A cloud edge-end based collaborative distributed learning defense method and system

Abstract

This invention relates to the field of computer system technology, specifically disclosing a collaborative distributed learning defense method and system based on cloud-edge-device architecture. During the learning process of a cloud server and n edge devices, after receiving the parameter differences from the n edge devices, the cloud server calculates the weight value of each edge device based on these parameter differences, and updates the parameters of the global model based on the weight values ​​of each edge device and the parameter differences. This invention hierarchically stratifies the parameter differences of each edge device according to the model structure, then clusters them to generate benign sets, enabling the detection of malicious edge devices at the layer space granularity, resulting in finer detection granularity. It utilizes cosine and Euclidean distance as two measurement scales to detect malicious edge devices at each layer, providing diverse detection scales. Different weights are assigned to different layers and different edge devices, making the weighting method more dynamic and flexible. The cloud server's runtime is linearly related only to the model parameters, resulting in low computational overhead.

Description

Technical Field

[0001] This invention relates to the field of computer system technology, and in particular to a collaborative distributed learning defense method and system based on cloud, edge, and terminal. Background Technology

[0002] In a collaborative distributed learning system between cloud, edge, and device, there are multiple edge devices and one cloud server. Each edge device has a local training dataset. In each communication, the cloud server distributes the global model from the cloud to the edge devices; then, each edge device uses its local training dataset to fine-tune the global model to train a model, called a local model, and submits the local model update to the cloud server; finally, the cloud server aggregates the local model updates from the edge devices and uses an aggregation rule to generate a new global model.

[0003] In cloud-edge-device collaborative distributed learning systems, some edge devices may take malicious actions to affect the training of the global model by modifying their local data or parts of the model. This type of attack is called poisoning attacks. Poisoning attacks are generally divided into two categories: data poisoning attacks and model poisoning attacks. Data poisoning attacks aim to contaminate training data, thereby corrupting the global model. Malicious edge devices can modify, add, or delete samples in their local training datasets, causing damage to the global model. The other type is model poisoning attacks, which can be further divided into untargeted attacks and targeted attacks, the latter also known as backdoor attacks, depending on the target. In untargeted attacks, malicious edge devices directly manipulate model updates, causing a significant decrease in the accuracy of the global model. In contrast, targeted attacks are more covert. Malicious edge devices modify local data and manipulate model updates to make the global model perform abnormally on specific target test samples (i.e., predict the target label set by the attacker), while the accuracy of other non-target samples remains unchanged.

[0004] The cloud-edge-device collaborative distributed learning defense method aims to prevent malicious model updates from being blocked by a central server when facing a limited number of malicious edge devices, by employing appropriate defense methods. Current defense methods can be summarized into five categories: distance-based, parameter-based, sign-flip-based, inversion-based, and principal component analysis-based.

[0005] Distance-based defense methods identify malicious devices by detecting distance differences between local model updates, where distances include L1 norm, L2 norm, cosine distance, etc. However, distance-based defense methods are vulnerable to custom attacks designed specifically for distance, leading to defense failure. Furthermore, different distance scales are suitable for different scenarios, but no single scale can fit all situations.

[0006] Parametric defense methods identify potential malicious devices by performing more detailed analysis of parameters updated in the local model. However, these methods suffer from several drawbacks. First, they incur high computational costs, especially when dealing with a large number of devices, where cloud server processing may become a bottleneck. Second, relying solely on analyzing locally updated parameters may not fully reflect device behavior, allowing malicious devices to evade detection.

[0007] Symbol-based defense methods distinguish between benign and malicious devices based on symbols updated in the local model. However, a limitation of symbol-based defense methods is that attackers can often exploit their characteristics to actively flip symbols to achieve their attack goals.

[0008] Inversion-based defense methods aim to invert potential triggers in each category and use adversarial training to eliminate backdoors in the global model. However, inversion-based defense methods incur significant computational overhead in inverting triggers, and adversarial training is costly. Furthermore, they lack flexibility, allowing attackers to adjust trigger positions in real time to launch attacks.

[0009] Principal component analysis (PCA)-based defense methods reduce the dimensionality of high-dimensional models using techniques such as eigenvalue decomposition (EVD) or singular value decomposition (SVD). The resulting parameters are then analyzed to identify and eliminate malicious devices. However, PCA-based defense methods are susceptible to decoy models. Attackers can manipulate decoy models to exhibit greater anomalies in the principal components, allowing the genuine malicious model to evade detection.

