A robust federated learning method and system based on frequency domain screening and model rollback

Abstract

The application provides a robust federated learning method and system based on frequency domain screening and model rollback, and belongs to the technical field of federated learning, which comprises the following steps: a center medical cloud server initializes a global model and parameters, and broadcasts the model parameters to each hospital client; the client trains a local gradient update based on the global model and uses a local medical image database to extract frequency domain features; the cloud server standardizes the frequency domain features, generates adaptive perceptual hash fingerprints using a four-level quantization mechanism, screens the fingerprints of each client based on the global center fingerprint to obtain a candidate running set; the cloud server calculates the global model displacement of the current trial run based on the candidate running set, combines the historical similarity distribution of each client, and performs audit evaluation of the client; and the cloud server performs decision and rollback reintegration based on the audit evaluation result. The application blocks the irreversible damage of malicious poisoning to the global model, and improves the robustness and precision.

Description

Technical Field

[0001] This invention belongs to the field of federated learning technology, and in particular relates to a robust federated learning method and system based on frequency domain filtering and model rollback. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In cross-hospital collaborative medical image analysis scenarios, existing centralized training requires sharing raw medical images containing serious privacy concerns, violating medical data privacy regulations. Furthermore, in federated learning across institutions, the central medical cloud server cannot directly observe the image data quality and training process of each participating hospital. If a maliciously controlled medical imaging workstation in the network launches a poisoning attack that disrupts the convergence of the diagnostic model, it can cause irreversible deviations in the global medical diagnostic network, potentially leading to serious medical accidents such as misdiagnosis and missed diagnosis.

[0004] First, traditional robust aggregation schemes suffer from computational efficiency bottlenecks. As the scale of deep neural networks in medical auxiliary diagnostic models increases and the dimensionality of parameter spaces expands, distance calculations and statistical analyses at the full parameter level lead to an exponential increase in computational complexity and memory consumption. This not only makes it difficult to adapt to the computationally limited resources of hospital edge devices but also severely impairs the real-time performance of federated learning systems, affecting the accuracy of assisted diagnosis. Second, most existing anti-poisoning schemes rely on single-round, static statistical threshold screening, resulting in overly flat defense logic. This single-layer defense mechanism becomes less effective against advanced, covert poisoning attacks. Attackers can employ a "long-term infiltration, sudden burst" strategy, posing as benign nodes in early communication rounds to upload normal gradient updates to establish historical trust and lower the system's defense threshold, then suddenly launching an attack in a specific round. This covert attack, characterized by initial normalcy and later mutation, can easily bypass traditional static coarse-grained screening mechanisms that rely solely on data from the current round, leading to the leakage of medical data privacy. Finally, existing federated learning defense architectures cannot detect model drift before the model is corrupted, nor can they undo aggregated operations that have already been performed. Usually, when the model is found to be corrupted, it is retrained, which wastes a lot of resources. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a robust federated learning method and system based on frequency domain filtering and model rollback. This addresses the problems of existing federated learning defense schemes, such as enormous computational overhead when dealing with the high-dimensional parameter space of medical images, susceptibility of single-layer static filtering mechanisms to poisoning attacks, and lack of aggregation fault tolerance leading to irreversible contamination. This invention significantly reduces computational load while enabling historical dynamic tracking and multi-dimensional behavior auditing of hospital edge computing devices. It also possesses automatic model rollback capabilities in the face of sudden attacks, protecting patient privacy while improving the accuracy of hospital-aided diagnostic models.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, this invention discloses a robust federated learning method based on frequency domain filtering and model rollback, comprising: The central medical cloud server initializes the global model and parameters, and broadcasts the model parameters to each hospital client. Based on the global model, the client uses a local medical image database to train and obtain local gradient updates and extract frequency domain features. After standardizing the frequency domain features, the cloud server uses a four-level quantization mechanism to generate an adaptive perceptual hash fingerprint, selects a global central fingerprint using the median principle, and filters the fingerprints of each client based on the global central fingerprint to obtain a candidate running set. The cloud server calculates the global model displacement for this round of trial operation based on the candidate running set, and performs audit evaluation of the client by combining the historical similarity distribution of each client; The cloud server makes decisions and rollbacks based on the audit assessment results.

