent devices repeatedly execute the offloading strategy within a certain time, that is, intelligent devices, as game participants, repeatedly execute the game process to obtain the maximum utility of offloading task distribution. Additionally, edge nodes, as game participants on the edge side, repeatedly execute their game processes to obtain the maximum utility of caching usage. On the intelligent device side, the evolutionary dynamic equation for each intelligent device's collaborative offloading is:

[0114]

[0115] where I(κ, H0) represents the utility of intelligent device collaborative offloading distribution under the credible state space H0. represents the average utility of intelligent device collaborative offloading distribution under the credible state space H0, and η is the edge - end collaboration factor. On the edge node side, the evolutionary dynamic equation for each edge node's collaborative offloading is:

[0116]

[0117] where I(ε, D0) represents the utility of edge node side collaborative caching under the credible state space D0. represents the average utility of edge node side collaborative caching under the credible state space D0. It can be seen from the evolutionary dynamic equation of each intelligent device's collaborative offloading that if the offloading task strategic action can lead to a value higher than the average utility, the probability of intelligent devices executing the offloading strategic action will increase. Similarly, it can be seen from the evolutionary dynamic equation of each edge node's collaborative offloading that if the task caching strategic action can lead to a value higher than the average utility, the probability of edge nodes executing the caching strategic action will increase.

[0118] When the offloading link of the intelligent device is not attacked, that is, in the trusted state space H0D0, the intelligent device and the edge node cooperate to take four different strategic actions The utility functions of these four offloading strategies are as follows:

[0119]

[0120]

[0121]

[0122]

[0123] Among them, C a represents the caching cost. represents the revenue obtained from offloading tasks. Here, the obtained cache is used as the revenue. The revenue obtained by the intelligent device and the edge node through cooperative offloading of tasks is:

[0124]

[0125] Among them, represents the probability of the edge node selecting the caching strategy action. Similarly, under the bilateral intelligent DDoS attack, the trusted state space is H1D1, and the utility calculation of the intelligent device is as follows:

[0126]

[0127]

[0128]

[0129]

[0130] Among them, represents the number of intelligent devices that choose to offload tasks.

[0131] In the trusted cooperation offloading guarantee model based on evolutionary game, the intelligent device, as a game participant, repeatedly executes the game process and selects its offloading strategy in the strategy action space Λ z to obtain the maximum utility of offloading task distribution. The repeated game process of the intelligent device using the offloading strategy is described by the dynamic replication equation. In the trusted state space H0D0, the utility of the intelligent device taking the strategy action κ is

[0132]

[0133] Among them,

[0134] In the trusted state space H1D1, the utility of the intelligent device taking the strategic action κ is:

[0135]

[0136] where p(ε, H1D1) = ρ ε P ca , g ca represents the amount of cached data, represents the total amount of offloaded data.

[0137] In the trusted state spaces H0D0 and H1D1, the average utility of the strategic action κ taken by the intelligent device is

[0138] I κ (H0D0, H1D1) = μ0I(κ, H0D0) + (1 - μ0)I(κ, H1D1)

[0139] where μ0 represents the probability of the trusted state space being H0D0, which is obtained from the observed information exchanged by the bilateral deployed IDS. 1 - μ0 represents the probability of the trusted state space being H1D1.

[0140] In the trusted state space H0D0, the utility of the intelligent device taking the strategic action is:

[0141]

[0142] In the trusted state space H1D1, the utility of the intelligent device taking the strategic action is:

[0143]

[0144] In the trusted state spaces H0D0 and H1D1, the average utility of the strategic action is:

[0145]

[0146] Finally, the dynamic replication equation of the strategic action space taken by the intelligent device in the trusted state spaces H0D0 and H1D1 is calculated as:

[0147]

[0148] In the trusted collaboration offloading guarantee model based on evolutionary game, the trusted state space during bilateral collaborative offloading and bilateral intelligent DDoS attacks is (H0D0, H1D1). In the trusted state space H0D0, the utility of the edge node side taking the strategic action ε is:

[0149]

[0150] Among them, Under the credible state space H1D1, the utility of the edge node taking the strategic action ε is:

[0151]

[0152] Among them, p(κ,H1D1) = ρ κ P of , v of represents the amount of data successfully distributed, represents the total amount of data distributed.

