A federated image classification system and method based on historical state backtracking and probabilistic sampling grouping

Abstract

This invention discloses a federated image classification system and method based on historical state backtracking and probabilistic sampling grouping, comprising: an aggregation server communicating with multiple image acquisition terminals; the aggregation server maintaining a first-in-first-out model buffer for each terminal; each terminal uploading behavioral feature vectors; the server calculating the similarity between vectors, sampling dynamic thresholds from a Beta distribution probability, and constructing collaborative groups for each terminal based on the comparison results; the server using the current model parameters of a terminal as an anchor point, selecting the historical model version with the smallest Euclidean distance from the anchor point in the buffers of each member terminal within its collaborative group, and weighting and aggregating them to generate personalized model parameters which are then distributed to the terminals for classifying the acquired images. This invention achieves flexible collaborative grouping through probabilistic sampling, breaking through the bottleneck of hard clustering collaboration, and suppresses overfitting and model drift through historical state backtracking, thereby improving the personalized performance and convergence stability of the image classification model.

Description

Technical Field

[0001] This invention belongs to the field of federated learning and image classification technology, and in particular relates to a federated image classification system and method based on historical state backtracking and probability sampling grouping. Background Technology

[0002] With the development of computer vision technology, image classification algorithms have been widely used in fields such as intelligent security, medical image analysis, and autonomous driving. Existing image classification systems typically employ a centralized modeling approach, aggregating image data collected from various terminal devices or data sources to a central server for unified training to construct a global classification model. However, in practical applications, image data collected from different data sources often exhibits significant distributional differences and heterogeneity. For example, the distribution of image categories contained in different terminal devices or user terminals shows a significant skew; some data sources mainly contain "cat" images, while others mainly contain "dog" images, resulting in the data from different user terminals exhibiting non-independent and identically distributed characteristics in both the label and feature spaces.

[0003] In the aforementioned scenarios, a single global model struggles to adapt to diverse data distributions simultaneously, tending to favor the dominant data source. This leads to decreased classification accuracy on some user terminals and a reduction in the overall model's generalization ability. Furthermore, the significant differences in data distribution across user terminals can introduce model update conflicts during training, increasing communication overhead and computational costs, thus impacting model training efficiency.

[0004] To protect user privacy, existing technologies typically employ a federated learning framework, where the model is trained locally on the user's device and the parameters are uploaded to a server for aggregation, forming a unified global model. However, in highly personalized scenarios, existing federated learning methods still have shortcomings.

[0005] Because existing methods use a single global model for parameter aggregation, when the distribution of user terminal data varies greatly, the global model tends to favor user terminals with larger data volumes or dominant distributions, making it difficult to take into account the feature distributions of other user terminals. This results in a significant decrease in classification accuracy on some user terminals and limits the overall model's generalization ability.

[0006] In a non-IID data environment, the local gradient update directions of different user terminals vary significantly. Existing aggregation strategies are prone to mutual interference during model fusion, leading to large performance fluctuations of the global model across different user terminals, and even unstable convergence or local degradation. For example, in the context of car recognition, even if both cameras are capturing images of "cars," the background features extracted by cameras in northern winters (snowy, low-light conditions) and southern summers (bright, lush vegetation) are completely different. Forced aggregation can cause a significant drop in image classification accuracy under specific lighting or background conditions.

[0007] To mitigate data heterogeneity, some existing technologies employ similarity-based clustering federated learning methods, dividing user terminals into multiple clusters by setting a fixed threshold. However, these methods typically use a "hard partitioning" strategy, preventing user terminals with similarity slightly below the threshold from participating in collaborative training, thus limiting potential cross-domain knowledge sharing capabilities. Furthermore, unreasonable clustering can lead to redundant training and communication, increasing system computational and communication overhead. Image data is high-dimensional and contains multiple attributes. For example, client A takes a picture of a "cat during the day," client B takes a picture of a "cat at night," and client C takes a picture of a "dog during the day." Clients A and B share the outline features of the "cat," and client C shares the low-level convolutional features of "daytime lighting."

