Long tail federated learning method and device based on positive incentive noise aggregation

Abstract

The application provides a long-tail federated learning method and device based on positive incentive noise aggregation, which are applied to federated model training in a heterogeneous long-tail data distribution scene. The method eliminates the cross error of traditional federated averaging through correct low-rank aggregation of double low-rank factors, constructs a heterogeneity perception graph based on the similarity of client task characteristics, constructs structured noise with far neighbor enhancement and near neighbor suppression on the basis of the heterogeneity perception graph, modulates the aggregation result, and combines singular value decomposition for rank constraint to maintain the low-rank structure of the LoRA parameter, thereby effectively improving the recognition accuracy of rare classes at the tail in the federated long-tail scene.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and distributed machine learning technology, and particularly to a long-tail federated learning method based on positive incentive noise aggregation, a long-tail federated learning device based on positive incentive noise aggregation, a computer-readable storage medium, and a computer device. Background Technology

[0002] Parameter-Efficient Fine-Tuning (PEFT) is a practical approach for Federated Learning (FL), with Low-Rank Adaptive Fine-Tuning (LoRA) becoming a mainstream choice due to its high communication efficiency and compatibility with large models. Early FL+LoRA work focused primarily on low-rank update transmission or decoupling sharing from client-specific adapters to balance personalization and efficiency. Recent methods have further introduced heterogeneous or structured aggregation mechanisms, allowing clients to use different ranks or selectively participate through rank alignment, weighting, and parameter selection. Despite these advancements, existing works generally treat heterogeneity as an unstable factor requiring correction, often suppressing aggregation artifacts without exploring its potential constructive role in long-tail representation learning.

[0003] Classic approaches to addressing class imbalance in flaccid systems fall into two main categories: one focuses on limiting local drift through optimization constraints, while the other performs server-side aggregation correction to mitigate non-IID bias. These methods improve training stability but often fail to explicitly consider the global long-tail distribution, resulting in poor performance for tail classes. Furthermore, strategies for handling class imbalance frequently rely on uploading class distributions for client-side selection or on server-side auxiliary data, which may violate privacy principles or assume unrealistic data availability. While centralized long-tail methods are relatively abundant, their transfer to flaccid systems lacking global statistics remains challenging. Therefore, effectively utilizing heterogeneous information among clients to improve long-tail class modeling capabilities without sharing raw data has become a pressing technical problem. Summary of the Invention

[0004] This invention aims to at least partially solve one of the technical problems in the aforementioned technologies. To this end, one objective of this invention is to propose a long-tail federated learning method based on positively excited noise aggregation. This method eliminates the cross-error of federated averaging through correct low-rank aggregation of dual low-rank factors, constructs a heterogeneity perception map based on the similarity of client task features, extracts distributed tail class complementary features using a structured noise construction mechanism of distant neighbor enhancement and near neighbor suppression, and combines SVD rank constraints to maintain the lightweight communication characteristics of LoRA, effectively improving the recognition accuracy of rare tail classes in federated long-tail scenarios.

[0005] A second objective of this invention is to provide a computer-readable storage medium.

[0006] The third objective of this invention is to provide a computer device.

[0007] The fourth objective of this invention is to propose a long-tail federated learning device based on positive excitation noise aggregation.

[0008] To achieve the above objectives, a first aspect of this invention proposes a long-tail federated learning method based on positively motivated noise aggregation. This method includes the following steps: broadcasting the global LoRA parameters for the current round to all participating clients, so that each client can load the global LoRA parameters locally for training; obtaining the locally updated low-rank factor matrix obtained by each client's independent training; calculating the LoRA weight increment for each client based on the low-rank factor matrix, and performing a weighted summation based on the sample size proportion of each client to obtain the correct low-rank aggregation matrix; calculating the cosine similarity between any two clients based on the low-rank factor matrix, and selecting a weighted summation for each client. The most similar clients form a nearest neighbor set. A set of the least dissimilar clients is formed to construct a client similarity graph. Based on this graph, the bad noise component and good noise component for each client are obtained, and these components are aggregated to form a global structured noise matrix. A fused weight increment matrix is ​​obtained from the correct low-rank aggregation matrix and the global structured noise matrix, and singular value decomposition is performed on this matrix to obtain a rank-constrained low-rank matrix. The rank-constrained low-rank matrix is ​​then reconstructed to obtain the updated global LoRA parameters for the next round, which are then broadcast to all participating clients until the model converges, ultimately resulting in a federated fine-tuning large model adapted for long-tailed heterogeneous scenarios.