[0153] Under the credible state spaces H0D0 and H1D1, the average utility of the strategic action ε taken by the edge node is:

[0154] I ε (H0D0,H1D1) = μ0I(ε,H0D0)+(1 - μ0)I(ε,H1D1)

[0155] Calculate that under the credible state space H0D0, the utility of the edge node taking the strategic action is:

[0156]

[0157] Calculate that under the credible state space H1D1, the utility of the edge node taking the strategic action is:

[0158]

[0159] Under the credible state spaces H0D0 and H1D1, the average utility of the strategic action taken by the intelligent device is:

[0160]

[0161] Thus, the dynamic replication equation of the strategic action space taken by the edge node under the credible state spaces H0D0 and H1D1 can be obtained as:

[0162]

[0163] In the above-mentioned trusted collaborative offloading guarantee model based on evolutionary game, under the bilateral intelligent DDoS attack, the intelligent device side and the edge node side cooperate to adjust their task offloading strategies, and observe the game utility in different trusted state spaces. The evolutionary game describes the game process of the intelligent device side and the edge node side as game participants to collaboratively offload tasks, so that the game participants on both sides reach an evolutionary strategy equilibrium of task offloading after a period of collaborative interaction. At this equilibrium point, the utilities of the game participants are the same, that is, the obtained benefit after the strategy selection is zero, and no game participant deviates from the equilibrium point to obtain a higher benefit. In the equilibrium state, the strategy action of the un-attacked intelligent device is κ, and the strategy action of the attacked intelligent device is The attacked edge node does not need to allocate cache either, and the un-attacked edge node allocates cache to store the federated learning task model. Thus, it can be obtained that in the equilibrium state of the evolutionary game, in order to save the energy resources for task distribution, the number of intelligent device side nodes adopting the offloading strategy action and the non-offloading strategy action will be optimized in the equilibrium state. In addition, in order to save cache resources and evenly utilize the task offloading cache on the edge node side, the number of nodes on the edge node side adopting the cache strategy action is also optimized in the game equilibrium state. From the benefit function of edge federated learning task offloading and distribution under intelligent DDoS attack, we have:

[0164] It can be seen from this that the benefit function of edge federated learning task offloading and distribution is a decreasing function. Under the intelligent DDoS attack, the more intelligent devices choose to offload tasks, the less benefit is obtained.

[0165] From the edge cache benefit function under intelligent DDoS attack, we have:

[0166]

[0167] It can be seen from this that the edge cache benefit function is a decreasing function. Under the intelligent DDoS attack, the more edge nodes choose to cache the offloaded tasks, the less benefit is obtained.

[0168] Regarding the stability analysis of the above-mentioned trusted collaborative offloading guarantee model based on evolutionary game, in the evolutionary game of bilateral collaborative strategy selection for the intelligent device side and the edge node side, under the bilateral intelligent DDoS attack, the game equilibrium (ESS) point is (ρ κ , ρ ε ) = (0, 0), which is the energy and cache threshold when obtaining the maximum benefit. The evolutionary game utility equation systems for the strategy selection of the intelligent device in the trusted state spaces H0D0 and H1D1 are as follows:

[0169]

[0170] And set the dynamic replication equation e of the strategic action space adopted by the intelligent device under the trusted state spaces H0D0 and H1D1 κ = 0, and the energy threshold condition of the intelligent device under the intelligent DDoS attack can be obtained:

[0171] Similarly, the evolutionary game utility equation system for the strategic choice of the edge node under the trusted state spaces H0D0 and H1D1 is as follows:

[0172]

[0173] And set the dynamic replication equation e of the strategic action space adopted by the edge node under the trusted state spaces H0D0 and H1D1 ε = 0, and the cache threshold condition of the edge node under the intelligent DDoS attack can be obtained:

[0174]

[0175] Algorithm 1: Cooperative Offloading Method on the Intelligent Device Side Based on Evolutionary Game

