Robust personalized federated learning method and system for resource-constrained internet of things
By employing methods such as anomaly detection, trusted screening, and robust aggregation, the problems of data heterogeneity and attack interference in federated learning in resource-constrained IoT are solved, enabling trusted screening and personalized adaptation in complex environments, thereby improving the security and stability of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-04-02
- Publication Date
- 2026-07-14
AI Technical Summary
Existing federated learning faces challenges such as data heterogeneity, attack interference, and personalized modeling in resource-constrained IoT scenarios, making it difficult to achieve reliable screening, robust aggregation, and personalized adaptation under conditions of non-independent and identically distributed data, partial participation, and coexistence of attacks.
By employing methods such as anomaly detection, trusted screening, historical suppression, reliability assessment, and robust aggregation, client updates are screened, evaluated, and aggregated through a central server to build personalized models, thereby enhancing the system's security and stability.
It improves the security and stability of the system in complex IoT environments, reduces the impact of malicious attacks, and enhances the robustness of the global model and the personalized adaptation capability of heterogeneous clients.
Smart Images
Figure CN122390109A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of federated learning, IoT edge intelligence, and artificial intelligence security, and particularly to a robust personalized federated learning method and system for resource-constrained IoT. Background Technology
[0002] With the rapid development of the Internet of Things (IoT) and edge intelligence, a large amount of data is continuously generated from sensors, mobile terminals, industrial equipment, and vehicle devices. Traditional centralized machine learning typically requires uploading the raw dataset to a server for unified training. However, in IoT scenarios, data is often limited by factors such as privacy protection, compliance constraints, transmission overhead, and unstable network connections, resulting in high centralized processing costs and implementation difficulties. Federated learning provides an effective path for collaborative modeling under decentralized data conditions by exchanging model updates without directly uploading the original data.
[0003] However, federated learning for resource-constrained IoT still faces significant security and reliability challenges. Since edge devices are typically deployed in open environments, attackers can manipulate the local training process to upload malicious model updates, thereby launching poisoning attacks or backdoor attacks and compromising global model performance. Meanwhile, client data in IoT scenarios often exhibits non-independent and identically distributed characteristics; updates between normal clients can already differ significantly. When combined with factors such as partial client involvement and collusion among multiple attackers, malicious updates become even more difficult to distinguish from normal heterogeneous updates.
[0004] Existing technologies have two main shortcomings: First, robust aggregation and anomaly detection methods usually rely on empirical parameters or manual thresholds, lack statistical interpretability, and are difficult to adapt to the dynamically changing IoT environment. Second, existing personalized federated learning methods usually separate robust defense from personalized modeling, making it difficult to simultaneously take into account the robustness of the global model and the personalized performance of local tasks.
[0005] Therefore, there is an urgent need for a robust personalized federated learning method and system for resource-constrained Internet of Things (IoT) to achieve trusted filtering of client updates, robust aggregation of the global model, and secure personalized adaptation to heterogeneous tasks under conditions of non-independent and identically distributed, partial participation, and coexistence of attacks. Summary of the Invention
[0006] This invention addresses the problems of heterogeneous client data, dynamic changes in participation status, and malicious update interference in existing federated learning methods. It proposes a robust and personalized federated learning method and system for resource-constrained IoT, including an anomaly detection module, a trusted screening module, a history suppression module, a reliability assessment module, a robust aggregation module, and a personalized construction module. In each round of communication, the central server receives model updates uploaded by the clients and sequentially executes anomaly detection, trusted screening, history suppression, reliability assessment, robust aggregation, and personalized construction to suppress anomalous updates, robustly optimize the global model, and personalize the tasks for heterogeneous clients. This system can be applied to federated learning scenarios consisting of a central server and multiple clients, and is suitable for scenarios where data is not independent and identically distributed, some clients participate in training, and there are potential poisoning attacks, collusion attacks, or backdoor attacks. It improves the security, stability, and applicability of federated learning systems in resource-constrained IoT scenarios, while balancing global collaborative training effects with the personalized needs of heterogeneous clients.
[0007] To achieve the above objectives, the present invention provides the following technical solution:
[0008] In a first aspect, the present invention provides a robust personalized federated learning method for resource-constrained Internet of Things, comprising the following steps:
[0009] Anomaly detection steps: The central server receives local model updates uploaded by participating clients, constructs anomaly scores for each client based on preset anomaly metrics, and standardizes the anomaly scores.