[0010] Existing defense methods detect malicious devices using various approaches such as distance-based, parameter-based, sign-flipping, inversion, and principal component analysis. However, they suffer from the following main problems: First, the detection granularity of these methods is too large to effectively identify fine-grained attackers. Second, the detection scale used is too singular to comprehensively identify differences between model updates. Third, the computational cost is high, and cloud servers and edge devices cannot support these computationally expensive methods. Summary of the Invention

[0011] This invention provides a collaborative distributed learning defense method and system based on cloud, edge, and terminal, which solves the technical problems of existing defense methods having excessively large detection granularity, single detection scale, and excessive computational cost.

[0012] To address the above technical problems, this invention provides a collaborative distributed learning defense method based on cloud-edge-device architecture, comprising the following steps:

[0013] S1. In each communication round, the cloud server broadcasts the parameters of the global model to n edge devices;

[0014] S2. Each edge device initializes the parameters of the received global model to its own parameters and updates its own parameters based on the local dataset.

[0015] S3. Each edge device sends the difference between its updated parameters and the parameters of the global model to the cloud server.

[0016] S4. The cloud server calculates the weight value of each edge device based on the parameter differences received from the n edge devices.

[0017] S5. Based on the weight value of each edge device and the difference in parameters uploaded by each edge device, the parameters of the global model are updated by aggregation.

[0018] Furthermore, step S4 specifically includes the following steps:

[0019] S41. The cloud server divides the parameter differences uploaded by each edge device into L layers according to the hierarchical structure of the network model.

[0020] S42. Calculate the paired cosine distance and paired Euclidean distance of the parameter differences for each layer of the n edge devices, respectively, to obtain the n parameter differences for each layer. 2 The pairwise cosine distances and n 2 A pair of European distances;

[0021] S43. For each layer of parameter difference n, respectively 2 The pairwise cosine distances and n 2 Clustering is performed using pairwise Euclidean distance to obtain corresponding cosine distance clusters and Euclidean distance clusters;

[0022] S44. Take the set of edge devices corresponding to the cluster with the most devices in the cosine distance cluster and the Euclidean distance cluster of each layer of parameter difference as the candidate benign set, and take the intersection of the two candidate benign sets as the benign set of parameter difference of that layer.

[0023] S45. Calculate the trust score of each edge device based on the benign set of parameter differences at each layer;

[0024] S46. Calculate the weight value of each edge device based on the trust score of each edge device.

[0025] Further, in step S45, the trust score of edge device i It is a binary classification definition, where β is 1 when the edge device i is in the benign set of the l-th layer, and 0 otherwise. β is a positive factor greater than 1, where i = 1, 2, ..., n.

[0026] Furthermore, β is an integer.

[0027] Further, in step S46, in the current round t, the weight value of edge device i is...

[0028] Furthermore, in step S5, the parameters w of the global model t Updated to w t-1 This represents the parameters of the global model before the update.

[0029] Further, in step S42, the paired cosine distance between edge device i and edge device j in the t-th round and the l-th layer... and Let represent the parameter difference of the l-th layer calculated by edge devices i and j in the t-th round, where <·,·> represents the inner product, and ‖·‖2 represents the L2 norm of the vector, and j = 1, 2, ..., n.

[0030] Further, in step S42, the paired Euclidean distance between edge device i and edge device j in the t-th round and the l-th layer...

[0031] Furthermore, in step S43, the K-means clustering method is used to analyze the parameter differences n at each layer. 2 The pairwise cosine distances and n 2 Clustering is performed using pairwise Euclidean distances.

[0032] The present invention also provides a cloud-edge-device collaborative distributed learning defense system, the key of which is: it includes a cloud server and n edge devices, and the cloud server and n edge devices learn according to the cloud-edge-device collaborative distributed learning defense method.

