Target detection model training method and system for adaptive noise control
By employing adaptive noise control and dynamic training strategies, the problem of poor adaptability of target detection models in autonomous driving systems to different scenarios is addressed, thereby improving model stability and privacy protection, and achieving efficient training and generalization capabilities in federated learning.
Patent Information
- Application Number
- CN202510988953.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-17
- Publication Date
- 2025-10-24
AI Technical Summary
In autonomous driving systems, under the traditional single-vehicle training mode, the object detection model has poor adaptability to different scenarios, resulting in a decline in the safety and generalization ability of autonomous driving. Furthermore, the fixed noise strategy in federated learning fails to effectively coordinate the optimization of privacy protection and object detection model performance.
An adaptive noise control method is adopted, which dynamically adjusts the global noise intensity by calculating the smooth similarity and decay factor of the global hash fingerprint, and optimizes the training strategy by combining the client's comprehensive score and activity type to improve model stability and privacy protection.
It achieves improved training efficiency and generalization ability of target detection models while ensuring data security, balancing privacy protection and model accuracy, and adapting to the autonomous driving needs of different scenarios.
Smart Images
Figure CN120833544A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence, and in particular to a target detection model training method and system with adaptive noise control. BACKGROUND
[0002] In an autonomous driving system (ADS system), a trained target detection model is usually used to identify vehicles, pedestrians, traffic signs, drivable areas, etc. around the vehicle based on vehicle perception data, and to predict the future motion trajectory and intention of the traffic participants around the vehicle based on the target recognition result, and to decide the candidate trajectory or driving strategy (such as acceleration, deceleration, lane change, turning suggestion) of the vehicle. Vehicles equipped with an ADS system are usually equipped with sensors (such as cameras, radars, lidar, etc.) and an autonomous driving platform (with a target detection model deployed internally). Vehicle perception data is not limited to camera image sequences, lidar point clouds, and millimeter wave radar data.
[0003] In fact, due to the limitations of region, time, weather, etc., the perception data of a single vehicle cannot cover all scenarios. For example, a vehicle that has been driving in the south for a long time lacks driving data on icy roads, and when it encounters a sudden snowstorm, its environmental perception and decision-making ability will decrease significantly, resulting in a decrease in the reliability of the ADS system. In the traditional single-vehicle training mode, the adaptability of the target detection model to different scenarios is poor, which seriously affects the safety and generalization ability of autonomous driving. In related technologies, a federated learning framework is used to train the target detection model. The autonomous driving computing platform of the vehicle acts as a client of federated learning (FL). The autonomous driving computing platforms of multiple vehicles each use the historical perception data collected during their driving to collaboratively train a shared, more powerful target detection model. After training, each vehicle downloads and deploys this shared, globally optimized global target detection model from the server. When driving, the vehicle independently uses this global target detection model to process its real-time perception data, obtains the target detection result, and makes a local driving decision (such as steering, throttle, and brake).
[0004] The perception data of the client is sensitive information, but federated learning has a risk of privacy leakage. Malicious clients may infer the sensitive information of other clients by analyzing the global parameter updates (such as gradients), so to improve the security of information transmission between the server and the client, noise is usually added to the global parameter updates issued by the server and the local parameter updates uploaded by the client. For global parameter updates, related technologies mostly use a fixed noise strategy, and noise control is not combined with the update trend of the target detection model, resulting in a failure to optimize privacy protection and target detection model performance simultaneously. SUMMARY
[0005] The application aims to at least solve the technical problems in the prior art, and provides a target detection model training method and system with adaptive noise control.
[0006] In a first aspect, the application provides a target detection model training method with adaptive noise control. The method comprises the following steps performed by a server in the tth global training:
[0007] Aggregating the tth local parameter update uploaded by the client after completing the local training in the tth global training to obtain the tth global parameter update;
[0008] Hash-compressing the tth global parameter update to obtain the tth global hash fingerprint;
[0009] Calculating the tth smooth similarity according to the tth global hash fingerprint, the (t-1)th global hash fingerprint and the (t-2)th global hash fingerprint;
[0010] Calculating the tth decay factor based on the tth smooth similarity, wherein the tth decay factor is negatively correlated with the tth smooth similarity;
[0011] Obtaining the (t+1)th global noise intensity based on the tth decay factor, wherein the (t+1)th global noise intensity is positively correlated with the tth decay factor;
[0012] Adding noise to the tth global parameter update according to the (t+1)th global noise intensity to obtain the (t+1)th global release parameter;
[0013] Broadcasting the (t+1)th global release parameter to the client;
[0014] wherein t is an integer greater than or equal to 2.