[0010] Trusted screening steps: Perform significance testing based on the standardized anomaly scores to identify and remove abnormal clients, and construct a set of trusted clients;
[0011] Historical suppression steps: Based on the historical anomaly records of each client, identify and restrict clients with persistent anomalies, and update the final set of trusted clients;
[0012] Reliability assessment steps: Perform directional consistency analysis on the model updates in the final set of trusted clients to generate reliability weights for each client;
[0013] Robust aggregation steps: Based on reliability weights, a robust loss function is used to perform weighted aggregation of model updates and update the global model;
[0014] Personalized construction steps: Based on the similarity of the client's update direction and the reliability weight, a personalized model is built for the trusted client.
[0015] Furthermore, in the anomaly detection step, the anomaly measurement includes the magnitude measurement and / or direction deviation measurement of the model update; the standardization process adopts a robust statistical method based on the median and median absolute deviation.
[0016] Furthermore, in the trusted screening step, a multiple verification screening is performed using an error detection rate control procedure; when the number of trusted clients after screening is lower than a preset lower limit, some clients are restored in order of increasing abnormality to ensure a minimum number of clients.
[0017] Furthermore, in the historical suppression step, a historical anomaly hit record is maintained for each client. The historical risk count of clients judged as anomalies in the current round is increased, and the historical count of clients not judged as anomalies is decreased by a decay factor. Based on the historical risk count, clients are marked as high-risk and their participation is restricted in subsequent rounds.
[0018] Furthermore, the reliability assessment steps include: normalizing the direction of the model update of the trusted client; randomly dividing the normalized update into multiple groups and calculating the center direction of each group; generating a reliability score based on the consistency between the update direction of each client and the center direction of each group, and converting the reliability score into a reliability weight.
[0019] Furthermore, in the robust aggregation step, the global robust update is calculated as follows:
[0020]
[0021] in, For the client Local updates, For the client Reliability weight, The robust loss function is solved using the iterative reweighted least squares method, and based on... Update global model .
[0022] Furthermore, the personalized construction step includes: selecting personalized neighbors for each trusted client based on the similarity of update directions and reliability weights among trusted clients; aggregating the neighbor models to obtain a neighbor-assisted model; and fusing the client's local model with the neighbor-assisted model to generate a personalized model.
[0023]
[0024] in, For client-side local models, For the neighbor aggregation model, This is the fusion coefficient.
[0025] Secondly, this invention proposes a robust personalized federated learning system for resource-constrained Internet of Things (IoT), comprising a central server and participating clients, wherein the central server is used to perform the initialization, distribution, aggregation, and updating of the global model, including:
[0026] Anomaly detection module: used to construct client anomaly scores based on preset anomaly metrics and perform standardization processing;
[0027] Trusted screening module: used to perform significance testing based on standardized anomaly scores and build a set of trusted clients;
[0028] Historical suppression module: used to identify and limit persistent abnormal clients based on historical anomaly records;
[0029] Reliability assessment module: used to perform directional consistency analysis on trusted client updates and generate reliability weights;
[0030] Robust aggregation module: used to perform weighted aggregation based on reliability weights using a robust loss function, and update the global model;
[0031] Personalization Module: Used to build personalized models for trusted clients based on update direction similarity and reliability weights;
[0032] The participating client is used to receive the global model, train it based on the local dataset, and upload local model updates.
[0033] Furthermore, the trusted screening module includes: an error detection rate control unit for performing multiple verification screening; and a minimum client guarantee unit for restoring a portion of clients in ascending order of anomaly severity when the number of trusted clients falls below a preset lower limit.
[0034] Furthermore, the reliability assessment module includes: a direction normalization unit, used to perform direction normalization processing on the model update of the trusted client; a random grouping unit, used to randomly divide the normalized update into multiple groups and calculate the center direction of each group; and a scoring unit, used to generate a reliability score based on the consistency between the update direction of each client and the center direction of each group and convert it into a reliability weight.
[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0036] (1) By performing trusted screening and continuous anomaly suppression on client-uploaded updates, this invention helps to reduce the adverse effects of poisoning attacks, collusion attacks and backdoor attacks on the federated training process, and improves the security and stability of the system in complex Internet of Things environments.
[0037] (2) The present invention adopts a reliable screening mechanism with statistical interpretability. Compared with existing methods that rely on empirical parameters or manual thresholds, it helps to more reasonably distinguish between abnormal clients and normal clients and reduce the risk of wrongly removing benign clients.
[0038] (3) By combining reliability assessment and robust aggregation mechanism, this invention can reduce the interference of residual abnormal updates, heavy-tailed outlier updates and cooperative attacks on the global model update direction, thereby improving the robustness of global model aggregation.