[0009] The long-tail federated learning method based on positive excitation noise aggregation proposed in this invention eliminates the cross error of federated average by correct low-rank aggregation of dual low-rank factors, constructs a heterogeneity perception map based on the similarity of client task features, extracts distributed tail class complementary features by adopting a structured noise construction mechanism of distant neighbor enhancement and near neighbor suppression, and maintains the lightweight communication characteristics of LoRA by combining SVD rank constraints, effectively improving the recognition accuracy of rare tail classes in federated long-tail scenarios.

[0010] In addition, the long-tail federated learning method based on positive excitation noise aggregation proposed in the above embodiments of the present invention may also have the following additional technical features: Optionally, the low-rank factor matrix includes a left low-rank factor matrix. and right low-rank factor matrix ,in, This indicates the client serial number.

[0011] Optionally, the LoRA weight increment for each client is calculated according to the following formula:

[0012] in, Indicates the first LoRA weight increment for each client.

[0013] Alternatively, the correct low-rank aggregation matrix can be obtained according to the following formula:

[0014] in, Represents a correct low-rank aggregate matrix. Indicates the first The percentage of samples from each client.

[0015] Optionally, the cosine similarity between clients can be calculated using the following formula:

[0016] in, This represents the cosine similarity between the i-th client and the j-th client. express The right low-rank factor matrix of each client, express The right low-rank factor matrix of each client.

[0017] Alternatively, the global structured noise matrix can be obtained according to the following formula:

[0018]

[0019]

[0020] in, This represents the bad noise component of the i-th client. This represents the good noise component of the i-th client. Represents the i-th client A set of nearest neighbors, Represents the i-th client The distant neighbor set, This represents the similarity weight coefficient between client i and its nearest neighbor s. This represents the distance weighting coefficient between client i and its distant neighbor j. Let i represent the left low-rank factor matrix of the i-th client. Let i represent the right low-rank factor matrix of the i-th client. This represents the left low-rank factor matrix of the s-th client. Let represent the right low-rank factor matrix of the s-th client. Let the left low-rank factor matrix of the j-th client be represented. Let the right low-rank factor matrix of the j-th client be represented. This represents the global structured noise matrix.

[0021] Alternatively, the weight increment matrix after fusion can be obtained according to the following formula:

[0022] in, This represents the weight increment matrix after fusion. Represents a correct low-rank aggregate matrix. Represents the global structured noise matrix. This indicates the noise intensity control parameters.

[0023] To achieve the above objectives, a second aspect of the present invention provides a computer-readable storage medium storing a long-tail federated learning program based on positive excitation noise aggregation, which, when executed by a processor, implements the long-tail federated learning method based on positive excitation noise aggregation as described above.

[0024] To achieve the above objectives, a third aspect of the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the long-tail federated learning method based on positive excitation noise aggregation as described above.

[0025] To achieve the above objectives, a fourth aspect of the present invention proposes a long-tail federated learning device based on positively motivated noise aggregation, comprising: a sending module for broadcasting the global LoRA parameters of the current round to all participating clients, so that each client can load the global LoRA parameters into its local machine for training; an acquisition module for acquiring the locally updated low-rank factor matrix obtained independently by each client; a correct aggregation module for calculating the LoRA weight increment of each client based on the low-rank factor matrix, and performing a weighted summation based on the sample size ratio of each client to obtain a correct low-rank aggregation matrix; and a similarity graph construction module for calculating the cosine similarity between any two clients based on the low-rank factor matrix, and selecting a similarity graph for each client. The most similar clients form a nearest neighbor set. A set of the least dissimilar clients is formed to construct a client similarity graph; a noise construction module is used to obtain the bad noise component and good noise component corresponding to each client based on the client similarity graph, and to summarize the bad noise component and good noise component of all clients to obtain a global structured noise matrix; a noise injection module is used to obtain a fused weight increment matrix based on the correct low-rank aggregation matrix and the global structured noise matrix, and to perform singular value decomposition on the fused weight increment matrix to obtain a rank-constrained low-rank matrix; an iterative training module is used to reconstruct the rank-constrained low-rank matrix to obtain the updated global LoRA parameters for the next round, and broadcast them to all participating clients until the model converges, finally obtaining a federated fine-tuning large model adapted to long-tail heterogeneous scenarios. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the long-tail federated learning method based on positive excitation noise aggregation according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the network model framework for a long-tail federated learning method based on positive excitation noise aggregation according to an embodiment of the present invention. Figure 3 This is a block diagram of a long-tail federated learning device based on positive excitation noise aggregation according to an embodiment of the present invention. Detailed Implementation