[0176] 1. Given time slot t = 0, randomly initialize the trusted state space and the strategic action space of the game participants. 2. For the given time slot t:

[0177] Step 1: The node on the intelligent device side senses the interference attack state of the intelligent DDoS attacker;

[0178] Step 2: The node on the edge node side senses the cache attack state of the intelligent DDoS attacker;

[0179] Step 3: The node on the intelligent device side and the node on the edge node side exchange the trusted state space;

[0180] Step 4:

[0181] If F i ≥ F th , then the trusted state space on the intelligent device side is H0;

[0182] Otherwise, the trusted state space on the intelligent device side is H1;

[0183] If D i < D th , then the trusted state space on the edge node side is D0;

[0184] Otherwise, the trusted state space on the edge node side is D1;

[0185] The intelligent device takes the action of distributing offloading tasks;

[0186] Calculate the game utility of the combined trusted state spaces H0D0 and H1D1 on the side of the computing intelligent device and the edge node

[0187]

[0188] Calculate the energy threshold condition;

[0189] If Then, stop;

[0190] Otherwise, go to step 4.

[0191] Algorithm 2: Edge Node Side Cooperative Offloading Method Based on Evolutionary Game

[0192] 1. Given time slot t = 0, randomly initialize the trusted state space and the strategic action space of the game participants

[0193] 2. For the given time slot t:

[0194] Step 1: The node on the intelligent device side senses the interference attack state of the intelligent DDoS attacker;

[0195] Step 2: The node on the edge node side senses the cache attack state of the intelligent DDoS attacker;

[0196] Step 3: The node on the intelligent device side and the node on the edge node side exchange the trusted state space;

[0197] Step 4:

[0198] If F i ≥ F th , then the trusted state space on the intelligent device side is H0;

[0199] Otherwise, the trusted state space on the intelligent device side is H1;

[0200] If D i < D th , then the trusted state space on the edge node side is D0;

[0201] Otherwise, the trusted state space on the edge node side is D1;

[0202] The edge node takes the action of offloading task caching;

[0203] Calculate the game utility of the combined trusted state spaces H0D0 and H1D1 on the side of the computing intelligent device and the edge node

[0204]

[0205] Calculate the edge cache threshold condition;

[0206] If Then, stop;

[0207] Otherwise, go to step 4.