[0015] In a first aspect, the application provides a system comprising a server and one or more clients in communication with the server, at least one of the one or more clients being an automatic driving platform of a vehicle, the server and the client completing target detection model training based on the target detection model training method with adaptive noise control according to any one of claims 1-9.
[0016] Beneficial technical effects of the present application: based on the federated learning framework, the target detection model training is carried out, in each round of global training, the server no longer adds noise with fixed noise intensity in the global parameter update issued, but first calculates the tth round of smooth similarity according to the tth round of global hash fingerprint, the t-1th round of global hash fingerprint and the t-2th round of global hash fingerprint, the global similarity is smoothed, avoiding the interference of the dramatic change of added noise caused by the abnormal update of single round of global training on the global strategy, which can improve the training stability, at the same time, the smooth similarity obtained after smoothing can better reflect the long-term trend of the target detection model update, then the tth round of decay factor is calculated based on the tth round of smooth similarity, the t+1th round of global noise intensity is positively related to the tth round of decay factor, it can be seen that the greater the tth round of smooth similarity, the smaller the t+1th round of global noise intensity obtained, indicating that the model training is relatively stable at this time, and less noise can achieve the purpose of privacy protection, the smaller the tth round of smooth similarity, the greater the t+1th round of global noise intensity obtained, indicating that the parameter update fluctuation is large at this time, and more noise needs to be added to achieve the purpose of privacy protection, realizing adaptive noise control, balancing privacy-target detection model precision, and improving training efficiency on the premise of ensuring data security. BRIEF DESCRIPTION OF DRAWINGS
[0017] Figure 1 is a flowchart of a target detection model training method with adaptive noise control in a preferred embodiment of the present application;
[0018] Figure 2 is a process diagram of the server obtaining the active type of the client in a preferred embodiment of the present application;
[0019] Figure 3 is a whole framework diagram of target detection model training in an application scenario of the present application;
[0020] Figure 4 is a framework diagram of the system in a preferred embodiment of the present application. DETAILED DESCRIPTION
[0021] The embodiments of the present application are described in detail below, examples of which are shown in the drawings, wherein the same or similar reference signs represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary, only for explaining the present application, and cannot be understood as limiting the present application.
[0022] In the description of the present application, it should be understood that the terms "longitudinal", "transverse", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, which are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.
[0023] In the description of the present application, unless otherwise specified and limited, it should be noted that the terms "mounting", "connection", "connection" should be understood broadly, for example, it can be mechanical connection or electrical connection, it can be the communication between the two elements, it can be direct connection or indirect connection through intermediate medium, and the specific meaning of the above terms can be understood by those skilled in the art according to the specific circumstances.
[0024] The present application discloses a target detection model training method of adaptive noise control, in a preferred embodiment, the method is: Figure 1 The server performs the following steps:
[0025] Step S1, aggregate the tth round of local parameter updates uploaded after the client completes the local training in the tth round of global training, and obtain the tth round of global parameter updates.
[0026] In the present embodiment, the server is not limited to a hardware server, a cloud server or a computer cluster. The client is not limited to an automatic driving computing platform of a vehicle, which can collect various types of perception data collected by sensors, such as camera image sequences, laser radar point clouds, millimeter wave radar data and the like. The target detection model is not limited to the existing YOLO series network, PointPillars network or VoxelNet network.
[0027] In the present embodiment, at the beginning of the tth round of global training, the server sends the tth round of global release parameters to the clients participating in the tth round of global training by broadcasting. Generally, after receiving the tth round of global release parameters, the client updates the network parameters of the t-1th round of local target detection model based on the tth round of global release parameters, trains the updated t-1th round of local target detection model using the locally stored perception data set, obtains the tth round of local target detection model, and uploads the tth round of local parameter updates to the server according to the tth round of local target detection model. After receiving the tth round of local parameter updates uploaded by all clients participating in the tth round of global training, the server is not limited to using weighted average or median aggregation and the like to obtain the tth round of global parameter updates.
[0028] Step S2, hashing compression is performed on the tth round of global parameter update to obtain a tth round of global hash fingerprint.