[0027] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0028] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the invention to those skilled in the art.

[0029] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0030] It should be noted that most noise learning methods achieve regularization or robustness enhancement through explicit random perturbations. In contrast, positively stimulated noise (π-noise) posits that certain perturbations can reduce task complexity while preserving task-related information, thus becoming beneficial inductive biases rather than pure interference. Based on entropy analysis of representation learning, this framework provides a criterion for distinguishing between "constructive noise" and "destructive error." Unlike works that rely on explicit noise injection, this application focuses on structured perturbations implicitly generated by heterogeneous aggregation and analyzes under what conditions they can bring entropy reduction benefits.

[0031] refer to Figure 1 As shown, the long-tail federated learning method based on positive excitation noise aggregation in this embodiment of the invention includes the following steps: S101 broadcasts the global LoRA parameters for the current round to all participating clients so that each client can load the global LoRA parameters into its local machine for training.

[0032] In other words, at the start of each round of federated communication, the server broadcasts the global LoRA parameters for the current round, including the low-rank factor matrix, to all participating clients. and Each client receives the shared parameters and loads them into its local model, combining them with the frozen backbone network parameters to form a complete trainable model structure. The client only trains the LoRA low-rank parameters locally, without updating the backbone network, to ensure controllable communication overhead.

[0033] During local training, each client performs several steps of stochastic gradient descent or other optimization algorithms based on its own private data to update its local LoRA parameters, thus obtaining the updated parameters. and It should be noted that, due to the significant differences in data distribution among different clients, the updated low-rank parameters of different clients usually deviate in both direction and magnitude. This difference forms the basis for subsequent heterogeneity modeling.

[0034] After completing local training, the client updates its low-rank factor matrix. Uploaded to the server, without uploading any raw data or gradient information, thus meeting privacy protection requirements.

[0035] S102, obtain the locally updated low-rank factor matrix obtained by each client through independent training.

[0036] As an example, the low-rank factor matrix includes a left low-rank factor matrix. and right low-rank factor matrix ,in, This indicates the client serial number.

[0037] S103. Calculate the LoRA weight increment for each client based on the low-rank factor matrix, and perform a weighted summation based on the sample size ratio of each client to obtain the correct low-rank aggregation matrix.

[0038] As an example, the LoRA weight increment for each client is calculated according to the following formula:

[0039] in, Indicates the first LoRA weight increment for each client.

[0040] As an example, the correct low-rank aggregation matrix is ​​obtained according to the following formula:

[0041] in, Represents a correct low-rank aggregate matrix. Indicates the first The percentage of samples from each client.

[0042] In other words, after receiving the low-rank factor matrix from all participating clients, the server first performs "correct low-rank aggregation". This aggregation method differs from the direct average weighting in traditional FedAvg; instead, it is calculated based on the LoRA incremental structure.

[0043] The LoRA weight increment for each client can be expressed as:

[0044] The server calculates based on the sample size proportion of each client. Perform a weighted summation to obtain the correct aggregation result:

[0045] This aggregation method preserves the independent low-rank structure of each client, and does not... and Averaging them separately avoids introducing additional crossover error. This step yields... This is considered the ideal baseline aggregation result, providing a basis for subsequent noise injection.