[0208] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and what is described in the above embodiments and the specification is only to illustrate the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will also have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A trustworthy offloading system for federated learning tasks in an edge-cloud collaborative environment, characterized in that, Including the following steps: (1) Under the bilateral intelligent DDoS attack, the intelligent device observes its trusted state space Γ s ={H0, H1}, and the edge node observes its trusted state space Γ e ={D0, D1} to collaboratively adjust its task offloading strategy, calculate the evolutionary game utility under different trusted state spaces, so that the game participants who cooperate bilaterally reach an evolutionary strategy equilibrium of task offloading after a period of cooperative interaction; at the equilibrium state, the unattacked intelligent device takes the strategic action κ, and the attacked intelligent device adopts the strategic action as The attacked edge node does not need to allocate cache either, and the unattacked edge node allocates cache to store the federated learning task model; (2) Obtain the optimal intelligent device offloading task distribution action and edge node offloading task caching action according to the step (1); On the premise of meeting the constraint conditions, minimize the offloading costs of all intelligent devices by using the optimal offloading task distribution strategy and caching strategy obtained in step (2). Minimizing the offloading costs of all intelligent devices is the optimization model. The total cost of all devices on the intelligent device side is denoted as: Among them, represents the set of intelligent devices, represents the set of edge federated nodes; Constraint (a) means that under the bilateral intelligent DDoS attack, the distribution time and caching time do not exceed the maximum completion time and D m represents the data size of the federated learning task model. In the offloading link attack perception module, the transmission rate of the task offloaded from the intelligent device m to the edge node at time slot t is: Among them, represents the channel bandwidth, and the intelligent device offloading task transmission decision represents that the intelligent device transmits the offloading task to the edge node, otherwise it represents not to transmit; represents the transmission power used by the intelligent device to offload the task to the edge node; represents the channel gain when the intelligent device offloads the task to the edge node; σ 2 represents Gaussian noise; J mn represents the interference attack intensity of the intelligent DDoS attacker on the link channel during the process of the mobile intelligent device offloading the task to the edge node; F0 represents that the intelligent device offloading link is not attacked, and F1 represents that the intelligent device offloading link is attacked; thus, the average offloading rate within time T is obtained as: s as follows: The intelligent device predefines a threshold F th to determine whether its offloading link is under interference attack, and the judgment conditions are as follows: H0: F i ≥ F th , then the offloading link of the intelligent device is not attacked; H1: F i < F th , then the offloading link of the intelligent device is attacked; In the edge node cache attack awareness module, the waiting time of the offloaded task in the cache is: where D w represents the task waiting to be processed in the cache. When the intelligent DDoS attacker launches a cache attack on the edge node, for the N s offloading tasks of the mobile intelligent device, the cache reception rate of the edge node is expressed as: Among them, α ∈ (0, 1) is a predefined parameter, and β τ (t) represents the number of tasks to be processed in the cache of the edge node at time slot t; Δμ(t) represents the growth rate of the number of cached tasks, represents the decreasing rate of the cached tasks; for the offloaded task τ, the edge caching decision is represented as a binary variable represents that the mobile intelligent device offloads the task τ to the edge node and caches it, otherwise represents that the edge node does not cache the offloaded task τ; within T s the average task offloading and caching rate of the edge node is: The edge node predefines a threshold D th to determine whether the cache is under a smart DDoS attack. When D0:D i < D th , then the cache of the edge node is not attacked. When D1:D i ≥ D th , then the cache of the edge node is attacked; Constraint (b) P mn Indicates that the transmission power constraint cannot exceed the maximum value (c) Indicates that the size of the cached data cannot exceed the cache capacity of the edge node Constraint (d) indicates that a smart device only selects one offloading and distribution strategy in a certain time slot; Constraint (e) indicates that the edge node only selects one caching strategy; By minimizing the total offloading task cost of all smart devices to ensure the quality of task offloading service, the total cost of all devices on the smart device side is denoted as: Among them, and is the offloading task distribution time and energy consumption factor, The intelligent device selects different weights according to the DDoS attack situation; for example: when the bilateral intelligent DDoS attacker attacks both the offloading task distribution link and the edge cache simultaneously, resulting in an increase in latency cost, the intelligent device selects a large weight To reduce latency, when the energy of the intelligent device is almost exhausted, a large weight is selected to reduce energy consumption; represents the energy consumption of the device.

2. The trustworthy offloading system for federated learning tasks in an edge-cloud collaborative environment according to claim 1, wherein The optimization module is to minimize the offloading costs of all intelligent devices. The optimization module is a trusted collaborative offloading guarantee model G based on evolutionary game, denoted as: G = (Σ, Γ, Λ, ρ, I) Among them, Σ represents the intelligent device and the edge node, which is the set of game participants; Γ: Γ = Γ s ×Γ e , where Γ s = {H0, H1} represents the trusted state space of the intelligent device; Γ e = {D0, D1} represents the trusted state space of the edge node; Λ: Λ = Λ z ×Λ x , represents the strategic action space of the intelligent device and the edge node; among them represents the strategic action space of the intelligent device, κ represents the intelligent device offloading task, represents that the intelligent device does not offload the task; represents the strategic action space of the edge node, ε represents the task of caching offloading, represents the task of not caching offloading; when the offloading task of the intelligent device is not attacked or attacked, that is, in the trusted state space H0D0 or H1D1, the intelligent device and the edge node cooperate to take four different strategic actions ρ: Γ → [0, 1], Λ → [0, 1] represents the probability distribution of the intelligent device and the edge node in the trusted state space and the strategic action space; among them, ρ = (ρ κ , ρ ε ), where ρ κ represents the proportion distribution of intelligent devices adopting the offloading task strategy κ, and ρ ε represents the proportion distribution of edge nodes adopting the caching strategy; I: I(Γ, Λ) represents the utility function of the game participants in the trusted state space and the strategic action space.

