Trust-driven federated learning approach
By employing a trust-driven federated learning approach, a comprehensive client score is calculated and sensitive layers are identified. Weights are dynamically adjusted, addressing the model instability caused by differences in client data quality and behavioral credibility, thus improving the robustness and efficiency of federated learning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI NORMAL UNIV
- Filing Date
- 2026-04-20
- Publication Date
- 2026-07-31
AI Technical Summary
In federated learning systems, differences in client data quality, computing power, and behavioral credibility lead to unstable model training. Some clients may upload low-quality or malicious updates, affecting the training stability and prediction performance of the global model.
We adopt a trust-driven federated learning approach, which identifies sensitive layers and performs differentiated aggregation by calculating the comprehensive score and aggregate weights of clients. We dynamically adjust the weights, suppress abnormal updates, encourage the participation of high-quality clients, and improve the robustness and generalization ability of the model.
It improves the robustness and generalization ability of the federated learning system, suppresses the negative impact of malicious clients, dynamically adapts to data heterogeneity during the training process, and improves overall training efficiency and model performance.
Smart Images
Figure CN122491544A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of federated learning technology, and more specifically, this invention relates to a trust-driven federated learning method. Background Technology
[0002] With the increasing demand for data privacy protection, federated learning, as a distributed machine learning framework, allows multiple participants to collaboratively train models without sharing the original data. It achieves joint modeling by exchanging model update information, thus balancing data utilization efficiency and privacy protection to some extent. Since McMahan et al. proposed the FedAvg algorithm, federated learning has received widespread attention in scenarios such as healthcare, finance, and intelligent recommendation.
[0003] However, in real-world federated learning systems, clients typically come from different institutions or terminal devices, resulting in variations in data quality, computing power, and behavioral credibility. Some clients may upload low-quality model updates due to poor data quality, insufficient training, or system malfunctions. In more complex scenarios, some participants may also upload abnormal updates through data poisoning or model poisoning, thereby affecting the training stability and predictive performance of the global model. Summary of the Invention
[0004] This invention provides a trust-driven federated learning method aimed at improving at least one of the aforementioned problems.
[0005] This invention is implemented as follows: a trust-driven federated learning method, the method is as follows:
[0006] (1) The client receives the global model parameters sent by the server and sets the current local model parameters. Update the received global model parameters, train the local model based on the local dataset, and upload the trained local model parameters to the server;
[0007] (2) The server calculates the comprehensive score of each client and the aggregate weight corresponding to the comprehensive score based on the local model parameters uploaded by the client and the current global model parameters;
[0008] (3) Aggregate the uploaded local model parameters of each layer based on the aggregation weight to form global model parameters of sensitive and non-sensitive layers, and send the current global model parameters to all clients.
[0009] Furthermore, the specific method for identifying sensitive layers in the global model is as follows:
[0010] (31) Evaluate the accuracy of the current global model on the validation set. ;
[0011] (32) Add Gaussian noise to the l-th layer of the global model to form a perturbed global model, and evaluate the accuracy of the perturbed global weight model on the same validation set. Calculate the sensitivity score of layer l. Restore the global model parameters of layer l and assign the sensitivity score of layer l. Add to list Se;
[0012] (33) Order Return to step (32) until the sensitivity scores of all layers of the global model are calculated, and return to list Se;
[0013] (34) The sensitivity scores in the normalized list Se are the normalized sensitivity scores. Greater than the set sensitivity threshold When the time comes, the l-th layer is marked as the sensitive layer S.
[0014] Furthermore, the overall score of the i-th client. The specific calculation formula is as follows:
[0015] ;
[0016] in, This represents the quality score of the i-th client. This represents the contribution score of the i-th client. This represents the trust score of the i-th client. The weight value, and .
[0017] Furthermore, in the current round hour, Total rounds, current round weight value The specific calculation formula is as follows:
[0018] ;
[0019] ;
[0020] ;
[0021] Current round At that time, the current round The weight value below The specific calculation formula is as follows:
[0022] ;
[0023] ;
[0024] ;
[0025] in, This represents the variance of the accuracy of the global model parameters over the most recent N rounds.