[0046] S104, Calculate the cosine similarity between any two clients based on the low-rank factor matrix, and select for each client. The most similar clients form a nearest neighbor set. The least dissimilar clients are used to form a distant neighbor set to construct a client similarity graph.

[0047] As an example, the cosine similarity between clients is calculated according to the following formula:

[0048] in, This represents the cosine similarity between the i-th client and the j-th client. express The right low-rank factor matrix of each client, express The right low-rank factor matrix of each client.

[0049] In other words, to identify the heterogeneous structure among different clients, the server constructs a similarity graph among the clients after aggregation. Specifically, the server utilizes the low-rank factor matrices uploaded by each client to characterize the client's feature space. As a feature representation, the similarity between any two clients is calculated. In this embodiment, cosine similarity can be used as the metric, i.e.:

[0050] Based on the calculated similarity matrix, the server selects several of the most similar clients as the nearest neighbor set and several of the least similar clients as the distant neighbor set for each client, thereby constructing a similarity graph reflecting the structural differences between the clients. This graph is used to subsequently distinguish between "beneficial heterogeneous relationships" and "redundant similar relationships," and is a key step in achieving structured noise construction in this application.

[0051] S105. Based on the client similarity map, obtain the bad noise component and good noise component corresponding to each client, and summarize the bad noise component and good noise component of all clients to obtain the global structured noise matrix.

[0052] As an example, the global structured noise matrix is ​​obtained according to the following formula:

[0053]

[0054]

[0055] in, This represents the bad noise component of the i-th client. This represents the good noise component of the i-th client. Represents the i-th client A set of nearest neighbors, Represents the i-th client The distant neighbor set, This represents the similarity weight coefficient between client i and its nearest neighbor s. This represents the distance weighting coefficient between client i and its distant neighbor j. Let i represent the left low-rank factor matrix of the i-th client. Let i represent the right low-rank factor matrix of the i-th client. This represents the left low-rank factor matrix of the s-th client. Let represent the right low-rank factor matrix of the s-th client. Let the left low-rank factor matrix of the j-th client be represented. Let the right low-rank factor matrix of the j-th client be represented. This represents the global structured noise matrix.

[0056] In other words, after obtaining the client similarity map, the server constructs structured heterogeneous noise based on the structural differences between clients. The noise construction follows the principle of "far neighbor enhancement and near neighbor suppression." For each client... Two types of cross-difference structures are defined: cross-difference from distant neighboring clients, referred to as "good noise"; and cross-difference from nearby neighboring clients, referred to as "bad noise". The basic structural form is as follows:

[0057] This difference structure effectively extracts cross-client information while eliminating diagonal redundancy. The server performs a weighted summation of the difference terms in both the far neighbor and near neighbor sets to obtain the local noise component for each client. Then, the local noise from all clients is aggregated to form a global structured noise matrix. The noise term is no longer a random perturbation, but a "heterogeneous cross-information reweighted combination" with a clear structural meaning.

[0058] S106. Based on the correct low-rank aggregation matrix and the global structured noise matrix, the fused weight increment matrix is ​​obtained, and singular value decomposition is performed on the fused weight increment matrix to obtain the low-rank matrix after rank constraint.

[0059] As an example, the weight increment matrix after fusion is obtained according to the following formula:

[0060] in, This represents the weight increment matrix after fusion. Represents a correct low-rank aggregate matrix. Represents the global structured noise matrix. This indicates the noise intensity control parameters.

[0061] In other words, after obtaining the structured noise matrix The server then injects it proportionally into the correct aggregation result. The final aggregation update format is:

[0062] in, This is a noise intensity control parameter used to adjust the injection amplitude of structured noise. When At that time, it degenerates into a traditional correct aggregation; when At that time, controlled heterogeneous crossover information is introduced. Based on theoretical analysis, within a small range... Under these conditions, the structured noise can reduce the task representation entropy and improve the tail class generalization ability.