[0008] If hard clustering is used, once client A is forcibly assigned to the "cat" group, it completely loses the opportunity to communicate "daytime lighting" features with client C. This results in the image model being extremely inadequate in extracting low-level general visual features (such as edges, shadows, and textures), thus limiting the generalization ability of the convolutional kernel.

[0009] Existing methods often rely on a single similarity metric (such as cosine similarity) for model aggregation. As the parameters gradually converge in the later stages of training, it becomes difficult to distinguish fine-grained differences. This can easily introduce model parameters that do not match the feature space of the target user terminal, leading to performance degradation of the aggregated model in local image classification tasks. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention proposes a federated image classification system and method based on historical state backtracking and probability sampling grouping, thereby resolving the issues present in the prior art.

[0011] To achieve the above objectives, this invention provides a federated image classification method based on historical state backtracking and probability sampling grouping, applied to a distributed system comprising an aggregation server and multiple image acquisition terminals, including: The aggregation server receives the behavioral feature vectors uploaded by each image acquisition terminal and calculates the similarity value between the behavioral feature vectors of each image acquisition terminal. Based on the similarity value and dynamic threshold, a collaborative group is constructed for each target image acquisition terminal; For each image acquisition terminal, the aggregation server uses the image classification model parameters uploaded by the image acquisition terminal in the current round as the search anchor point, and selects the historical image classification model version based on geometric distance from the model buffers of each member image acquisition terminal in its collaboration group. The aggregation server performs weighted aggregation on the selected historical image classification model versions to generate personalized image classification model parameters, and sends the personalized image classification model parameters to the corresponding image acquisition terminal for the terminal to classify the acquired images.

[0012] Optionally, the process of generating the behavioral feature vector includes: The aggregation server distributes a common reference image sample set to each image acquisition terminal; Each image acquisition terminal uses its local image classification model to perform forward reasoning on the public reference image sample set to obtain a logical value vector; The mean of the logical value vectors calculated by each image acquisition terminal is uploaded to the aggregation server as the behavioral feature vector.

[0013] Optionally, the similarity value is cosine similarity.

[0014] Optionally, a collaborative group is constructed for each target image acquisition terminal based on the similarity value and a dynamic threshold, including: The aggregation server obtains the dynamic threshold by randomly sampling from the Beta distribution; For any two image acquisition terminals, if the similarity value between the two image acquisition terminals is greater than the dynamic threshold, then one of the image acquisition terminals will be included in the collaborative group of the other image acquisition terminal.

[0015] Optionally, the process of selecting a version of the historical image classification model includes: The aggregation server obtains the image classification model parameters uploaded by the image acquisition terminal in the current round as the search anchor point; The aggregation server traverses the model buffers of each member image acquisition terminal in the collaborative group of the image acquisition terminal, calculates the Euclidean distance between the search anchor point and the parameters of each historical image classification model stored in the model buffer, and selects the historical image classification model version with the smallest Euclidean distance.

[0016] Optionally, the model buffer is a first-in-first-out queue, and each model buffer has a preset storage depth. When the number of stored historical image classification model parameters exceeds the storage depth, the earliest stored historical image classification model parameters are removed.

[0017] Optionally, the aggregation server performs weighted aggregation on the selected historical image classification model versions, including: The aggregation server calculates the similarity value between the target image acquisition terminal and the image acquisition terminals of each member in its collaborative group; the aggregation server uses the similarity value as a weight to perform a weighted summation on each selected historical image classification model version to generate personalized image classification model parameters.