[0026] Furthermore, the aggregate weight of the i-th client It is expressed as follows:
[0027] ;
[0028] in, Indicates the temperature coefficient. This represents the overall score of all clients participating in this round of aggregation, excluding the i-th client, and the temperature coefficient for the current round t. .
[0029] Furthermore, the formation process of the global model parameters for the sensitive and non-sensitive layers is as follows:
[0030] Through normalized aggregate weights Local model parameters for each client at the same layer A weighted average is performed to form the initial global model parameters for each layer. ;
[0031] Extract the penultimate layer local model parameters from the local models uploaded by the i-th and j-th clients. , Calculate the feature similarity between the i-th client and the j-th client. Based on feature similarity Calculate the initial global model parameters Adjustment amount Adjusted global model parameters ;
[0032] The parameters of the global model at layer l are represented as follows:
[0033] ;
[0034] in, , The weight value, and .
[0035] Further adjustments The specific calculation formula is as follows:
[0036] ;
[0037] in, Indicates the strength parameter, Indicates the feature similarity threshold; This indicates that the feature similarity is greater than the similarity threshold. The number of clients; Indicates weight, ; This represents the pruning difference between the i-th client and the j-th client on the local model parameters at layer l. , This represents the global model parameters of the current layer l. This indicates a cropping process.
[0038] Furthermore, quality rating The specific calculation formula is as follows:
[0039] ;
[0040] in, This represents the accuracy weight for the current round t. This represents the stability weight for the current round t. This represents the trend weight in the current round t. This indicates a smoothing term to prevent the denominator from being zero.
[0041] Furthermore, the contribution score of the i-th client. The specific calculation formula is as follows:
[0042] ;
[0043] in, , The weights representing the update magnitude and the consistency weights are... This represents the update magnitude of the i-th client. This represents the directional consistency score of the i-th client;
[0044] Calculate the l-th layer local model parameters for the i-th client. and the global model parameters of the current layer l Difference range Normalized difference magnitude The parameter update magnitude of the i-th client is obtained by taking the arithmetic mean of the normalized difference magnitudes of all layers. The update magnitude score of the i-th client ;
[0045] Directional Consistency Score Calculate the l-th layer local model parameters for the i-th client. Relative to the global model parameters of the current layer l The l-th layer parameter update vector , Calculate the parameter update vector of the l-th layer. Cosine similarity to the current global model vector at layer l The direction matching degree of the i-th client is obtained by taking the arithmetic mean of the cosine similarity of all layers. .
[0046] Furthermore, the trust score of the i-th client. The specific calculation formula is as follows:
[0047] ;
[0048] This represents the historical stability score of the i-th client. This represents the abnormal score of the i-th client. This represents the distribution consistency of the i-th client;
[0049] Historical stability score The specific calculation formula is as follows:
[0050] ;
[0051] in, Let represent the variance of the accuracy of the local model for the i-th client in the most recent N rounds.
[0052] The abnormal score of the i-th client The specific calculation formula is as follows:
[0053] ;
[0054] in, Let represent the accuracy of the local model of the i-th client in the most recent round. This represents the average accuracy of the local model for the i-th client over the most recent N rounds.
[0055] Distribution consistency of the i-th client The specific calculation formula is as follows:
[0056] ;
[0057] in, This represents the number of sample categories on the i-th client's local machine. This indicates the number of sample categories in the server. This represents the set of sample labels.
[0058] The trust-driven federated learning method provided by this invention has the following beneficial technical effects:
[0059] (1) The server evaluates the predictive performance of each client model and calculates the behavioral similarity index by analyzing its historical update trajectory, thereby forming a comprehensive trust score, which serves as the weight basis for global model aggregation. The higher the score of the client, the greater the impact of its local model update on the global model. The purpose is to improve the robustness and generalization ability of the federated learning system. In addition, the trust mechanism can suppress the negative impact of abnormal or malicious clients on the global model as much as possible; encourage clients with excellent performance and data distribution that are more in line with the global distribution to participate in aggregation; dynamically adapt to data heterogeneity and changes in client behavior during the training process, and improve the overall training efficiency and model performance. (2) In view of the sensitivity differences of each layer in the model, a sensitive layer identification and differential aggregation mechanism is designed. When it is found that some layers have a greater impact on the verification performance, a conservative aggregation strategy is used for these layers. By perturbing the weights of each layer and evaluating the performance decline, the aggregation weights of the sensitive layers are dynamically adjusted to prevent unstable updates from damaging the model. Attached Figure Description
[0060] Figure 1 A flowchart of a trust-driven federated learning method provided in an embodiment of the present invention. Detailed Implementation
[0061] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.