[0063] Due to noise injection, the matrix The pre-defined low-rank structure requirement may no longer be met. To ensure that the rank of the LoRA parameter does not expand, rank constraint processing is required. The server updates the matrix. Performing singular value decomposition (SVD) yields:

[0064] Then only the former was retained. The maximal singular values ​​and their corresponding singular vectors yield a rank of . The approximate matrix is ​​obtained. This operation ensures both communication efficiency and maintains the consistency of the LoRA structure. This step ensures that this application introduces structured noise without increasing model complexity or parameter size.

[0065] S107. The low-rank matrix after rank constraint is reconstructed to obtain the updated global LoRA parameters for the next round, and then broadcast to all participating clients until the model converges, finally obtaining a federated fine-tuning large model adapted to long-tail heterogeneous scenarios.

[0066] In other words, after completing the SVD rank constraint, the server reconstructs a new low-rank factor matrix based on the decomposition results. Specifically, the singular value matrix can be distributed using the square root to obtain the new:

[0067] The newly obtained low-rank factor matrix constitutes the global LoRA parameters for the next round. The server broadcasts this to all clients, initiating the next round of federated training.

[0068] Specifically, such as Figure 2As shown, it fully illustrates the complete workflow of the long-tail federated learning method based on positively motivated noise aggregation. The left side shows the client-side workflow of the federated training, which demonstrates the basic architecture of distributed training: it includes K independent clients (such as mobile phones, edge devices, etc.), each client holds local private long-tail distributed data, and only trains the LoRA low-rank adapter parameters. The backbone network of the large model is globally frozen. At the beginning of each training round, the client receives the global LoRA parameters broadcast by the server, completes several training steps based on local data, and only uploads the updated low-rank factors ( This process does not disclose any raw data or gradient information, meeting the privacy requirements of federated learning; all client upload parameters are aggregated and then fed into the server-side global aggregation module. The middle section describes: The first step is correct low-rank aggregation: the server receives all client uploads (…). After that, weights are assigned based on the sample size of each client. First, calculate the independent LoRA update for each client. Then, a weighted sum is taken to obtain the benchmark aggregation result without cross-error. This operation avoids the additional errors caused by averaging A and B separately in traditional FedAvg; the second step is noise injection fusion: the benchmark aggregation result is... Structured noise generated by the heterogeneity sensing module By coefficient Weighted fusion yields a total update with gain. ,in, The first step is to control the noise injection intensity as an adjustable hyperparameter. The third step is rank-constrained projection: Singular value decomposition (SVD) is performed on the fused update matrix, retaining only the first r largest singular values ​​and their corresponding vectors. This ensures that the rank of the output LoRA parameters is consistent with the initial setting, without increasing model complexity or communication overhead. The processed parameters will be broadcast to all clients as the global LoRA parameters for the next round. The right-hand side describes the construction of heterogeneity-aware noise, realizing the function of extracting positive excitation noise from heterogeneity: First, based on the right factor uploaded by all clients... Constructing a client-side similarity graph: As a client-side feature, pairwise cosine similarity is calculated, and the client with the highest similarity is selected. A neighbor and the least similar A distant neighbor Subsequently, noise is decomposed based on similarity relationships: bad noise is extracted from the nearest neighbor set. The updates from similar clients have a high degree of overlap, and the overlapping items are redundant interference that needs to be suppressed; good noise should be extracted from the distant neighbor set. The cross terms of clients with large differences contain rich tail class feature information, which are positive gains and need to be enhanced. Finally, the noise components of all clients are aggregated (good noise minus bad noise) to obtain global structured noise N, which is injected into the benchmark aggregation result to achieve the goal of transforming heterogeneity from training interference into long-tail performance gain.

[0069] In summary, this application proposes a long-tail federated learning method based on positively motivated noise aggregation. The overall process includes: (1) correct low-rank aggregation (2) constructing client similarity graph (3) decomposing "good noise" and "bad noise" (4) structured noise synthesis (5) injection aggregation (6) SVD rank-constrained projection. First, the standard FedAvg aggregation mechanism updates the decomposition and cross-term structure modeling: Under standard FedAvg, the LoRA factor is averaged across clients. For clarity, consider two clients (K=2) with aggregate weights of... and At this point, the LoRA update can be written as:

[0070] The last item is the ideal aggregation. Cross-client interactions that do not exist in the code are called cross noise. All off-diagonal terms are collected as theoretical cross noise.