[0029] In the embodiment, the server only saves the low-dimensional hash fingerprints of the global parameter updates in each round of global training to realize the isolation of similarity calculation, which is not limited to using the existing SimHash algorithm to obtain the global hash fingerprints, and realizes the mapping of high-dimensional global parameter updates to a low-dimensional space for efficient representation of the key features of the global parameter updates. The calculation of the subsequent global similarity, the current similarity and the historical similarity is completely constructed in the low-dimensional hash fingerprint space, so that the calculation process is completely decoupled from the storage, transmission and processing of the original global parameter updates, and the dual optimization of data privacy and calculation efficiency is achieved. Preferably, the server constructs a fingerprint record table, and each column of the fingerprint record table stores a global hash fingerprint. In an example, the storage format of each column is: {round ID: hash fingerprint}.
[0030] Step S3, the tth round of smooth similarity is calculated according to the tth round of global hash fingerprint, the (t-1)th round of global hash fingerprint and the (t-2)th round of global hash fingerprint.
[0031] In the embodiment, preferably, step S3 includes:
[0032] Step S31, the current similarity is calculated by the tth round of global hash fingerprint and the (t-1)th round of global hash fingerprint.
[0033] Step S32, the historical similarity is calculated by the (t-1)th round of global hash fingerprint and the (t-2)th round of global hash fingerprint.
[0034] Step S33, the tth round of smooth similarity is obtained by fusing the current similarity and the historical similarity with a preset smoothing factor.
[0035] For example, the tth round of global hash fingerprint is represented as HF(U t ), the (t-1)th round of global hash fingerprint is represented as HF(U t -1 ), and the (t-2)th round of global hash fingerprint is represented as HF(U t-2 ). The Hamming distance HD(HF(U t ), HF(U t-1 )) of the tth round of global hash fingerprint and the (t-1)th round of global hash fingerprint is calculated, HD(,) represents the Hamming distance function, J represents the length of the global hash fingerprint, which is a positive integer; j represents the bit index of the global hash fingerprint, and |·| represents the absolute value. Then the current similarity sim t is: Similarly, the historical similarity sim t-1 is: The tth round of smoothing similarity S_sim t S_sim t = a sim t-1 + (1-a) sim t , where a represents a smoothing factor, a e [0, 1], and the smoothing factor is used to balance the contribution of the current similarity and the historical similarity.
[0036] In this embodiment, instead of using only the current similarity to obtain the decay factor, the smoothing similarity, that is, the fusion value of the current similarity and the historical similarity, is used as the smoothing similarity, so that the smoothing similarity can better reflect the long-term trend of the target detection model parameter update, which can avoid the interference of the abnormal update of the single round of global training on the noise coefficient, thereby improving the training stability.
[0037] Step S4, calculating the tth round of decay factor based on the tth round of smoothing similarity, wherein the tth round of decay factor is negatively correlated with the tth round of smoothing similarity.
[0038] Exemplarily, the mapping relationship between the tth round of decay factor and the tth round of smoothing similarity is not limited to a linear function with a negative slope or an inverse proportional relationship.
[0039] Step S5, obtaining the (t+1)th round of global noise intensity based on the tth round of decay factor, wherein the (t+1)th round of global noise intensity is positively correlated with the tth round of decay factor.
[0040] Exemplarily, the mapping relationship between the (t+1)th round of global noise intensity and the tth round of decay factor is not limited to a linear function with a positive slope or a positive proportional relationship.
[0041] Step S6, adding noise in the tth round of global parameter update according to the (t+1)th round of global noise intensity to obtain the (t+1)th round of global release parameter.
[0042] In this embodiment, the noise added in the tth round of global parameter update satisfies a Gaussian distribution. Exemplarily, the (t+1)th round of global release parameter is represented as:
[0043]
[0044] where N(0, σ t+1 2 ) represents a Gaussian distribution random variable with a mean of 0 and a standard deviation of the tth round of global noise intensity σ t+1 , that is, random noise.
[0045] Step S7, broadcasting the tth round of global release parameter to the client; wherein t is an integer greater than or equal to 2.