[0033] This invention provides a collaborative distributed learning defense method and system based on cloud-edge-device architecture. During the learning process between a cloud server and n edge devices, after receiving the parameter differences from the n edge devices, the cloud server calculates the weight value of each edge device based on these differences. Then, it updates the parameters of the global model based on the weight values ​​of each edge device and the parameter differences. This invention hierarchically categorizes the parameter differences uploaded by each edge device according to the network model's layer structure, then clusters them to generate a benign set, enabling the detection of malicious edge devices at the layer-space granularity, resulting in finer detection granularity. It utilizes cosine and Euclidean distance as two measurement scales to detect malicious edge devices at each layer, providing diverse detection scales. Different weights are assigned to different layers and different edge devices, making the weighting method more dynamic and flexible. The cloud server's runtime is linearly related only to the model parameters, resulting in low computational overhead. Attached Figure Description

[0034] Figure 1 This is a diagram of an existing cloud-edge-device-based collaborative distributed learning system architecture provided in this embodiment of the invention.

[0035] Figure 2 This is a flowchart of a collaborative distributed learning defense method and system based on cloud, edge, and terminal provided in an embodiment of the present invention;

[0036] Figure 3 This is a detailed process example diagram of steps S41 to S44 provided in the embodiments of the present invention;

[0037] Figure 4 These are specific process examples of steps S4 to S5 provided in the embodiments of the present invention. Detailed Implementation

[0038] The embodiments of the present invention are described in detail below with reference to the accompanying drawings. The embodiments are given for illustrative purposes only and should not be construed as limiting the present invention. The accompanying drawings are for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0039] Existing cloud-edge-device collaborative distributed learning system architectures, such as Figure 1 As shown, the system includes a cloud server and n edge devices. Each edge device deploys a local model, and the cloud server deploys a global model. The local and global models use the same model architecture. The learning process between the cloud server and the n edge devices mainly consists of four steps:

[0040] ① In each communication round t, the cloud server transmits the parameters w of the global model via the network. t-1 Broadcast to n edge devices.

[0041] ② Each edge device i (i = 1, 2, ..., n) first uses the parameters w received from the global model. t-1 Parameters of its own local model Initialization is performed, with Then, E rounds of iterative training are performed. In each round, their local training dataset is divided into several batches according to the batch size N. Let B be the set of these batches. For each batch b∈B, stochastic gradient descent is performed to update the parameters of the local model.

[0042]

[0043] Where t represents the current communication round, and η is the learning rate used for training. This represents the average gradient of edge device i on batch b in the current iteration.

[0044] ③ After the edge device completes local training, it updates the local model via the network. Send to the server, denoted as

[0045] ④ The cloud server obtains a new global model through a specific aggregation algorithm, such as average aggregation.

[0046]

[0047] exist Figure 1 Based on this, the present invention designs a collaborative distributed learning defense method based on cloud, edge, and endpoint, as exemplified by... Figure 2 As shown, the steps include:

[0048] S1. In each communication round t, the cloud server will transfer the parameters w of the global model. t-1 Broadcast to n edge devices;

[0049] S2. Each edge device i (i = 1, 2, ..., n) will receive the parameters w of the global model. t-1 Initialize to its own parameters And based on the local dataset, adjust its own parameters. Update;

[0050] S3, Each edge device i will update its own parameters With the parameters w of the global model t-1 The difference between Send to the cloud server;

[0051] S4, The cloud server is based on the parameter differences received from n edge devices. to Calculate the weight value for each edge device;

[0052] S5. Based on the weight value of each edge device and the difference in parameters uploaded by each edge device, the parameters of the global model are updated by aggregation.

[0053] Compared with steps ① to ④, the improvement of the cloud-edge-device-based collaborative distributed learning defense method designed in this invention lies in steps S4 and S5.

[0054] Specifically, step S4 includes the following steps:

[0055] S41. The cloud server, according to the hierarchical structure of the network model, processes the parameter differences uploaded by each edge device. Divided into L (≥2) layers;

[0056] S42. Calculate the paired cosine distance and paired Euclidean distance of the parameter differences for each layer of the n edge devices, respectively, to obtain the n parameter differences for each layer. 2 The pairwise cosine distances and n 2 A pair of European distances;

[0057] S43. For each layer of parameter difference n, respectively 2 The pairwise cosine distances and n 2 Clustering is performed using pairwise Euclidean distance to obtain corresponding cosine distance clusters and Euclidean distance clusters;

[0058] S44. Take the set of edge devices corresponding to the cluster with the most devices in the cosine distance cluster and the Euclidean distance cluster of each layer of parameter difference as the candidate benign set, and take the intersection of the two candidate benign sets as the benign set of parameter difference of that layer.