[0039] (4) Based on the robust optimization of the global model, this invention further constructs a personalized defense perception model, which helps to balance the global collaborative training effect and the personalized adaptation requirements of heterogeneous client tasks, and is more suitable for actual deployment in resource-constrained IoT scenarios. Attached Figure Description
[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0041] Figure 1 This is a schematic diagram of the overall architecture of a robust personalized federated learning system for resource-constrained Internet of Things provided in an embodiment of the present invention.
[0042] Figure 2 This is a schematic diagram of a robust personalized federated learning method for resource-constrained Internet of Things provided in an embodiment of the present invention.
[0043] Figure 3 A schematic diagram illustrating the key mechanism of a robust personalized federated learning method for resource-constrained Internet of Things provided in this embodiment of the invention. Detailed Implementation
[0044] To better understand this technical solution, the method of the present invention will be described in detail below with reference to the accompanying drawings.
[0045] Example 1
[0046] This invention provides a robust personalized federated learning system for resource-constrained Internet of Things (IoT), the overall architecture of which is as follows: Figure 1 As shown, it includes a central server and participating clients. The central server is used to perform the initialization, distribution, aggregation, and updating of the global model, including:
[0047] Anomaly detection module: used to construct client anomaly scores based on preset anomaly metrics and perform standardization processing;
[0048] Trusted screening module: used to perform significance testing based on standardized anomaly scores and build a set of trusted clients;
[0049] Historical suppression module: used to identify and limit persistent abnormal clients based on historical anomaly records;
[0050] Reliability assessment module: used to perform directional consistency analysis on trusted client updates and generate reliability weights;
[0051] Robust aggregation module: used to perform weighted aggregation based on reliability weights using a robust loss function, and update the global model;
[0052] Personalization Module: Used to build personalized models for trusted clients based on update direction similarity and reliability weights;
[0053] The participating client is used to receive the global model, train it based on the local dataset, and upload local model updates.
[0054] The trusted screening module includes: an error detection rate control unit for performing multiple verification screenings; and a minimum client protection unit for restoring a portion of clients in ascending order of anomaly severity when the number of trusted clients falls below a preset lower limit.
[0055] The reliability assessment module includes: a direction normalization unit, used to perform direction normalization processing on the model update of trusted clients; a random grouping unit, used to randomly divide the normalized update into multiple groups and calculate the center direction of each group; and a scoring unit, used to generate a reliability score based on the consistency between the update direction of each client and the center direction of each group and convert it into a reliability weight.
[0056] Example 2
[0057] The robust personalized federated learning method for resource-constrained Internet of Things provided by this invention, such as Figure 2 As shown, it includes the following steps:
[0058] Step S1: Initialize the federated learning system.
[0059] Let the client set be The number of communication rounds is The central server initializes the global model as follows: In the first In round-robin communication, the central server selects a subset of clients from all clients to participate in training. and the current global model The data is distributed to the participating clients. Each participating client performs local training based on its own local dataset, obtains local model parameters, and calculates the model update for this round. Local updates are represented as
[0060]
[0061] in, For the client In the The model parameters are displayed after local training. Each participating client uploads its corresponding model update to the central server as input for subsequent processing.
[0062] In step S2, the central server performs anomaly detection in the trusted screening process on the uploaded model update.
[0063] To identify client updates that deviate from the normal distribution, the central server first constructs anomaly scores for each participating client. Preferably, the magnitude of the model update is used as the anomaly measure, i.e.
[0064]
[0065] in, Indicates the client In the The outlier score of the round. When it is necessary to enhance the identification capability of directional anomalies or cooperative attacks, an outlier score can also be constructed by combining the updated direction deviation. Subsequently, the central server performs robust standardization on the outlier scores of all clients to reduce the impact of extreme values on the statistical discrimination results. The standardized statistic can be expressed as:
[0066]
[0067] in, The median of the abnormal scores in the current round. This represents the median absolute deviation. To prevent stable terms with a denominator of zero, this step maps the anomaly levels of different clients to a unified statistical scale, providing a basis for subsequent significance screening. Furthermore, significance indicators corresponding to each client can be constructed based on the standardized statistics.
[0068] Step S3: Construct a trusted client set.