[0062] Figure 1 The flowchart of the trust-driven federated learning method provided in this embodiment of the invention is as follows:
[0063] (1) The client receives the global model parameters sent by the server. , local model parameters Update to global model parameters The local model is trained based on the local dataset. After local training is complete, the parameters of the trained local model are... Uploaded to the server;
[0064] (2) The server is based on the local model parameters uploaded by each client. The overall score for the client is calculated based on the current global model parameters, forming the aggregate weight for the corresponding client;
[0065] In this embodiment of the invention, the comprehensive score of the i-th client participating in federated learning is... It consists of three parts, namely, the client's quality score. Contribution score and trust rating Overall score The specific calculation formula is as follows:
[0066] ;
[0067] in, The weight value, and .
[0068] In this embodiment of the invention, the weight value Adaptive adjustments were made in earlier rounds, and in the current round... hour, Total rounds, current round weight value The specific calculation formula is as follows:
[0069] ;
[0070] ;
[0071] ;
[0072] In the initial stage, the weight of quality scores will be reduced, while the weight of contribution scores and trust scores will be increased to encourage exploration. The value range is [0.25, 0.40]. The value range is [0.375, 0.45]. The value range is [0.225, 0.30].
[0073] Later As the model converges, the focus shifts to leveraging high-quality updates to improve final performance. At this point, the weight of the quality score will gradually increase, while the weights of the contribution score and trust score will decrease accordingly. The weight value below The specific calculation formula is as follows:
[0074] ;
[0075] ;
[0076] ;
[0077] in, This represents the variance of the accuracy of the global model parameters over the most recent N rounds.
[0078] The dynamic weight adjustment mechanism described above ensures that the most appropriate aggregation focus is adopted at different stages of training, balancing the breadth of model exploration and the depth of convergence. Overall Score Through dynamic weighting, clients that are considered "more valuable" at each stage have been preliminarily selected. However, the overall score... The absolute value differences may be large (the initial client ratings varied significantly), and directly using a standard softmax algorithm could lead to over-concentration or over-dispersion of weights. The temperature coefficient T further "calibrates" the weight allocation by adjusting the distribution steepness, making it more consistent with... The adjustment goals are consistent. A Softmax function with a temperature coefficient T is used to comprehensively score each client. Normalization is performed to form the aggregate weight of each client. The aggregate weight of the i-th client It is expressed as follows:
[0079] ;
[0080] in, This represents the temperature coefficient, which changes dynamically with each cycle. This represents the overall score of all clients participating in this round of aggregation, excluding the i-th client.
[0081] Early stage from low temperature (from Initially, the overall ratings of clients may vary significantly. It's crucial to quickly identify and utilize high-quality clients while suppressing low-quality or malicious clients. Gradually increase the temperature later. ( The temperature coefficient of the current round t. Some low-rated clients may possess unique data distributions that can be mined. Furthermore, it doesn't rely solely on a few clients, but rather leverages the local knowledge from all clients comprehensively.
[0082] In this embodiment of the invention, quality scoring Contribution score and trust rating The specific formation method is as follows:
[0083] (21) Quality rating Calculate a quality score based on the accuracy of the local model on the local validation set and training stability, and calculate the accuracy of the local model for the i-th client in the most recent round. Fluctuations in accuracy and accuracy increment (i.e., the difference in accuracy between the first and last rounds), quality score The specific calculation formula is as follows:
[0084] ;
[0085] in, This represents the accuracy weight for the current round t. This represents the stability weight for the current round t. This represents the trend weight in the current round t. This indicates a smoothing term to prevent the denominator from being zero. .