[0071] Theoretical cross term is defined as: , And reparameterize it as follows: , This achieves controllable and weighted distribution, eliminates diagonal redundancy, and preserves heterogeneous cross-linking information.

[0072] Next, the positive excitation noise theory proves: According to linear perturbation analysis, if the representation transformation is: , The entropy change is then: , when and Under spectral constraints (such as a limited range of eigenvalues), there exists a small range of λ such that: , Furthermore, it is demonstrated that heterogeneous cross-noise can reduce task entropy, which is π-noise (positive excitation noise).

[0073] Then, the monotonicity law of heterogeneous gain is proposed: Define the heterogeneity index:

[0074] Under certain smoothness conditions and within a small perturbation range, it is proven that... , Thus, under certain conditions, we obtain the law: the higher the degree of heterogeneity, the greater the theoretical gain.

[0075] Finally, the client-side similarity graph construction and structured "good noise - bad noise" mechanism: Projection matrix based on LoRA The cosine similarity is used to calculate pairwise similarity. For each client i, select... The most similar clients are used as the nearest neighbor set. ,as well as The least dissimilar clients are used as the far neighbor set. The normalized similarity score is used as a weighting coefficient to aggregate across noise. This graph encodes the heterogeneous structure upon which the noise constructs depend.

[0076] Based on the similarity graph, two classes of local cross-noise components are defined. Bad noise characterizes interference from strongly similar clients; in contrast, good noise is defined on dissimilar clients.

[0077] Local definition:

[0078]

[0079] Global noise:

[0080] Final aggregation formula: .

[0081] As a specific embodiment, the method of this application can be specifically applied to the field of distributed traffic violation recognition. Currently, the data collected by traffic cameras on different road sections and at different locations in the city exhibits strong heterogeneous long-tail characteristics: more than 90% of the data from cameras on core main roads are common violation types such as running red lights and illegal parking (head categories), while cameras at locations such as suburban side roads and around scenic spots collect a large number of rare violation types such as non-motorized vehicles occupying motor vehicle lanes, construction vehicles spilling waste, and identification of cloned license plates (tail categories). Moreover, the data from each camera belongs to sensitive traffic management information and cannot be centrally stored for training.

[0082] Through implementation and deployment logic: Client-side deployment: A fine-tuned large model is deployed at the edge nodes of 200 traffic cameras throughout the city. All nodes share a unified Transformer backbone network (with parameters frozen), and only the LoRA low-rank adapter is trained. Each camera is trained locally based on locally collected violation data, and only the A and B low-rank factors of LoRA are uploaded to the city traffic management center server. The original images and labeled data are completely retained locally, meeting data security and privacy requirements.

[0083] Server Aggregation: The traffic management center server adopts the PIN-LoRA aggregation mechanism. First, it performs correct low-rank aggregation based on the daily average data volume of each camera to obtain a baseline update. Then, it constructs a similarity map based on the similarity of the LoRA parameters of each camera. Cameras with large differences in location and different collection scenes are marked as distant neighbors, while cameras in the same road segment and similar scenes are marked as near neighbors. Positive excitation noise is constructed through distant neighbor enhancement and near neighbor suppression. After being injected into the aggregation results according to the preset λ coefficient, it is returned to all camera nodes after SVD rank constraint.

[0084] Iterative training: A federated update is performed every 24 hours. After 5 iterations, the model converges and all camera nodes obtain unified global LoRA parameters.

[0085] The technical effects that PIN-LoRA can achieve in this scenario include: (1) Significantly improved tail category recognition rate: The recognition accuracy of rare tail violation types such as engineering vehicle spillage, cloned vehicles, and non-motorized vehicles occupying lanes has increased from 28% of the traditional FedAvg scheme to 67%, exceeding the baseline method by nearly 40 percentage points, meeting the law enforcement scenario's need to capture rare violations; the recognition rate of common head violation types remains above 95% without decline, achieving performance balance. (2) Strong engineering implementation adaptability: The entire training process does not require adjusting the LoRA rank size, and the number of parameters transmitted in each round is only 1% of the full parameter fine-tuning, without needing to modify the computing power and bandwidth resources of existing cameras; no cross-node data sharing is required, fully complying with the safety and compliance requirements of traffic management data. (3) Strong heterogeneity adaptability: When new locations are added, or the traffic characteristics of a certain road segment change seasonally (such as changes in the types of violations around scenic spots during the peak tourist season), there is no need to retrain the entire data, and only 3 rounds of iteration are needed to adapt to the new data distribution. The gain is more obvious in scenarios with higher heterogeneity, which is in line with the characteristics of dynamic changes in urban traffic data.