[0069] The central server calculates the significance index for each client based on the standardized statistics corresponding to the anomaly scores of each client, and performs multiple tests based on the Benjamini-Hochberg error detection rate control procedure to determine the set of anomaly clients to be removed in the current round. The remaining clients that were not identified as abnormal are used to construct the initial set of trusted clients. To avoid insufficient trusted clients due to excessive screening when some clients participate in or are attacked at a high rate, this invention further sets up a minimum retention mechanism. When the number of trusted clients after screening falls below a preset lower limit, the central server recovers some clients from the boundary clients according to their degree of anomaly, from low to high, to ensure that subsequent robust aggregation can be executed normally. Through the above steps, the central server obtains an initial set of trusted clients and a corresponding set of trusted updates.
[0070] Step S4: Perform continuous anomaly suppression based on historical information.
[0071] The central server maintains a historical anomaly hit record for each client to identify persistent malicious or recurring abnormal behavior. For clients judged as abnormal in the current round, their historical risk count is increased; for clients not judged as abnormal, their historical count is decreased by a decay factor. If a client is repeatedly judged as abnormal in multiple consecutive rounds, the central server marks it as a high-risk client and restricts its entry into the trusted client set in subsequent rounds. By introducing a historical accumulation mechanism, this invention can not only identify single-round anomaly updates but also provide long-term suppression of more covert and longer-lasting attacks, thereby improving the stability of the trusted set. After historical anomaly suppression, the central server obtains the final trusted client set and uses the model update corresponding to the final trusted client set as input for the robust reliability assessment in step S5.
[0072] Step S5: Perform a robust reliability assessment on trusted client updates.
[0073] After steps S3 and S4, the central server only performs subsequent aggregation on model updates within the final set of trusted clients. To reduce the impact of update magnitude differences on the judgment results, the central server first normalizes the direction of each trusted client's update; then, it randomly divides these updates into several groups and calculates the center direction of each group's update. For each trusted client, the central server compares the consistency between its update direction and the center directions of each group, and forms a reliability score for that client accordingly. This score is then converted into a reliability weight used in subsequent robust aggregation. This random grouping is only used for reliability assessment to enhance the identification of colluding clients and heavy-tailed anomalous updates, and is not used as the basis for subsequent personalized neighbor partitioning. A higher reliability score indicates that the client's update is more consistent with the majority of trusted clients, and should be given a higher weight in global aggregation. In this way, the interference of a small number of anomalous updates on the aggregation direction can be reduced, allowing truly trusted updates to play a more stable role in global optimization.
[0074] Step S6: Perform robust aggregation and update the global model.
[0075] After obtaining the reliability scores of each trusted client and converting them into reliability weights, the central server further calculates the global update for this round through weighted robust estimation. Its goal is to comprehensively utilize client reliability information and residual suppression mechanisms to reduce the impact of outlier updates and residual anomalous updates on the aggregation results. The global robust update is represented as...
[0076]
[0077] in, Indicates the client Reliability weight, The robust loss function is defined. The central server uses an iterative reweighted least squares method to solve for the robust update, and then uses the obtained result... Update the global model, i.e. Through this step, the central server can maintain good robustness against heavy-tailed anomalies, coordinated attacks, and residual outliers without knowing the proportion of attackers in advance.
[0078] Step S7: Construct a personalized model for defense awareness.
[0079] After completing the global robust update, the central server continues to utilize the reliability and update direction information obtained in the previous stages to build personalized models for trusted clients. Specifically, the central server first calculates the update direction similarity among trusted clients, and then combines this with the reliability weights of each client to form neighbor candidate relationships. For each trusted client, several clients with update directions closer to its own and higher reliability are selected as personalized neighbors, and these neighbor models are aggregated to obtain the corresponding neighbor auxiliary model for that client. Finally, the client's local model and the neighbor auxiliary model are fused to generate the personalized model for that client in the [missing information - likely a specific phase or stage]. The personalized model at the end of the round, i.e.
[0080]
[0081] in, For client-side local models, For the neighbor aggregation model, This is the fusion coefficient. Through this step, the client can prioritize utilizing trusted neighbor information that is similar to its own task distribution and has been verified by defense, thereby improving personalized adaptation capabilities while reducing the risk of malicious clients spreading adverse effects through collaborative relationships.
[0082] Step S8, repeat the iteration until training ends.
[0083] The central server broadcasts the updated global model to the next round of participating clients and repeats steps S1 to S7 until the maximum number of communication rounds is reached or the preset convergence condition is met. After training, the system outputs the final global model and the personalized model set corresponding to each trusted client. Through the above process, this invention sequentially performs three stages: trust screening, robust aggregation, and personalized model construction. This allows the anomaly screening results, reliability weights, and update direction information to be continuously reused in subsequent stages, thus forming a seamless federated training process that balances the robustness of the global model with the personalized performance of the clients.