[0059] S45. Calculate the trust score of each edge device based on the benign set of parameter differences at each layer;

[0060] S46. Calculate the weight value of each edge device based on the trust score of each edge device.

[0061] In step S41, for non-targeted attacks, the attacker's goal is to significantly reduce the accuracy of the global model. The essence of a non-targeted attack is to create a significant deviation in the aggregated model before and after the attack through a carefully designed model. Compared to a benign model, these carefully designed models have a larger angular deviation, indicating that they are likely to threaten the collaborative distributed learning system between the cloud and edge. Therefore, to detect and mitigate these deviations, the parameter differences uploaded by the edge devices are obtained. to Then, the cloud server is layered according to the model's structure.

[0062] For ease of understanding, this embodiment illustrates the specific process of steps S41 to S44 with an example of 3 edge devices and 3 layers for each edge device.

[0063] exist Figure 3 In the example, assuming edge device 3 is attacked (marked in red), specifically its local model's third layer is attacked (marked in red), then the uploaded third-layer parameter discrepancies are problematic, but these cloud servers are unaware of this. After applying a cloud-edge collaborative distributed learning defense method, the third-layer parameter discrepancies should be completely eliminated.

[0064] In step S42, the parameter difference at each layer of edge device i is... ( to In this embodiment, paired cosine distance is used to identify malicious edge devices that launch non-directional attacks, and paired Euclidean distance is used to identify malicious edge devices that launch directional attacks.

[0065] The cloud server uses the following formula to calculate the pairwise cosine distance between all local model updates at each layer:

[0066]

[0067] in, and Let represent the parameter difference between edge devices i and j calculated for the l-th layer in the t-th round, where j = 1, 2, ..., n. <·,·> represent the inner product, and ‖·‖2 represents the L2 norm of the vector, also known as the Euclidean distance. Let represent the cosine distance between edge device i and edge device j at layer l in round t. This scale, by measuring the angular difference in parameter differences from different edge devices at each layer, helps identify anomalous updates that may trigger attacks. Subsequently, the cloud server iteratively calculates the parameter differences of n edge devices at each layer using formula (3). to The pairwise cosine distance between them is denoted as exist Figure 3 In the example, the paired cosine distances of the three edge devices in the first layer The paired cosine distance of the three edge devices on the second layer The paired cosine distance of the three edge devices in the third layer Each layer has 9 cosine distance values.

[0068] For backdoor model poisoning attacks, attackers typically employ two types of attack strategies. The first strategy involves training a backdoor model with high attack impact, which differs significantly from a benign model. In this case, attackers usually use data with a high Poison-Data-Rate (PDR) or consume a large number of local training cycles.i Represents the clean dataset of edge device i. If the dataset represents the contaminated dataset, then the merged poisoned dataset D′ is... i The PDR is given by the following formula:

[0069]

[0070] Where || represents the size of the dataset.

[0071] Another type of attack strategy involves training a weak backdoor model on low PDR data, but these models can increase their contribution to the aggregate model by expanding the parameters of the local model.

[0072] For backdoor attacks, regardless of the strategy employed, the Euclidean distance deviates significantly from the Euclidean distance updated by the benign model. Therefore, to mitigate the impact of backdoor attacks, this embodiment uses the following formula to calculate the pairwise Euclidean distance of the parameter differences at each layer for n edge devices:

[0073]

[0074] in, Let represent the Euclidean distance between edge device i and edge device j in the t-th round and the l-th layer. Pairwise Euclidean distance can capture richer features and subtle differences, thus enabling more detailed and nuanced anomaly detection. Similar to the method of calculating pairwise cosine distance, the cloud server iteratively uses formula (5) to calculate the pairwise Euclidean distance between the parameter differences of n edge devices in each layer, denoted as . exist Figure 3 In the example, the pairwise Euclidean distance of the three edge devices in the first layer The paired European distance of the three edge devices on the second layer The paired Euclidean distance of the three edge devices on the third layer Each layer has 9 Euclidean distance values.