[0086] Weight , , The aggregation round changes dynamically with the aggregation round t, where t is the current aggregation round: earlier ( Increase the weight of stability and trend in the later stages. This increases the weight of accuracy and focuses on high-performance clients.
[0087] In the early stages, the weights at round t Weight Weight The specific calculation formula is as follows:
[0088] ;
[0089] ;
[0090] ;
[0091] Later weighting Weight Weight .
[0092] (22) Contribution score The contribution score reflects the value of client updates to the global model and consists of an update magnitude score and a directional consistency score. The contribution score for the i-th client... The specific calculation formula is as follows:
[0093] ;
[0094] in, , The weights representing the update magnitude and the consistency weights are... This represents the update magnitude of the i-th client. This represents the directional consistency score of the i-th client.
[0095] Calculate the l-th layer local model parameters for the i-th client. and the global model parameters of the current layer l Difference range To eliminate the impact of differences in parameter scale at different levels, Normalization was performed, and the magnitude of the difference after normalization was determined. The parameter update magnitude of the i-th client is obtained by taking the arithmetic mean of the normalized difference magnitudes of all layers. , The update magnitude score of the i-th client ;
[0096] Directional Consistency Score Using Sigmoid mapping: Calculate the l-th layer local model parameters for the i-th client. Relative to the global model parameters of the current layer l The l-th layer parameter update vector , Calculate the parameter update vector of the l-th layer. Cosine similarity to the current global model vector at layer l Measure the orientation matching degree of the l-th layer. cosine similarity The larger the value, the more consistent the directions. The direction matching degree of the i-th client is obtained by taking the arithmetic mean of the cosine similarities of all layers. , .
[0097] In this embodiment of the invention, weight Weight Dynamically changes with each round, in the early stage ( (Give higher weight to directional consistency to encourage diverse exploration, including weights for update magnitude and consistency at round t.) , The specific calculation formula is as follows:
[0098] ;
[0099] ;
[0100] As the global model approaches its convergence region, to accelerate convergence and fully utilize high-quality updates, the weights... and The values were set to 0.7 and 0.3. This ratio was set based on the following considerations: in the later stages of optimization, the update magnitude often contributes more directly and significantly to the convergence speed than subtle alignment of directions. Based on empirical observations, in the later stages of training, when the model is close to the optimal solution, ablation experiments on multiple benchmark tasks showed that the contribution ratio of the two was approximately 7:3, at which point the model achieved the best convergence performance.
[0101] (23) Trust rating This is used to evaluate client reliability, based on the client's historical behavioral stability, current performance anomalies, and data distribution consistency. The historical stability score of the local model is also considered. The formula for calculating the consistency of the behavior of the i-th client is as follows:
[0102] ;
[0103] in, Let represent the variance of the accuracy of the local model for the i-th client in the most recent N rounds (e.g., N=3). The smaller the value, the more stable the client's behavior, and the higher the score. The anomaly score for the i-th client. The specific calculation formula is as follows:
[0104] ;
[0105] in, Let represent the accuracy of the local model of the i-th client in the most recent round. This represents the average accuracy of the local model for the i-th client over the most recent N rounds.
[0106] The normalization metric, similar to univariate Mahalanobis distance, measures the degree to which client performance deviates from the overall distribution. The closer the client accuracy is to the mean, the higher the anomaly score. Furthermore, data quality is assessed by evaluating the KL divergence between the client data distribution and the global reference distribution, thus obtaining the distribution consistency. The score is obtained by calculating the KL divergence between two distributions, representing the distributional consistency of the i-th client. The specific calculation formula is as follows:
[0107] ;
[0108] in, This represents the number of sample categories on the i-th client's local machine. This indicates the number of sample categories in the server. This represents the set of sample labels.
[0109] In summary, trust scoring Since historical stability and anomaly directly reflect the reliability of client performance, the weight of historical stability and anomaly is set to 40%. Distribution consistency is used as an auxiliary indicator to avoid excessive reliance on data distribution. The weight of distribution consistency is set to 20%. This weight configuration is optimal in various Non-IID scenarios and balances the consideration of the client's historical performance and current state.