[0086] In addition, the present invention also proposes a computer-readable storage medium storing a long-tail federated learning program based on positive excitation noise aggregation, which, when executed by a processor, implements the long-tail federated learning method based on positive excitation noise aggregation as described above.

[0087] In addition, this invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the long-tail federated learning method based on positive excitation noise aggregation as described above.

[0088] Figure 3 This is a block diagram of an image rigorous classification device for open-world semi-supervised learning according to an embodiment of the present invention. Figure 3 As shown, the device includes: a sending module 10, an acquisition module 20, a correct aggregation module 30, a similarity graph construction module 40, a noise construction module 50, a noise injection module 60, and an iterative training module 70.

[0089] The system comprises the following modules: Sending module 10 broadcasts the global LoRA parameters for the current round to all participating clients, allowing each client to load the global LoRA parameters locally for training; Acquisition module 20 acquires the locally updated low-rank factor matrix obtained by each client's independent training; Correct aggregation module 30 calculates the LoRA weight increment for each client based on the low-rank factor matrix and performs a weighted summation based on the sample size proportion of each client to obtain the correct low-rank aggregation matrix; and Similarity graph construction module 40 calculates the cosine similarity between any two clients based on the low-rank factor matrix and selects a similarity graph for each client. The most similar clients form a nearest neighbor set. The least dissimilar clients form a distant neighbor set to construct a client similarity graph; the noise construction module 50 is used to obtain the bad noise component and good noise component corresponding to each client based on the client similarity graph, and to summarize the bad noise component and good noise component of all clients to obtain a global structured noise matrix; the noise injection module 60 is used to obtain the fused weight increment matrix based on the correct low-rank aggregation matrix and the global structured noise matrix, and to perform singular value decomposition on the fused weight increment matrix to obtain a low-rank matrix with rank constraints; the iterative training module 70 is used to reconstruct based on the low-rank matrix with rank constraints to obtain the updated global LoRA parameters for the next round, and broadcast them to all participating clients until the model converges, finally obtaining a federated fine-tuned large model adapted to long-tail heterogeneous scenarios.

[0090] It should be noted that the explanations and descriptions of the embodiments of the long-tail federated learning method based on positive excitation noise aggregation described above also apply to the long-tail federated learning device based on positive excitation noise aggregation in this embodiment, and will not be repeated here.

[0091] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0092] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0093] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0094] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0095] It should be noted that any reference signs placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.

[0096] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0097] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0098] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0099] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0100] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0101] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms should not be construed as necessarily referring to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0102] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A long-tail federated learning method based on positive excitation noise aggregation, characterized in that, Includes the following steps: The global LoRA parameters for the current round are broadcast to all participating clients so that each client can load the global LoRA parameters into its local machine for training. Obtain the locally updated low-rank factor matrix obtained by each client through independent training; The LoRA weight increment for each client is calculated based on the low-rank factor matrix, and a weighted sum is performed according to the sample size ratio of each client to obtain the correct low-rank aggregation matrix. Calculate the cosine similarity between any two clients based on the low-rank factor matrix, and select a factor for each client. The most similar clients form a nearest neighbor set. The least dissimilar clients form a distant neighbor set to construct a client similarity graph; Based on the client similarity map, the bad noise component and good noise component corresponding to each client are obtained, and the bad noise component and good noise component of all clients are summarized to obtain the global structured noise matrix. The fused weight increment matrix is ​​obtained based on the correct low-rank aggregation matrix and the global structured noise matrix, and singular value decomposition is performed on the fused weight increment matrix to obtain the low-rank matrix after rank constraint. The low-rank matrix after rank constraint is reconstructed to obtain the updated global LoRA parameters for the next round, and then broadcast to all participating clients until the model converges, finally obtaining a federated fine-tuning large model adapted to long-tail heterogeneous scenarios.