[0046] Preferably, see Figure 1 After step S7, the server further includes judging whether a global training stop condition is reached, if yes, the server ends the training, and obtains the trained target detection model based on the t-th global parameter update. The global training stop condition is not limited to t being greater than or equal to a preset maximum number of global training rounds. If not, the server enters the (t+1)-th global training. Correspondingly, in the (t+1)-th global training, the client performs: receiving the (t+1)-th global publishing parameter, and updating the network parameters of the t-th local target detection model based on the (t+1)-th global publishing parameter, training the updated t-th local target detection model by using the locally stored perception dataset, obtaining the (t+1)-th local target detection model, and obtaining the (t+1)-th local parameter update according to the (t+1)-th local target detection model. Preferably, the (t+1)-th local parameter update is uploaded to the server after adding noise, so as to increase data security.
[0047] In a preferred embodiment, in step S4, the t-th decay factor is calculated based on the t-th smooth similarity, including:
[0048] In step S41, a power adjustment parameter β is set according to the privacy protection level.
[0049] In the embodiment, the power adjustment parameter β is greater than 0, and the privacy protection level includes a high privacy protection level and a low privacy effective update level. When the high privacy protection level is used, β>1, and when the low privacy effective update level is used, 0<β<1. The power adjustment parameter β is mainly used to control the sensitivity of the decay factor to the global similarity.
[0050] In step S42, the t-th decay factor is calculated according to the following formula:
[0051]
[0052] Wherein, α1 represents a first decay rate, and α1>0. The greater α1 is, the faster the noise intensity decreases with the increase of the smooth similarity; S_sim t represents the t-th smooth similarity. β controls the concave-convex of the decay curve. When β>1, the higher the smooth similarity is, the faster the decay factor decreases than when 0<β<1. The decrease of the decay factor leads to a sharp decrease of the global noise intensity. Therefore, the decay factor curve is steeper in the high smooth similarity area. When the smooth similarity increases slightly, the decay factor rapidly decreases, and the global noise intensity sharply decreases, thereby highlighting the effective update, which is suitable for scenarios that emphasize effective update. When 0<β<1, the decay factor curve is generally smoother. The fluctuation of the decay factor caused by the change of the smooth similarity is small, and the noise decreases slowly. In the low smooth similarity area, a larger noise is still retained to strengthen privacy protection, which is suitable for scenarios that emphasize privacy protection.
[0053] In one example, a1=1. When β=2, when S_sim t is 0.25, ξ t is calculated as 0.94, when S_sim t is 0.81, ξ t is calculated as 0.6, and it can be seen that the smoothing similarity S_sim t decreases by 0.34 from low to high. When β=0.5, when S_sim t is 0.25, ξ t is calculated as 0.67, when S_sim t is 0.81, ξ t is calculated as 0.53, and it can be seen that the smoothing similarity S_sim t decreases by 0.14 from low to high.
[0054] In the present embodiment, adaptive noise control is implemented, the attenuation factor of the tth round adopts a fractional power attenuation function, and the noise change can be made gentle.
[0055] In a preferred embodiment, in step S5, the t+1th global noise intensity σ t+1 is obtained based on the tth attenuation factor.
[0056] σ t+1 = σ min + (σ max - σ min )·ξ t
[0057] wherein σ max represents a noise intensity upper limit value, σ min represents a noise intensity lower limit value, and ξ t represents the tth attenuation factor.
[0058] In the present embodiment, the noise intensity lower limit value σ min is set as a noise lower limit protection, to avoid the case that the noise may approach zero when the smoothing similarity is extremely high, resulting in failure of noise protection, and to ensure that even in the most extreme case, the basic privacy protection requirement can be met. Through the above noise intensity calculation formula, the privacy-precision is balanced, and the calculation efficiency is improved. While ensuring data security, the communication efficiency is improved, the privacy leakage risk is reduced, and the efficiency and stability of data transmission in the vehicle networking environment are effectively improved.
[0059] Existing technologies rely solely on the similarity or data volume of the current round to screen clients, without establishing a multi-dimensional evaluation system. This makes it impossible to dynamically adapt to the client's actual needs, resulting in low resource utilization. In a preferred embodiment, a dynamic training control method based on global-historical collaborative filtering is provided, which is implemented by the server. Specifically, in the tth round of global training, the server also performs the following steps:
[0060] Step S8, obtaining a comprehensive score of the i-th client in the t-th round based on at least one of the four indicators of the i-th client in the t-th round: global similarity, data stability, historical contribution, and computing power.