[0075] In step S43, after the server calculates the pairwise cosine distance and pairwise Euclidean distance for each layer, the server uses the K-means clustering algorithm to cluster the pairwise cosine distance and pairwise Euclidean distance for each layer to detect malicious layers. Since most edge devices in the cloud-edge-device collaborative distributed learning system are benign, the clusters with more devices in the clustering results are considered as the set of benign devices, i.e., the candidate benign set. In other words, each layer has two pairs of nodes whose pairwise cosine distances are C0 and C1. t(l) And paired Euclidean distance E t(l) The generated candidate benign set. Figure 3 In the example, the first layer consists of paired cosine distances C t(1) And paired Euclidean distance E t(1)The generated candidate benign sets are represented as [2,3] and [1,3], respectively, where 1, 2, and 3 refer to edge devices 1, 2, and 3, respectively. The second layer consists of paired cosine distances C. t(2) And paired Euclidean distance E t(2) The generated candidate benign sets are represented as [1,3] and [1,3], respectively. The third layer consists of pairwise cosine distances C. t(3) And paired Euclidean distance E t(3) The generated candidate benign sets are represented as [1,2,3] and [1,2], respectively.

[0076] To comprehensively evaluate edge devices at each layer, the intersection of two candidate benign sets for each layer is taken as the benign set for that layer. Figure 3 In the example, the benign set of the first layer is [3], the benign set of the second layer is [1,3], and the benign set of the third layer is [1,2]. After the above steps, it can be seen that the benign set of the third layer [1,2] does not include edge device 3, thus successfully removing the third layer parameter difference of edge device 3.

[0077] Step S45 uses a layer-based depth and benign set dynamic scoring strategy to calculate the trust score of each edge device. Step S46 uses a dynamic aggregation method based on the trust score to calculate the weight value of each edge device. Figure 4 Based on Figure 3 The flowcharts for steps S45 and S46 are shown below. Generally, deeper layers in a model are more important than shallower layers because in neural network models, such as DNNs (Deep Neural Networks), shallow layers typically learn low-level features, such as edges and textures, while deeper layers combine these features to learn more complex high-level concepts, such as objects and their parts. Therefore, deeper layers are generally more important to model performance than shallower layers. At the same time, although the output layer may be the most important layer, the filtering results of other layers cannot be ignored, as attackers can successfully compromise the model by attacking other layers. Therefore, this embodiment considers the parameters of all layers and dynamically assigns scores based on layer depth; deeper layers are selected as benign and receive higher scores. The scoring rules for edge device i are as follows:

[0078]

[0079] Among them, Score i This represents the trust score of edge device i. This is a binary classification definition: β is 1 if edge device i belongs to the benign set at layer l, and 0 otherwise. β is a positive factor greater than 1; when β is greater than 1, deeper layers are more important. For ease of calculation, β is rounded to an integer. Finally, malicious edge devices are filtered out using the trust score of each edge device.

[0080] exist Figure 4 In the example, the value of β is set to 2. For edge device 1 (child node 1, corresponding to 1 in the benign set), the trust score is... Using the same calculation formula, the trust score Score2 of edge device 2 (child node 2, corresponding to 2 in the benign set) is β. 3 =8, the trust score Score3 of edge device 3 (child node 3, corresponding to 3 in the benign set) = β 1 +β 2 =2+4=6.

[0081] In step S46, after obtaining the trust score of each edge device, this embodiment adopts a dynamic model aggregation strategy based on dynamic scores. First, the weight value of each edge device is calculated. This embodiment uses the following formula to calculate the weight value of terminal device i.

[0082]

[0083] Finally, in step S5, the traditional federated averaging algorithm averages all edge devices, which ignores the contribution of each edge device to the model. This project uses the trust score of each edge device as a weighted value; the higher the score, the higher the weight, to reflect the importance of edge devices that contribute more. The global model's parameters are updated as follows:

[0084]

[0085] in, For edge device i, update the local model parameters in round t, w t These are the parameters after the t-th round of aggregation for the cloud server.

[0086] Based on the above method, this embodiment of the invention also provides a cloud-edge-device collaborative distributed learning defense system, which includes a cloud server and n edge devices, wherein the cloud server and n edge devices learn according to the above-described cloud-edge-device collaborative distributed learning defense method.