2. The long-tail federated learning method based on positive excitation noise aggregation as described in claim 1, characterized in that, The low-rank factor matrix includes the left low-rank factor matrix. and right low-rank factor matrix ,in, This indicates the client serial number.

3. The long-tail federated learning method based on positive excitation noise aggregation as described in claim 2, characterized in that, The LoRA weight increment for each client is calculated using the following formula: in, Indicates the first LoRA weight increment for each client.

4. The long-tail federated learning method based on positive excitation noise aggregation as described in claim 3, characterized in that, The correct low-rank aggregation matrix can be obtained using the following formula: in, Represents a correct low-rank aggregate matrix. Indicates the first The percentage of samples from each client.

5. The long-tail federated learning method based on positive excitation noise aggregation as described in claim 1, characterized in that, Calculate the cosine similarity between clients using the following formula: in, This represents the cosine similarity between the i-th client and the j-th client. express The right low-rank factor matrix of each client, express The right low-rank factor matrix of each client.

6. The long-tail federated learning method based on positive excitation noise aggregation as described in claim 1, characterized in that, The global structured noise matrix is ​​obtained using the following formula: in, This represents the bad noise component of the i-th client. This represents the good noise component of the i-th client. Represents the i-th client A set of nearest neighbors, Represents the i-th client The distant neighbor set, This represents the similarity weight coefficient between client i and its nearest neighbor s. This represents the distance weighting coefficient between client i and its distant neighbor j. Let i represent the left low-rank factor matrix of the i-th client. Let i represent the right low-rank factor matrix of the i-th client. This represents the left low-rank factor matrix of the s-th client. Let represent the right low-rank factor matrix of the s-th client. Let the left low-rank factor matrix of the j-th client be represented. Let the right low-rank factor matrix of the j-th client be represented. This represents the global structured noise matrix.

7. The long-tail federated learning method based on positive excitation noise aggregation as described in claim 1, characterized in that, The weight increment matrix after fusion is obtained using the following formula: in, This represents the weight increment matrix after fusion. Represents a correct low-rank aggregate matrix. Represents the global structured noise matrix. This indicates the noise intensity control parameters.

8. A computer-readable storage medium, characterized in that, It stores a long-tail federated learning program based on positive excitation noise aggregation, which, when executed by a processor, implements the long-tail federated learning method based on positive excitation noise aggregation as described in any one of claims 1-7.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the long-tail federated learning method based on positive excitation noise aggregation as described in any one of claims 1-7.

10. A long-tail federated learning device based on positive excitation noise aggregation, characterized in that, include: The sending module is used to broadcast the global LoRA parameters for the current round to all participating clients, so that each client can load the global LoRA parameters into its local machine for training. The acquisition module is used to acquire the locally updated low-rank factor matrix obtained by each client through independent training. The correct aggregation module is used to calculate the LoRA weight increment of each client based on the low-rank factor matrix, and perform weighted summation based on the sample size ratio of each client to obtain the correct low-rank aggregation matrix. The similarity graph construction module is used to calculate the cosine similarity between any two clients based on the low-rank factor matrix, and selects a similarity graph for each client. The most similar clients form a nearest neighbor set. The least dissimilar clients form a distant neighbor set to construct a client similarity graph; The noise construction module is used to obtain the bad noise component and good noise component corresponding to each client based on the client similarity map, and to summarize the bad noise component and good noise component of all clients to obtain a global structured noise matrix. The noise injection module is used to obtain the fused weight increment matrix based on the correct low-rank aggregation matrix and the global structured noise matrix, and to perform singular value decomposition on the fused weight increment matrix to obtain a low-rank matrix after rank constraint. The iterative training module is used to reconstruct the low-rank matrix after rank constraint to obtain the updated global LoRA parameters for the next round, and broadcast them to all participating clients until the model converges, finally obtaining a federated fine-tuning large model adapted to long-tail heterogeneous scenarios.

Images