[0061] Among them, the global similarity of the i-th client in the t-th round is The tth round of local parameter updates uploaded by the i-th client Hash fingerprint of The similarity HF(U t ),Right now:
[0062]
[0063] Global similarity of the i-th client in round t The consistency between the client and the global trend can be measured. The updates provided by highly similar clients are more in line with the global optimization direction.
[0064] Among them, the data stability of the tth round of the i-th client It is calculated based on the historical local verification accuracy of the i-th client. It is dynamically adjusted based on changes in local verification accuracy to avoid frequently fluctuating clients dominating training. The calculation formula is:
[0065]
[0066] represents the local verification accuracy obtained by the k-th round of global training on the i-th client, represents the local verification accuracy of the i-th client in the k+1th round of global training. The local verification accuracy is reported in plain text by the client and does not affect parameter privacy. T represents the sliding window size and the range of rounds used to calculate the stability index. |·| represents the absolute value, which is used to capture the fluctuation of the accuracy.
[0067] Among them, the historical contribution of the i-th client in the t-th round is Obtained based on the multi-round historical global similarity of the i-th client. Clients with good long-term performance contribute more to the model and should continue to receive priority participation.
[0068]
[0069] wherein, θ represents a similarity threshold value; is an indicator function, when , the indicator function takes the value 1, when , the indicator function takes the value 0.
[0070] wherein, the computing capacity Comp i of the i-th client is calculated by the following formula:
[0071]
[0072] Reported_FLOPS i represents the reported CPU computing power and / or GPU computing power of the i-th client. C represents a client set.
[0073] In the embodiment, preferably, the comprehensive score of the i-th client in the t-th round is a weighted sum of the global similarity, the data stability, the historical contribution degree and the computing capacity of the i-th client in the t-th round, forming a comprehensive client evaluation framework. Specifically, the comprehensive score of the i-th client in the t-th round is calculated by the following formula:
[0074]
[0075] w1, w2, w3, w4 represent the first weight coefficient, the second weight coefficient, the third weight coefficient and the fourth weight coefficient respectively.
[0076] The four indicators of global similarity (measuring the consistency of the client with the global trend, the updates provided by the client with high similarity are more consistent with the global optimization direction, and should be given higher weight), data stability (based on accuracy change dynamic adjustment, avoiding frequent fluctuation of the client leading the training), historical contribution degree (the client with long-term good performance contributes more to the model, and should continue to obtain the priority participation right) and computing capacity (in actual deployment, the client with strong computing capacity can complete the training faster, and should be given more participation opportunities to improve the overall efficiency) are comprehensively considered to evaluate the client.
[0077] Step S9, obtaining the active type of the i-th client in the t+1-th round according to the comparison result of the comprehensive score of the i-th client in the t-th round and the active threshold value.
[0078] Step S10, sending the active type of the i-th client in the t+1-th round to the i-th client.
[0079] Correspondingly, in the t-th round of global training, the i-th client performs local training based on the training strategy corresponding to the active type of the i-th client in the t-th round; wherein, i is a client index, and is a positive integer.
[0080] In the embodiment, preferably, the activity threshold value comprises a first activity threshold value and a second activity threshold value, the first activity threshold value is greater than the second activity threshold value; the first activity threshold value and the second activity threshold value can be artificially set.
[0081] See Figure 2 According to the comparison result of the comprehensive score of the i-th client in the t-th round and the activity threshold value, the server obtains the global activity type of the i-th client in the t+1-th round, including:
[0082] When the comprehensive score of the i-th client in the t-th round is greater than or equal to the first activity threshold value, the activity type of the i-th client in the t+1-th round is a high activity client, and the training strategy corresponding to the high activity client is: the high activity client does not perform local training, and uploads the local parameter update obtained by the high activity client in the global training in the last round (for example, if the current is the t-th round of global training, the last round is the t-1-th round of global training), that is, does not perform training, directly reuses the historical local parameter update, reduces the local training calculation amount and communication overhead, and can reduce the requirement for the computing power and energy consumption of the vehicle client.
[0083] When the comprehensive score of the i-th client in the t-th round is less than the first activity threshold value and greater than or equal to the second activity threshold value, the activity type of the i-th client in the t+1-th round is a medium activity client, and the training strategy corresponding to the medium activity client is: the medium activity client only adjusts the output layer parameters of the target detection model in local training, and the output layer can be a fully connected layer. Compared with complete training, the parameter update amount is less, and the communication data amount is compressed.