[0018] This invention also provides a federated image classification system based on historical state backtracking and probability sampling grouping, for implementing the above method, including: An aggregation server and multiple image acquisition terminals that are connected to the aggregation server via a network; The aggregation server includes a flexible collaborative grouping logic module and a historical version backtracking aggregation module; The flexible collaborative grouping logic module is used to construct collaborative groups for each image acquisition terminal based on the behavioral feature vectors uploaded by each image acquisition terminal. The historical version backtracking aggregation module is used to select the historical image classification model version with the smallest geometric distance from the current round image classification model parameters of each image acquisition terminal in the model buffer of each member image acquisition terminal in the collaboration group of each image acquisition terminal, and perform weighted aggregation on the selected historical image classification model versions to generate personalized image classification model parameters. The historical image classification model parameters are stored in each model buffer in a first-in-first-out manner.

[0019] Optionally, the flexible collaborative grouping logic module includes a behavior feature conversion unit and a probability discrimination logic unit; The behavior feature conversion unit is used to map the high-dimensional logical values ​​uploaded by each image acquisition terminal into a one-dimensional distribution representation vector. The probability discrimination logic unit is used to randomly sample from the Beta distribution to obtain a dynamic threshold, and to construct a cooperative group for each image acquisition terminal based on the comparison result of the cosine similarity between the one-dimensional distribution representation vectors of each image acquisition terminal and the dynamic threshold.

[0020] Optionally, the historical version backtracking aggregation module includes a multi-dimensional first-in-first-out parameter cache structure and an Euclidean distance minimization search unit; The multi-dimensional first-in-first-out parameter cache structure is used to save the historical image classification model parameters uploaded by each image acquisition terminal in continuous multi-round communication in a first-in-first-out manner. The Euclidean distance minimization search unit is used to select the historical image classification model version with the smallest Euclidean distance to the search anchor point from the multi-dimensional first-in-first-out parameter cache structure of each member image acquisition terminal in the collaborative group of the target image acquisition terminal, and to perform weighted aggregation on the selected historical image classification model versions to generate personalized image classification model parameters.

[0021] Compared with the prior art, the present invention has the following advantages and technical effects: This invention introduces Beta distribution probability sampling to replace fixed-threshold hard clustering, allowing terminals with overlapping visual features to participate in collaboration with dynamic probabilities, forming a flexible collaborative topology, breaking down collaboration silos and enhancing cross-domain knowledge transfer capabilities. It maintains a FIFO historical model buffer for each terminal, using the current model as an anchor point to select the best-matching historical version through Euclidean distance backtracking. The smoothness of the time dimension offsets immediate random biases, effectively suppressing local overfitting and model drift, and improving convergence stability. A dual-metric coupling structure of cosine similarity coarse selection and Euclidean distance fine filtering is constructed to accurately identify small deviations in weight magnitudes in the later stages of training, avoiding the introduction of irrelevant knowledge and suppressing negative transfer. While maintaining the same communication overhead as existing methods, it reduces repeated local training caused by model drift at the terminal, lowers computational load and communication rounds, and improves system convergence efficiency. Attached Figure Description

[0022] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is an overall framework diagram of an embodiment of the present invention; Figure 2 This is a schematic diagram of the hierarchical grouping strategy according to an embodiment of the present invention; Figure 3 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0025] Example 1 like Figure 3 As shown, this embodiment provides a federated image classification method based on historical state backtracking and probability sampling grouping, including: Step 1: Behavioral feature extraction based on logical values; The aggregation server distributes a public reference image sample set containing multiple benchmark categories. Each image acquisition terminal (user terminal) uses a locally trained image classification model (such as CNN or Vision Transformer) to perform forward inference on the sample set, extracting the mean vector of logits before the network output layer. The Logits vector represents the activation preference of the current terminal model for different visual categories, and it is uploaded to the aggregation server as a prior estimate of the local image data distribution of the terminal.

[0026] The purpose of distributing a public reference image sample set containing multiple benchmark categories is to estimate the data distribution for each user terminal. The public dataset is an estimate of the datasets from all user terminals; essentially, it's a merged dataset with a portion taken from each user terminal. Different models can estimate the model differences by performing inference on the same data.

[0027] This step uses the public dataset for inference on each user terminal. After inference, instead of calculating probabilities using the logic layer, the probabilities are directly used as logic values ​​to estimate the parameter offset of each model. This step obtains the average of multiple logic values ​​obtained from multiple forward inference iterations for each user terminal.