[0110] (3) Aggregate the uploaded local model parameters of each layer based on the aggregation weight to form global model parameters of sensitive and non-sensitive layers. Send the current global model parameters to all clients and return to step (1) until the aggregation round reaches the set aggregation round.
[0111] In this embodiment of the invention, the method for identifying sensitive layers in the current global model is as follows:
[0112] After calculating the aggregate weights for each client, the server uses its own validation set (a portion of the minst dataset) to identify sensitive layers. The specific identification method is as follows:
[0113] (31) Evaluate the accuracy of the current global model (without noise) on the validation set. —Input the validation set from the server into the current global model to obtain the prediction results. Compare the prediction results with the true labels of the validation set to calculate the accuracy of the current global model. ;
[0114] (32) Based on the weights of the l-th layer of the global model norm and parameter size Calculate the standard deviation of noise , ,in, This represents the number of weight parameters in the l-th layer of the global model. Indicates hyperparameters, The generated values have a mean of 0 and a standard deviation of 0. Gaussian noise ;
[0115] (33) Add Gaussian noise to the l-th layer of the global model. The parameters of the l-th layer global model after adding noise are... A perturbed global model is formed, and the accuracy of the perturbed global weight model is evaluated on the same validation set. Calculate the sensitivity score for layer l: Restore the weights of the l-th layer to And score the sensitivity of the lth layer. Add to list Se;
[0116] (34) Order Return to step (32) until the sensitivity scores of all layers of the global model have been calculated, and return to list Se. , This represents the total number of layers in the global model.
[0117] Because the scale parameters and functions of different layers may differ, the absolute values of the sensitivity scores for each layer in list Se may also differ. Direct cross-layer comparisons can lead to judgment bias. Therefore, it is necessary to normalize the sensitivity scores for each layer in table Se to the [0,1] interval. After normalization, the sensitivity score of the l-th layer is used as... This means that the closer the normalized value is to 1, the higher the sensitivity of the corresponding layer to model performance. The sensitivity score for layer l is... Greater than the set sensitivity threshold If the layer is sensitive, then the layer 1 is marked as sensitive layer S; otherwise, it is not marked and is considered a non-sensitive layer.
[0118] Sensitive threshold Dynamically adjusts with each round, current round Sensitive threshold The normalized sensitivity score of the current round Above the sensitivity threshold The layer is marked as a sensitive layer. : In subsequent aggregation processes, a conservative update strategy is adopted for this layer, and non-sensitive layers directly use the aggregated weights.
[0119] The server collects all local model parameters uploaded by clients and aggregates the weights using softmax normalization. Through normalized aggregate weights Local model parameters for each client at the same layer A weighted average is performed to form the initial global model parameters for each layer. The specific expression is as follows:
[0120] ;
[0121] in, This represents the l-th layer of the local model parameters uploaded by the i-th client.
[0122] To further utilize client information with similar data distributions, the initial global model parameters are adjusted based on feature representation similarity. Adjustments were made to enhance the global model's ability to capture similar data patterns. The server validation dataset was used to extract the penultimate layer local model parameters from the local models uploaded by the i-th and j-th clients. , Feature extraction is achieved through forward propagation from the model input layer to the penultimate layer. The feature similarity between the i-th client and the j-th client is calculated using cosine similarity, and the specific calculation formula is as follows:
[0123] ;
[0124] in, To prevent smooth terms with a denominator of zero, The closer the feature similarity value is to 1, the more similar the data distribution is between the i-th client and the j-th client. The initial global model parameters are then adjusted based on the feature similarity. For each layer l, calculate the initial global model parameters. Adjustment amount Adjusted global model parameters Among them, adjustment amount The specific calculation formula is as follows:
[0125] ;
[0126] in, Indicates the strength parameter, ; Indicates the feature similarity threshold; This indicates that the feature similarity is greater than the similarity threshold. The number of clients; Indicates weight, ; This represents the pruning difference between the i-th and j-th clients in the l-th layer local model parameters. To prevent interference from outliers, the l-th layer local model parameters of the i-th and j-th clients are... Perform cropping: , This represents the global model parameters of the current layer l. This indicates a pruning process, where clients with more similar data distributions have a higher weighting for the impact of parameter differences on the global model, and the adjustment range is greater.