[0084] When the comprehensive score of the i-th client in the t-th round is less than the second activity threshold value, the activity type of the i-th client in the t+1-th round is a low activity client, and the training strategy corresponding to the low activity client is: the low activity client adjusts all layer parameters of the target detection model in local training.
[0085] In the embodiment, the training strategy is dynamically adjusted according to the activity type of the client, the resource utilization of the client is optimized, the real demand of the client can be adapted, the client with high-quality data but limited computing resources is not ignored, the client with strong computing ability but poor data quality is not selected frequently, the waste of invalid computing resources is avoided, the generalization ability of the target detection model is improved, and the resource utilization rate is improved.
[0086] In order to dynamically switch the first activity threshold value and the second activity threshold value according to the global training round and the verification accuracy of the global target detection model, and achieve the purpose of stage adaptive threshold adjustment by using the training strategy of “first focusing and then generalization”. In a preferred embodiment, in the t-th round of global training, the server also performs:
[0087] Calculate the first activity threshold value
[0088]
[0089] Computing the second active threshold
[0090]
[0091] wherein τ 0,high represents an initial threshold of the first active threshold; τ 0,mid represents an initial threshold of the second active threshold; α high represents a time decay coefficient of the first active threshold; α mid represents a time decay coefficient of the second active threshold; β high represents a precision compensation coefficient of the first active threshold; β mid represents a precision compensation coefficient of the second active threshold. represents the validation accuracy of the t-th round of global target detection model, and e represents a natural constant.
[0092] In the embodiment, α high and α mid are used to control the speed of the first active threshold and the second active threshold over time. β high and β mid are the adjustment strength of the target detection model to the first active threshold and the second active threshold. When the performance of the target detection model is poor (the validation accuracy is low), the first active threshold and the second active threshold are increased to relax the conditions (more updates are introduced); when the performance of the target detection model is good (the validation accuracy is high), the first active threshold and the second active threshold are reduced to maintain the screening standard. α high , α mid , β high and β mid may be set artificially according to experiments, business needs or experience.
[0093] In the embodiment, the activity determination threshold (the first active threshold and the second active threshold) is dynamically adjusted according to the global training round and the validation accuracy of the global target detection model, so as to realize the purpose of "focusing first and then generalization enhancement", that is, the first active threshold filters high-quality updates in the early stage, and the second active threshold introduces diverse updates in the later stage.
[0094] Specifically, the model parameters are randomly initialized at the beginning of training, and the uneven distribution of data has a great impact on convergence. Strict screening can ensure that only high-quality clients (i.e., clients with high comprehensive scores) participate, avoiding the dominance of noise data in the global model; as the training progresses and the model gradually converges, more diverse data is needed to participate to improve the generalization ability, and relaxing the threshold can include more contributions from clients to prevent the model from overfitting.
[0095] In the initial stage of global training: high-quality clients are equivalent to high-active clients.
[0096] Since both the initial first active threshold and the second active threshold are high, only high-quality clients with high comprehensive scores will be classified as high-active, so they are different descriptions of the same group in the initial stage (high-quality: attribute of comprehensive score; high-active: result of classification based on the first active threshold and the second active threshold)
[0097] In the middle and later stages of training: the range of high-active clients expands.
[0098] As training progresses, the dynamic first active threshold and the second active threshold decrease, and more moderately comprehensive-score clients are classified as high-active, but at this time, high-quality clients still belong to the core group of high-active.
[0099] It can be seen that the initial stage of global training emphasizes “high-quality client participation”, which does not require it to perform local training in each round, but achieves contribution through reuse of historical updates. This approach not only reduces resource consumption, but also ensures that the global model is always optimized based on high-quality data. Only clients that show high-quality updates through complete training (low-active strategy) in the early global training rounds can accumulate a high enough comprehensive score (such as: historical contribution, which depends on the number of historical rounds with a similarity exceeding a threshold) in subsequent rounds, and then be classified as high-active. If a client has never participated in complete training, it cannot obtain a high comprehensive score.
[0100] In this embodiment, the model parameters are randomly initialized in the initial training stage, and uneven data distribution has a large impact on convergence. Strict screening can ensure that only high-quality clients participate, avoiding the dominance of noisy data in the global model; as the model gradually converges, more diverse data is needed to participate to improve the generalization ability, and relaxing the threshold can include more contributions from clients to prevent model overfitting.