[0028] Step 2: Probabilistic user terminal selection based on Beta distribution; The aggregation server receives feature vectors uploaded by all participating user terminals. (Logical mean vector), and use cosine similarity to calculate the similarity between any two user terminals. With user terminal Cosine similarity between .

[0029] The variable The range of values ​​is The closer the value is to 1, the more similar the image category distributions or visual feature spaces identified by the two user terminals. The aggregation server starts from the Beta distribution. Random sampling dynamic threshold .like Then the terminal Join Terminal Collaboration group .

[0030] When the parameters in the Beta distribution When the Beta distribution is at this point, the graph generally shows a left skew (or positive skew, with a tail to the right), meaning the probability density is mainly concentrated near the 0 end. For example, when... At this point, the probability density of the Beta distribution is mainly concentrated around 0.2. This ensures that most similar models are grouped into the same group while also introducing noise appropriately. In conventional threshold grouping, user terminals often become increasingly sparse. Therefore, this embodiment uses the Beta distribution to introduce heterogeneous user terminals to increase robustness and reduce sparsity.

[0031] In image classification tasks, this not only clusters terminals with similar shooting scenes (such as those shooting indoor scenes), but also allows terminals with potential "complementary visual features" to participate in collaboration with a certain probability through dynamic thresholds, which greatly enhances the image classification model's ability to recognize long-tail categories and complex lighting environments.

[0032] Step 3: Historical checkpoint backtracking based on Euclidean distance; The aggregation server maintains a length of [length missing] for each user terminal. Historical model buffer. (Based on user terminal) The latest personalized image model weights uploaded in the current communication round are used as the search anchor. This anchor represents the latest learning state of the user terminal in extracting local image features. The aggregation server traverses the historical model buffer of each member in the collaboration group, which stores the member user terminal's most recent data in a FIFO (First-In-First-Out) manner. A snapshot of the wheel's parameters. Calculate and select the model version with the smallest Euclidean distance from the anchor point from the collaborating members' historical database.

[0033] With user terminal In the current communication round Uploaded latest personalized model weights This serves as the search anchor. The anchor represents the user's current, most up-to-date local learning status. The aggregation server iterates through the collaboration groups. Each member Historical model buffer This buffer stores the most recent data of the user terminal in a FIFO (First-In-First-Out) manner. Wheel parameter snapshot In the collaborative members Select the model version with the smallest Euclidean distance from the historical database. Lock in the historical state that best fits the anchor point weights to avoid potential overfitting bias in the latest round. For example... Figure 2 As shown, for anchor user terminals The collaborating group selects the user terminal marked by the red line, while for other anchor user terminals, the collaborating group selects other user terminals. The selected model version and user terminal may be different, which avoids the model shift that may be caused by a small number of abnormal local images (such as extremely blurry or occluded image samples) in the latest round, thus avoiding overfitting bias.

[0034] Unlike the cosine similarity used for macroscopic grouping in Step 2 (which focuses on the direction of parameter evolution), this step uses Euclidean distance for accuracy calibration at the microscopic level. Its physical significance lies in capturing the absolute geometric displacement of the weight parameters in hyperspace. In the later stages of image classification model training, the cosine similarity of different model parameters converges (direction alignment). At this point, Euclidean distance can more accurately identify pathological fluctuations caused by overfitting through subtle differences in the magnitude of convolution kernels or attention weights, thereby extracting truly generalizable image contour and texture features.