[0127] If the l-th layer is identified as a sensitive layer, the global model parameters are not used directly. Instead, it adopts a conservative update strategy (initial global model parameters). With the adjusted global model parameters The weights are mixed in a ratio of 0.2:0.8 to ensure stability. The global model parameters of the l-th layer are represented as follows:
[0128] ;
[0129] in, , The weight value, and Finally, the global model parameters for each layer are obtained.
[0130] Sensitive layers are crucial to model performance; drastic changes in their weights can lead to significant fluctuations or a decline in model performance. Assigning a higher proportion of old weights ( = 0.8) helps maintain the stability of these key layer parameters and prevents them from being overly affected by potential negative updates caused by data heterogeneity or noise. This approach represents a cautious acceptance of new aggregated information; the system allows sensitive layers to learn from new training data, but the learning "steps" are small, ensuring updates are gradual and smooth. Too low... This could cause the sensitive layer to learn too slowly, making it difficult to adapt to real changes in data distribution; while an excessively high [sense level]... A value of 0.5 might not effectively mitigate the risks associated with unstable updates. 0.2 provides a compromise that is considered relatively robust in most cases. Other non-sensitive layers directly use the finely tuned value. As the updated model parameters of layer l, they suppress the impact on the core layer when facing abnormal or poisoned updates, thereby improving the overall robustness of the model.
[0131] The trust-driven federated learning method provided by this invention has the following beneficial technical effects:
[0132] (1) The server evaluates the predictive performance of each client model and calculates the behavioral similarity index by analyzing its historical update trajectory, thereby forming a comprehensive trust score, which serves as the weight basis for global model aggregation. The higher the score of the client, the greater the impact of its local model update on the global model. The purpose is to improve the robustness and generalization ability of the federated learning system. In addition, the trust mechanism can suppress the negative impact of abnormal or malicious clients on the global model as much as possible; encourage clients with excellent performance and data distribution that are more in line with the global distribution to participate in aggregation; dynamically adapt to data heterogeneity and changes in client behavior during the training process, and improve the overall training efficiency and model performance. (2) In view of the sensitivity differences of each layer in the model, a sensitive layer identification and differential aggregation mechanism is designed. When it is found that some layers have a greater impact on the verification performance, a conservative aggregation strategy is used for these layers. By perturbing the weights of each layer and evaluating the performance decline, the aggregation weights of the sensitive layers are dynamically adjusted to prevent unstable updates from damaging the model. (3) The introduced trust score mechanism provides a basis for the aggregation decision on the server side, making the federated learning process no longer a complete "black box" and helping to understand the contribution of each client. (4) The trust assessment mechanism can effectively identify and suppress malicious or abnormal client updates. The sensitive layer identification and differentiated aggregation strategy adopts a more conservative update for the key layers that affect the global performance in the model, reducing the possibility of unstable updates destroying the important features that have been learned.
[0133] The present invention has been described by way of example. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.
Claims
1. A trust-driven federated learning method, characterized in that, The method is as follows: (1) The client receives the global model parameters sent by the server, updates the current local model parameters to the received global model parameters, trains the local model based on the local dataset, and uploads the trained local model parameters to the server. (2) The server calculates the comprehensive score of each client and the aggregate weight corresponding to the comprehensive score based on the local model parameters uploaded by the client and the current global model parameters; (3) Aggregate the uploaded local model parameters of each layer based on the aggregation weight to form global model parameters of sensitive and non-sensitive layers, and send the current global model parameters to all clients.
2. The trust-driven federated learning method of claim 1, wherein, The specific method for identifying sensitive layers in the global model is as follows: (31) evaluating the accuracy of the current global model on the validation set ; (32) Add Gaussian noise to the l-th layer in the global model to form a perturbed global model, and evaluate the accuracy of the perturbed global weight model on the same validation set , calculate the sensitivity score of the l-th layer , restore the global model parameters of the l-th layer, and add the sensitivity score of the l-th layer to the list Se; (33) Order Return to step (32) until the sensitivity scores of all layers of the global model are calculated, and return to list Se; (34) The sensitivity scores in the normalized list Se are the normalized sensitivity scores. Greater than the set sensitivity threshold When the time comes, the l-th layer is marked as the sensitive layer S.