[0101] In the prior art, the data distribution robustness of clients is insufficient, and there is a lack of active adaptation mechanism, relying only on passive indicators such as model accuracy decline to trigger data enhancement or retraining, lacking an active data distribution monitoring mechanism, and being unable to adjust the training strategy in a timely manner to cope with dynamic changes (such as label shift in Non-IID scenarios). Based on this, in a preferred embodiment, in the tth round of global training, the ith client performs the following steps before performing local training:
[0102] Step D1, calculate the KL divergence of the current local data distribution of the ith client with respect to the historical local data distribution of the ith client, and the KL divergence is represented as D KL(P||Q), wherein P is the current local data distribution of the i-th client, and Q is the historical local data distribution of the i-th client.
[0103] Step D2, mapping the KL divergence to the numerical interval [0, 1] to obtain a mapped value as the drift proportion. The mapping function can be linear mapping. Preferably, the drift proportion Drift_Prop is calculated as follows:
[0104]
[0105] wherein D KL_max is a preset maximum KL divergence, such as the KL divergence value calculated when P is a completely uniform distribution and Q is an extremely skewed distribution as D KL_max , representing the "theoretical maximum difference". min(·, 1) is to prevent the proportion from exceeding the limit when the KL divergence is abnormally large, ensuring that Drift_Prop ∈ [0, 1].
[0106] Step D3, adjusting the active type of the i-th client in the t-th round according to the drift proportion, and the client performs local training according to the training strategy corresponding to the adjusted active type. Specifically, the active type of the i-th client in the t-th round is adjusted according to the "proportion interval-active level" correspondence relationship shown in Table 1 in step D3.
[0107] Table 1 "Proportion Interval-Active Level" Correspondence Table
[0108]
[0109]
[0110] In the embodiment, the client detects the data distribution change through the KL divergence, i.e., performs data drift detection, automatically triggers complete training and reports a warning to the server, the server adjusts the active type of the client according to the drift proportion, enhances the robustness of the global target detection model to sudden changes in data, effectively deals with the Non-IID data scene, and improves the model generalization ability. When the data distribution of the client changes suddenly, the existing scheme does not appear the problem of deviation of the global target detection model caused by lag response, and even converges to a local optimum, which can effectively avoid the local optimum.
[0111] Figure 3 An application scenario of a target detection model training is shown.
[0112] The application also discloses a system. In a preferred embodiment, please see Figure 4The system comprises a server and one or more clients in communication with the server, at least one of the one or more clients being an automatic driving platform of a vehicle, and the server and the client complete training of the target detection model based on the adaptive noise control target detection model training method.
[0113] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", "an implementation", "a preferred implementation" or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.
[0114] Although the embodiments of the present application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the claims and their equivalents.
Claims
1. A method for training a target detection model with adaptive noise control, characterized in that, In the tth round of global training, the server performs the following steps: aggregate the tth round of local parameter updates uploaded by the clients after completing local training in the tth round of global training to obtain the tth round of global parameter updates; hash compress the tth round of global parameter updates to obtain the tth round of global hash fingerprints; calculate the tth round of smooth similarity according to the tth round of global hash fingerprints, the (t-1)th round of global hash fingerprints and the (t-2)th round of global hash fingerprints; calculate the tth round of decay factor based on the tth round of smooth similarity, wherein the tth round of decay factor is negatively correlated with the tth round of smooth similarity; obtain the (t+1)th round of global noise intensity based on the tth round of decay factor, wherein the (t+1)th round of global noise intensity is positively correlated with the tth round of decay factor; add noise in the tth round of global parameter updates according to the (t+1)th round of global noise intensity to obtain the (t+1)th round of global release parameters; broadcast the (t+1)th round of global release parameters to the clients; wherein t is an integer greater than or equal to 2. 2.The method of claim 1, wherein, The calculation of the tth round of smooth similarity according to the tth round of global hash fingerprints, the (t-1)th round of global hash fingerprints and the (t-2)th round of global hash fingerprints includes: calculate the current similarity by the tth round of global hash fingerprints and the (t-1)th round of global hash fingerprints; calculate the historical similarity by the (t-1)th round of global hash fingerprints and the (t-2)th round of global hash fingerprints; fuse the current similarity and the historical similarity by using a preset smoothing factor to obtain the tth round of smooth similarity.