3. The federated learning task trusted offloading system in an edge-cloud collaborative environment according to claim 2, wherein For the trusted collaborative offloading guarantee model G based on evolutionary game, intelligent devices, as game participants, repeatedly execute the game process and select their offloading strategies in the strategy action space Λ z to obtain the maximum utility of offloading task distribution. The repeated game process of intelligent devices using offloading strategies is described by the dynamic replication equation. Under the trusted state spaces H0D0 and H1D1, the dynamic replication equation of the strategy action space adopted by intelligent devices is as follows: Among them, η is the edge collaboration factor. Under the trusted state spaces H0D0 and H1D1, the average utility of the strategic action κ taken by the intelligent device is I κ (H0D0, H1D1) = μ0I(κ, H0D0) + (1 - μ0)I(κ, H1D1); where μ0 represents the probability of the trusted state space being H0D0, which is obtained from the observation information exchanged by the two-sided deployed IDS; 1 - μ0 represents the probability of the trusted state space being H1D1; under the trusted state space H0D0, the utility of the intelligent device taking the strategic action κ is: Among them, T s represents the task offloading time of the intelligent device; Among them, C a represents the caching cost; represents the benefit obtained from offloading the task. Here, the obtained cache is used as the benefit. The benefit obtained by the intelligent device and the edge node through collaborative task offloading is Among them, represents the probability of the edge node that selects the caching policy action; I(κ, H1D1) represents the utility of the intelligent device taking the strategy action κ in the trusted state space H1D1, and where p(ε,H1D1) = ρ ε P ca , g ca represents the amount of cached data, represents the total amount of offloaded data; represents the number of intelligent devices that select offloading tasks; represents the average utility of the strategic actions taken by the intelligent devices in the trusted state spaces H0D0 and H1D1, and and where represents the utility of the strategic actions taken by the intelligent devices in the trusted state space H0D0, and and Among them, C a represents the cache cost; Denote the utility of the intelligent device taking strategic actions in the trusted state space H1D1 and Among them, 4. The federated learning task trusted offloading system in an edge-cloud collaborative environment according to claim 2, wherein For the trusted collaborative offloading guarantee model G based on evolutionary game, the edge node, as a game participant on the edge side, selects its caching strategy action in the strategy action space Λ by repeatedly executing its game process to obtain the maximum utility of cache usage; the repeated game process of the edge node using the caching strategy is described by the dynamic replication equation. Under the trusted state spaces H0D0 and H1D1, the dynamic replication equation of the strategy action space adopted by the intelligent device is as follows: x On it, to obtain the maximum utility of cache usage; by using the dynamic replication equation to describe the repeated game process of the edge node using the caching strategy, under the trusted state spaces H0D0 and H1D1, the dynamic replication equation of the strategy action space adopted by the intelligent device is: Among them, under the trusted state spaces H0D0 and H1D1, I ε (H0D0, H1D1) is the average utility of the strategic action ε taken by the edge node, and I ε (H0D0, H1D1) = μ0I(ε, H0D0) + (1 - μ0)I(ε, H1D1); I(ε, H0D0) represents the utility when the strategic action taken on the edge node side is ε under the trusted state space H0D0, and Among them, and where C a represents the caching cost; represents the benefit obtained from offloading tasks. Here, the obtained cache is used as the benefit, and the benefit obtained by the intelligent device and the edge node through collaborative offloading tasks is Among them, represents the probability of the edge node selecting the caching policy action; I(ε, H1D1) represents the utility of the edge node taking the strategy action ε in the trusted state space H1D1, and where p(κ,H1D1) = ρ κ P of , v of represents the amount of successfully distributed data, represents the total amount of distributed data; represents the number of intelligent devices that select offloading tasks; represents the average utility of the strategic actions taken by intelligent devices in the trusted state spaces H0D0 and H1D1, and where where represents the utility of the strategic action taken by the edge node in the trusted state space H0D0, and where Among them, Denote the utility of the edge node taking strategic actions under the trusted state space H1D1 and Among them,

Images