[0035] Step 4: Personalized adaptive aggregation; After completing the historical model version screening at the micro level (step 3), the aggregation server extracts the macro-group cosine similarity calculated in step 2. For a given anchor user terminal (target terminal device)... and any member user terminal in its collaboration group The cosine similarity This directly characterizes the similarity between the two in terms of local image data distribution features. To transform this similarity into a reasonable aggregation weight, the recommendation server assigns all members within the collaboration group to the anchor user terminal. The set of cosine similarities is input into the Softmax function for normalization. (Member user terminal) Corresponding aggregate weight The calculation formula is as follows: ; in, Indicates the anchor user terminal The system effectively creates a gradient in similarity scores by leveraging the exponential amplification effect and mapping normalization properties of the Softmax function. This ensures that collaborative models with visual feature distributions more similar to the target terminal device are assigned greater weights during the aggregation phase; conversely, weights are appropriately attenuated to achieve precise "soft collaboration." Finally, the aggregation server utilizes the calculated normalized weights... The classification model parameters for the high-quality historical images of each terminal selected in step 3 are... Perform a weighted summation to generate and distribute data to user terminals specific to the anchor point. Personalized update model : .

[0036] This embodiment also provides a federated image classification system based on historical state backtracking and probability sampling grouping to implement the above method, such as... Figure 1 As shown, it includes: An aggregation server and multiple image acquisition terminals that are connected to the aggregation server via a network; The aggregation server includes a flexible collaborative grouping logic module and a historical version backtracking aggregation module; The flexible collaborative grouping logic module is used to construct collaborative groups for each image acquisition terminal based on the behavioral feature vectors uploaded by each image acquisition terminal. The historical version backtracking aggregation module is used to select the historical image classification model version with the smallest geometric distance from the current round image classification model parameters of each image acquisition terminal in the model buffer of each member image acquisition terminal in the collaboration group of each image acquisition terminal, and perform weighted aggregation on the selected historical image classification model versions to generate personalized image classification model parameters. The historical image classification model parameters are stored in each model buffer in a first-in-first-out manner.

[0037] The flexible collaborative grouping logic module includes a behavior feature conversion unit and a probability discrimination logic unit; The behavior feature conversion unit is used to map the high-dimensional logical values ​​uploaded by each image acquisition terminal into a one-dimensional distribution representation vector. The probability discrimination logic unit is used to randomly sample from the Beta distribution to obtain a dynamic threshold, and to construct a cooperative group for each image acquisition terminal based on the comparison result of the cosine similarity between the one-dimensional distribution representation vectors of each image acquisition terminal and the dynamic threshold.

[0038] The historical version backtracking aggregation module includes a hierarchical first-in-first-out parameter cache structure and an Euclidean distance minimization search unit. The hierarchical first-in-first-out parameter cache structure is used to save the historical image classification model parameters uploaded by each image acquisition terminal in continuous multi-round communication in a first-in-first-out manner. The Euclidean distance minimization search unit is used to select the historical image classification model version with the smallest Euclidean distance to the search anchor point from the hierarchical first-in-first-out parameter cache structure of each member image acquisition terminal in the collaborative group of the target image acquisition terminal, and to perform weighted aggregation on the selected historical image classification model versions to generate personalized image classification model parameters.

[0039] Furthermore, the flexible collaborative grouping logic module based on Beta distribution sampling defines the dynamic construction structure of the cross-user terminal collaborative graph. It includes a behavioral feature transformation unit (for mapping high-dimensional Logits to one-dimensional distributed representation vectors) and a probabilistic discriminant logic unit. This discriminant logic unit does not use fixed numerical values ​​as decision boundaries; instead, it uses an instantaneous scalar generated by a random number generator following a Beta distribution as a dynamic threshold, comparing it in real-time with the cosine similarity between user terminals. This logically forms a non-deterministic, flexible collaborative topology that allows for cross-domain knowledge penetration.

[0040] Furthermore, the historical version backtracking aggregation module based on the weight space trajectory defines a data processing structure for vertical correction within the parameter space. It consists of a nested multi-dimensional first-in-first-out parameter cache structure and an Euclidean distance minimization search algorithm. This module uses the latest parameter tensor of the current target user terminal as the logical anchor point. By traversing the historical weight tensors at each level in the associated user terminal cache queue, it performs geometric distance comparisons based on weight magnitudes. The aim is to retrieve and pinpoint the historical parameter version that best matches the current anchor point state from the timeline evolution trajectory for precise aggregation.