3. The trust-driven federated learning method as described in claim 1, characterized in that, Overall rating of the i-th client The specific calculation formula is as follows: ； in, This represents the quality score of the i-th client. This represents the contribution score of the i-th client. This represents the trust score of the i-th client. The weight value, and .
4. The trust-driven federated learning method as described in claim 3, characterized in that, In the current round hour, Total rounds, current round weight value The specific calculation formula is as follows: ； ; ； Current round At that time, the current round The weight value below The specific calculation formula is as follows: ； ； ； in, This represents the variance of the accuracy of the global model parameters over the most recent N rounds.
5. The trust-driven federated learning method as described in claim 1, characterized in that, The aggregate weight of the i-th client It is expressed as follows: ； in, Indicates the temperature coefficient. This represents the overall score of all clients participating in this round of aggregation, excluding the i-th client, and the temperature coefficient for the current round t. .
6. The trust-driven federated learning method as described in claim 1, characterized in that, The formation process of global model parameters for sensitive and non-sensitive layers is as follows: Through normalized aggregate weights Local model parameters for each client at the same layer A weighted average is performed to form the initial global model parameters for each layer. ; Extract the penultimate layer local model parameters from the local models uploaded by the i-th and j-th clients. , Calculate the feature similarity between the i-th client and the j-th client. Based on feature similarity Calculate the initial global model parameters Adjustment amount Adjusted global model parameters ; The parameters of the global model at layer l are represented as follows: ； in, , The weight value, and .
7. The trust-driven federated learning method as described in claim 6, characterized in that, Adjustment amount The specific calculation formula is as follows: ； in, Indicates the strength parameter, Indicates the feature similarity threshold; This indicates that the feature similarity is greater than the similarity threshold. The number of clients; Indicates weight, ; This represents the pruning difference between the i-th and j-th clients on the local model parameters at layer l. , This represents the global model parameters of the current layer l. This indicates a cropping process.
8. The trust-driven federated learning method as described in claim 3, characterized in that, Quality rating The specific calculation formula is as follows: ； in, This represents the accuracy weight for the current round t. This represents the stability weight for the current round t. This represents the trend weight in the current round t. This indicates a smoothing term to prevent the denominator from being zero.
9. The trust-driven federated learning method as described in claim 3, characterized in that, Contribution score of the i-th client The specific calculation formula is as follows: ； in, , The weights representing the update magnitude and the consistency weights are... This represents the update magnitude of the i-th client. This represents the directional consistency score of the i-th client; Calculate the l-th layer local model parameters for the i-th client. and the global model parameters of the current layer l Difference range Normalized difference magnitude The parameter update magnitude of the i-th client is obtained by taking the arithmetic mean of the normalized difference magnitudes of all layers. The update magnitude score of the i-th client ; Directional Consistency Score Calculate the l-th layer local model parameters for the i-th client. Relative to the global model parameters of the current l-th layer The l-th layer parameter update vector , Calculate the parameter update vector of the l-th layer. Cosine similarity to the current global model vector at layer l The direction matching degree of the i-th client is obtained by taking the arithmetic mean of the cosine similarities of all layers. .
10. The trust-driven federated learning method as described in claim 3, characterized in that, Trust score of the i-th client The specific calculation formula is as follows: ； This represents the historical stability score of the i-th client. This represents the abnormal score of the i-th client. This represents the distribution consistency of the i-th client; Historical stability score The specific calculation formula is as follows: ； in, Let represent the variance of the accuracy of the local model for the i-th client in the most recent N rounds. The abnormal score of the i-th client The specific calculation formula is as follows: ； in, Let represent the accuracy of the local model of the i-th client in the most recent round. This represents the average accuracy of the local model for the i-th client over the most recent N rounds. Distribution consistency of the i-th client The specific calculation formula is as follows: ； in, This represents the number of sample categories on the i-th client's local machine. This indicates the number of sample categories in the server. This represents the set of sample labels.