[0087] In summary, the cloud-edge-device collaborative distributed learning defense method and system provided by this invention involves the following steps during the learning process of a cloud server and n edge devices. After receiving the parameter differences from the n edge devices, the cloud server calculates the weight value of each edge device based on these differences, and then updates the parameters of the global model based on the weight values ​​of each edge device and the parameter differences. This invention hierarchically categorizes the parameter differences uploaded by each edge device according to the network model's hierarchical structure, then clusters them to generate a benign set, enabling the detection of malicious edge devices at the layer-space granularity, resulting in finer detection granularity. It utilizes cosine and Euclidean distance as two measurement scales to detect malicious edge devices at each layer, providing diverse detection scales. Different weights are assigned to different layers and different edge devices, making the weighting method more dynamic and flexible. The cloud server's runtime is linearly related only to the model parameters, resulting in low computational overhead.

[0088] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A collaborative distributed learning defense method based on cloud-edge-device architecture, characterized in that, Including the following steps: S1. In each communication round, the cloud server broadcasts the parameters of the global model to n edge devices; S2. Each edge device initializes the parameters of the received global model to its own parameters and updates its own parameters based on the local dataset. S3. Each edge device sends the difference between its updated parameters and the parameters of the global model to the cloud server. S4. The cloud server calculates the weight value of each edge device based on the parameter differences received from the n edge devices. S5. Based on the weight value of each edge device and the difference in parameters uploaded by each edge device, the parameters of the global model are updated by aggregation. Step S4 specifically includes the following steps: S41. The cloud server divides the parameter differences uploaded by each edge device into L layers according to the hierarchical structure of the network model. S42. Calculate the paired cosine distance and paired Euclidean distance of the parameter differences for each layer of the n edge devices, respectively, to obtain the n parameter differences for each layer. 2 The pairwise cosine distances and n 2 A pair of European distances; S43. For each layer of parameter difference n, respectively 2 The pairwise cosine distances and n 2 Clustering is performed using pairwise Euclidean distance to obtain corresponding cosine distance clusters and Euclidean distance clusters; S44. Take the set of edge devices corresponding to the cluster with the most devices in the cosine distance cluster and the Euclidean distance cluster of each layer of parameter difference as the candidate benign set, and take the intersection of the two candidate benign sets as the benign set of parameter difference of that layer. S45. Calculate the trust score of each edge device based on the benign set of parameter differences at each layer; S46. Calculate the weight value of each edge device based on the trust score of each edge device.

2. The cloud-edge-device-based collaborative distributed learning defense method according to claim 1, characterized in that: In step S45, the trust score of edge device i , It is a binary classification definition, when the edge device In the The value is 1 when the layer is a benign set, and 0 otherwise. It is a positive factor greater than 1, i=1,2,…,n.

3. The cloud-edge-device-based collaborative distributed learning defense method according to claim 2, characterized in that: It is an integer.

4. The cloud-edge-device-based collaborative distributed learning defense method according to claim 3, characterized in that: In step S46, the weight value of edge device i in the current round t is... .

5. The cloud-edge-device-based collaborative distributed learning defense method according to claim 4, characterized in that: In step S5, the parameters of the global model Updated to , This represents the parameters of the global model before the update. Indicates the first Wheel end equipment The calculated parameter difference.

6. A cloud-edge-device-based collaborative distributed learning defense method according to any one of claims 1 to 5, characterized in that: In step S42, the edge device and edge devices In the Round Paired cosine distance of layers , and Indicates the first Wheel end equipment and Calculated Layer parameter differences, Represents the inner product. Let L2 be the L2 norm of the vector, j = 1, 2, ..., n.

7. The cloud-edge-device-based collaborative distributed learning defense method according to claim 6, characterized in that: In step S42, the edge device and edge devices In the Round Paired Euclidean distance of layers .

8. The cloud-edge-device-based collaborative distributed learning defense method according to claim 7, characterized in that: In step S43, the K-means clustering method is used to analyze the parameter differences n of each layer. 2 The pairwise cosine distances and n 2 Clustering is performed using pairwise Euclidean distances.

9. A cloud-edge-device-based collaborative distributed learning defense system, characterized in that: It includes a cloud server and n edge devices, and the cloud server and n edge devices learn according to the cloud-edge collaborative distributed learning defense method according to any one of claims 1 to 8.

Images