[0084] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A robust, personalized federated learning method for resource-constrained Internet of Things (IoT), characterized in that, Includes the following steps: Anomaly detection steps: The central server receives local model updates uploaded by participating clients, constructs anomaly scores for each client based on preset anomaly metrics, and standardizes the anomaly scores. Trusted screening steps: Perform significance testing based on the standardized anomaly scores to identify and remove abnormal clients, and construct a set of trusted clients; Historical suppression steps: Based on the historical anomaly records of each client, identify and restrict clients with persistent anomalies, and update the final set of trusted clients; Reliability assessment steps: Perform directional consistency analysis on the model updates in the final set of trusted clients to generate reliability weights for each client; Robust aggregation steps: Based on reliability weights, a robust loss function is used to perform weighted aggregation of model updates and update the global model; Personalized construction steps: Based on the similarity of the client's update direction and the reliability weight, a personalized model is built for the trusted client.
2. The robust personalized federated learning method for resource-constrained Internet of Things according to claim 1, characterized in that, In the anomaly detection step, the anomaly measurement includes the magnitude measurement and / or direction deviation measurement of the model update; the standardization process adopts a robust statistical method based on the median and median absolute deviation.
3. The robust personalized federated learning method for resource-constrained Internet of Things according to claim 1, characterized in that, In the trusted screening step, an error detection rate control procedure is used to perform multiple verification screenings; when the number of trusted clients after screening is lower than a preset lower limit, some clients are restored in order of abnormality from low to high to ensure a minimum number of clients.
4. The robust personalized federated learning method for resource-constrained Internet of Things according to claim 1, characterized in that, In the historical suppression step, a historical anomaly hit record is maintained for each client. The historical risk count of a client that is judged as an anomaly in the current round is increased, and the historical count of a client that is not judged as an anomaly is decreased by a decay factor. Based on the historical risk count, the client is marked as high risk and its participation is restricted in subsequent rounds.
5. The robust personalized federated learning method for resource-constrained Internet of Things according to claim 1, characterized in that, The reliability assessment steps include: normalizing the direction of the model update of the trusted client; randomly dividing the normalized update into multiple groups and calculating the center direction of each group; generating a reliability score based on the consistency between the update direction of each client and the center direction of each group, and converting the reliability score into a reliability weight.
6. The robust personalized federated learning method for resource-constrained Internet of Things according to claim 1, characterized in that, In the robust aggregation step, the global robust update is calculated as follows: ， in, For the client Local updates, For the client Reliability weight, The robust loss function is solved using the iterative reweighted least squares method, and based on... Update global model .
7. The robust personalized federated learning method for resource-constrained Internet of Things according to claim 1, characterized in that, The personalized construction steps include: selecting personalized neighbors for each trusted client based on the similarity of update directions and reliability weights among trusted clients; aggregating the neighbor models to obtain a neighbor-assisted model; and merging the client's local model with the neighbor-assisted model to generate a personalized model. ， in, For client-side local models, For the neighbor aggregation model, This is the fusion coefficient.
8. A robust personalized federated learning system for resource-constrained Internet of Things, characterized in that, It includes a central server and participating clients, whereby the central server is used to perform the initialization, distribution, aggregation, and updating of the global model, including: Anomaly detection module: used to construct client anomaly scores based on preset anomaly metrics and perform standardization processing; Trusted screening module: used to perform significance testing based on standardized anomaly scores and build a set of trusted clients; Historical suppression module: used to identify and limit persistent abnormal clients based on historical anomaly records; Reliability assessment module: used to perform directional consistency analysis on trusted client updates and generate reliability weights; Robust aggregation module: used to perform weighted aggregation based on reliability weights using a robust loss function, and update the global model; Personalization Module: Used to build personalized models for trusted clients based on update direction similarity and reliability weights; The participating client is used to receive the global model, train it based on the local dataset, and upload local model updates.
9. The robust personalized federated learning system for resource-constrained Internet of Things according to claim 8, characterized in that, The trusted screening module includes: an error detection rate control unit for performing multiple verification screenings; and a minimum client protection unit for restoring a portion of clients in ascending order of anomaly severity when the number of trusted clients falls below a preset lower limit.
10. The robust personalized federated learning system for resource-constrained Internet of Things according to claim 8, characterized in that, The reliability assessment module includes: a direction normalization unit, used to perform direction normalization processing on the model update of trusted clients; a random grouping unit, used to randomly divide the normalized update into multiple groups and calculate the center direction of each group; and a scoring unit, used to generate a reliability score based on the consistency between the update direction of each client and the center direction of each group and convert it into a reliability weight.