[0041] Existing technologies only have parameter receiving registers for a single round in the recommendation server, which are then overwritten after aggregation. This invention establishes a FIFO (First-In-First-Out) queue buffer structure, which is logically bound to each user terminal ID, forming a time-series candidate pool with a depth of L.

[0042] When a collaborative user terminal overfits in the current round t, its latest weight parameters The geometric distance between the target anchor point and the target anchor point in the weight space increases non-linearly. This invention utilizes this structured buffer in conjunction with a Euclidean distance minimization search method. Its principle is to use the smoothness of historical states to counteract the random noise of the immediate state. The objective effect is the ability to automatically capture the model version with the best generalization performance from historical trajectories, significantly correcting the drift of locally trained models.

[0043] This invention establishes a FIFO (First-In, First-Out) model buffer on the server side, bound to the terminal ID, physically constructing a temporal-depth "visual feature memory" for each terminal. When a collaborative terminal overfits in the current round due to fitting random noise in the local image (such as overactivating the recognition of specific noise points or meaningless backgrounds), the absolute displacement of its deep convolutional kernels or attention weights in the parameter space increases non-linearly. At this time, the system uses Euclidean distance minimization search to automatically backtrack and capture the model version of that terminal in the historical trajectory that is not contaminated by the noise of the current batch and has the best generalization performance. This mechanism utilizes the smoothness of image feature learning over time, successfully offsetting the immediate damage caused by single local image data.

[0044] This invention employs a two-tiered pipeline structure of "coarse selection based on similarity + fine filtering based on Euclidean distance" to transform from a "single-dimensional quantity" to a "hierarchical screening structure".

[0045] In image classification tasks, while maintaining the same communication overhead as the FedAvg method, this invention effectively reduces the impact of local overfitting noise in each training round by improving the server-side model processing. This invention evaluates multiple historical versions during model evolution by calculating the Euclidean distance between a preset anchor model and historical models, selecting historical models with high matching degree and better stability for aggregation, thereby avoiding the direct use of potentially biased model parameters from the current round.

[0046] By utilizing historical version information during model evolution, the current model is screened and corrected, thereby reducing the adverse effects of random fluctuations and improving the stability of model training. Experimental results on the CIFAR-10 image classification dataset show that the method of this invention significantly improves classification accuracy compared to excellent methods such as FedAP, with an accuracy improvement of approximately 5%, thus demonstrating superior performance in image classification tasks, as shown in Table 1.

[0047] Table 1 The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A federated image classification method based on historical state backtracking and probability sampling grouping, applied to a distributed system containing an aggregation server and multiple image acquisition terminals, characterized in that, include: The aggregation server receives the behavioral feature vectors uploaded by each image acquisition terminal and calculates the similarity value between the behavioral feature vectors of each image acquisition terminal. Based on the similarity value and dynamic threshold, a collaborative group is constructed for each target image acquisition terminal; For each image acquisition terminal, the aggregation server uses the image classification model parameters uploaded by the image acquisition terminal in the current round as the search anchor point, and selects the historical image classification model version based on geometric distance from the model buffers of each member image acquisition terminal in its collaboration group. The aggregation server performs weighted aggregation on the selected historical image classification model versions to generate personalized image classification model parameters, and sends the personalized image classification model parameters to the corresponding image acquisition terminal for the terminal to classify the acquired images.

2. The federated image classification method based on historical state backtracking and probability sampling grouping according to claim 1, characterized in that, The process of generating the behavioral feature vector includes: The aggregation server distributes a common reference image sample set to each image acquisition terminal; Each image acquisition terminal uses its local image classification model to perform forward reasoning on the public reference image sample set to obtain a logical value vector; The mean of the logical value vectors calculated by each image acquisition terminal is uploaded to the aggregation server as the behavioral feature vector.

3. The federated image classification method based on historical state backtracking and probability sampling grouping according to claim 1, characterized in that, The similarity value is cosine similarity.