3. The method of claim 1, wherein, The calculation of the tth round of decay factor based on the tth round of smooth similarity includes: set the power adjustment parameter β according to the privacy protection level; calculate the tth round of decay factor according to the following formula: wherein a1 represents a first decay rate; S_sim t represents the tth round of smoothing similarity.
4. The method of claim 1, wherein, The t+1th global noise intensity σ is obtained based on the tth decay factor t+1 : σ t+1 = σ min + (σ max - σ min ) · ξ t where σ max represents an upper limit value of noise intensity, σ min represents a lower limit value of noise intensity, ξ t represents a tthround attenuation factor.
5. A method of training a target detection model for adaptive noise control according to one of claims 1 to 4, characterized in that, In the tth round of global training, the server also performs the following steps: obtain the comprehensive score of the i th client in the tth round according to at least one of the four indicators of the global similarity of the i th client in the tth round, data stability, historical contribution and computing power, wherein the global similarity of the i th client in the tth round is the similarity between the hash fingerprint of the tth round of local parameter updates uploaded by the i th client and the tth round of global hash fingerprints, the data stability of the i th client in the tth round is calculated based on the historical local verification accuracy of the i th client, and the historical contribution of the i th client in the tth round is obtained according to the historical global similarity of the i th client; obtain the active type of the i th client in the (t+1)th round according to the comparison result of the comprehensive score of the i th client in the tth round and the active threshold; send the active type of the i th client in the (t+1)th round to the i th client; In the tth round of global training, the i th client performs local training based on the training strategy corresponding to the active type of the i th client in the tth round; wherein i is the client index and is a positive integer.
6. The method of claim 5, wherein the target detection model is trained using a self- adaptive noise control. The comprehensive score of the i th client in the tth round is the weighted sum of the global similarity, data stability, historical contribution and computing power of the i th client in the tth round.
7. The method of claim 5, wherein the target detection model is trained using a self- adaptive noise control. The active threshold includes a first active threshold and a second active threshold, and the first active threshold is greater than the second active threshold; The comparison result of the comprehensive score of the i-th client in the t-th round and the active threshold value is obtained. The global active type of the i-th client in the t+1-th round is obtained, including: When the comprehensive score of the i-th client in the t-th round is greater than or equal to the first active threshold value, the active type of the i-th client in the t+1-th round is a high active client, and the training strategy corresponding to the high active client is that the high active client does not perform local training, and uploads the local parameter update obtained by the high active client in the last round of global training. When the comprehensive score of the i-th client in the t-th round is less than the first active threshold value and greater than or equal to the second active threshold value, the active type of the i-th client in the t+1-th round is a medium active client, and the training strategy corresponding to the medium active client is that the medium active client only adjusts the output layer parameters of the target detection model in local training. When the comprehensive score of the i-th client in the t-th round is less than the second active threshold value, the active type of the i-th client in the t+1-th round is a low active client, and the training strategy corresponding to the low active client is that the low active client adjusts all layer parameters of the target detection model in local training.
8. The method of claim 7, wherein the target detection model is trained using a self- adaptive noise control. In the t-th round of global training, the server further performs: computing a first active threshold computing a second activity threshold wherein τ 0,high denotes an initial threshold value of the first active threshold; τ 0,mid denotes an initial threshold value of the second active threshold; α high denotes a time decay coefficient of the first active threshold; α mid denotes a time decay coefficient of the second active threshold; β high denotes a precision compensation coefficient of the first active threshold; β mid denotes a precision compensation coefficient of the second active threshold; denotes a validation accuracy of the t-th round global target detection model, and e denotes a natural constant.
9. The target detection model training method of claim 6 or 7 or 8, wherein, In the t-th round of global training, the i-th client further performs the following steps before performing local training: Calculate the KL divergence of the current local data distribution of the i-th client with respect to the historical local data distribution of the i-th client; Map the KL divergence to the numerical interval [0, 1], and use the mapping value as the drift ratio; Adjust the active type of the i-th client in the t-th round according to the drift ratio.
10. A system, characterized by The system comprises a server and one or more clients in communication with the server, at least one of the one or more clients being an automatic driving platform of a vehicle, and the server and the client complete target detection model training based on the adaptive noise control target detection model training method of any one of claims 1-9.
Citation Information
Cited By
Federal learning-based human body activity identification method and system for Internet of Things equipment
CN121167329A