4. The federated image classification method based on historical state backtracking and probability sampling grouping according to claim 1, characterized in that, Based on the similarity value and dynamic threshold, a collaborative group is constructed for each target image acquisition terminal, including: The aggregation server obtains the dynamic threshold by randomly sampling from the Beta distribution; For any two image acquisition terminals, if the similarity value between the two image acquisition terminals is greater than the dynamic threshold, then one of the image acquisition terminals will be included in the collaborative group of the other image acquisition terminal.

5. The federated image classification method based on historical state backtracking and probability sampling grouping according to claim 1, characterized in that, The process of selecting a version of a historical image classification model includes: The aggregation server obtains the image classification model parameters uploaded by the image acquisition terminal in the current round as the search anchor point; The aggregation server traverses the model buffers of each member image acquisition terminal in the collaborative group of the image acquisition terminal, calculates the Euclidean distance between the search anchor point and the parameters of each historical image classification model stored in the model buffer, and selects the historical image classification model version with the smallest Euclidean distance.

6. The federated image classification method based on historical state backtracking and probability sampling grouping according to claim 1, characterized in that, The model buffer is a first-in-first-out queue. Each model buffer has a preset storage depth. When the number of stored historical image classification model parameters exceeds the storage depth, the earliest stored historical image classification model parameters are removed.

7. The federated image classification method based on historical state backtracking and probability sampling grouping according to claim 1, characterized in that, The aggregation server performs weighted aggregation on the selected historical image classification model versions, including: The aggregation server calculates the similarity value between the target image acquisition terminal and the image acquisition terminals of each member in its collaborative group; the aggregation server uses the similarity value as a weight to perform a weighted summation on each selected historical image classification model version to generate personalized image classification model parameters.

8. A federated image classification system based on historical state backtracking and probability sampling grouping, used to implement the method described in any one of claims 1-7, characterized in that, include: An aggregation server and multiple image acquisition terminals that are connected to the aggregation server via a network; The aggregation server includes a flexible collaborative grouping logic module and a historical version backtracking aggregation module; The flexible collaborative grouping logic module is used to construct collaborative groups for each image acquisition terminal based on the behavioral feature vectors uploaded by each image acquisition terminal. The historical version backtracking aggregation module is used to select the historical image classification model version with the smallest geometric distance from the current round image classification model parameters of each image acquisition terminal in the model buffer of each member image acquisition terminal in the collaboration group of each image acquisition terminal, and perform weighted aggregation on the selected historical image classification model versions to generate personalized image classification model parameters. The historical image classification model parameters are stored in each model buffer in a first-in-first-out manner.

9. The federated image classification system based on historical state backtracking and probability sampling grouping according to claim 8, characterized in that, The flexible collaborative grouping logic module includes a behavior feature conversion unit and a probability discrimination logic unit; The behavior feature conversion unit is used to map the high-dimensional logical values ​​uploaded by each image acquisition terminal into a one-dimensional distribution representation vector. The probability discrimination logic unit is used to randomly sample from the Beta distribution to obtain a dynamic threshold, and to construct a cooperative group for each image acquisition terminal based on the comparison result of the cosine similarity between the one-dimensional distribution representation vectors of each image acquisition terminal and the dynamic threshold.

10. The federated image classification system based on historical state backtracking and probability sampling grouping according to claim 8, characterized in that, The historical version backtracking aggregation module includes a multi-dimensional first-in-first-out parameter cache structure and an Euclidean distance minimization search unit; The multi-dimensional first-in-first-out parameter cache structure is used to save the historical image classification model parameters uploaded by each image acquisition terminal in continuous multi-round communication in a first-in-first-out manner. The Euclidean distance minimization search unit is used to select the historical image classification model version with the smallest Euclidean distance to the search anchor point from the multi-dimensional first-in-first-out parameter cache structure of each member image acquisition terminal in the collaborative group of the target image acquisition terminal, and to perform weighted aggregation on the selected historical image classification model versions to generate personalized image classification model parameters.

Images