[0007] Secondly, this invention discloses a robust federated learning system based on frequency domain filtering and model rollback, comprising: The initialization module is configured to: initialize the global model and parameters of the central medical cloud server, and broadcast the model parameters to each hospital client. The gradient frequency domain lossless mapping module is configured such that: the client, based on the global model, uses a local medical image database to train and obtain local gradient updates and extract frequency domain features. The frequency domain structure coarse screening module is configured as follows: after the cloud server standardizes the frequency domain features, it generates an adaptive perceptual hash fingerprint using a four-level quantization mechanism, selects a global center fingerprint using the median principle, and filters the fingerprints of each client based on the global center fingerprint to obtain a candidate running set. The trial run aggregation module is configured to: the cloud server calculates the global model displacement for this round of trial run based on the candidate run set, and performs audit evaluation of the client by combining the historical similarity distribution of each client; The decision-making and rollback reassembly module is configured such that the cloud server performs decision-making and rollback reassembly based on the audit assessment results.

[0008] Thirdly, the present invention discloses an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when run by the processor, complete the steps of the robust federated learning method based on frequency domain filtering and model rollback described above.

[0009] Fourthly, the present invention discloses a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the robust federated learning method based on frequency domain filtering and model rollback described above.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention overcomes the limitations of traditional defense schemes that rely solely on single-round static statistical features. By establishing historical behavioral profiles of clients and implementing mutation audits, it can keenly detect abnormal perturbations in similarity indicators across historical dimensions. For potentially maliciously controlled imaging workstations in medical networks, this invention can effectively identify and eliminate latent malicious nodes that initially establish trust through careful camouflage during training but then suddenly launch attacks with significant scaling factors in later stages, or attempt to cause cumulative contamination by subtly skewing the model in each round. This greatly enhances the survivability of medical auxiliary diagnostic models in complex adversarial environments, preventing malicious tampering of the global medical knowledge base that could lead to misdiagnosis or missed diagnosis.

[0011] Medical auxiliary diagnosis demands extremely high model accuracy; irreversible model contamination can lead to serious medical accidents. This invention innovatively implements a trial-run-audit-rollback architecture during the aggregation phase. By conducting trial aggregation and model drift monitoring before formal updates, it empowers the system to reverse erroneous decisions. Once an unexpected shift is detected in the global medical diagnostic network, the system immediately triggers a panic alarm and rolls back to a safe snapshot state, completely blocking the path of malicious gradients being incorporated into the global model. In the medical collaborative diagnostic industry, where irreversible contamination is a concern, this ensures the absolute purity of model parameters, avoids the enormous waste of medical computing power caused by retraining after model contamination, and ultimately achieves high-precision and robust convergence of the global auxiliary diagnostic model.

[0012] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0013] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0014] Figure 1 This is a schematic diagram of the robust federated learning method based on frequency domain filtering and model rollback as described in Embodiment 1 of the present invention. Detailed Implementation

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0016] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0017] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0018] Example 1 In one or more embodiments, a robust federated learning method based on frequency domain filtering and model rollback is disclosed, targeting collaborative scenarios in medical institutions. The central server is materialized as a central medical cloud server platform, and the client is materialized as hospital edge computing devices participating in collaborative training. The local data processed is a patient medical image dataset with privacy protection restrictions, such as... Figure 1 As shown, it includes the following steps: Step S1: The central medical cloud server initializes the global model and parameters, and broadcasts the model parameters to the clients, i.e., the edge computing devices of each hospital; Specifically, the healthcare federal learning system includes Each hospital edge computing device (client) is denoted as... and a central medical cloud server The parameters of the global medical imaging diagnostic model are represented as follows: ,in This represents the total dimension of the model parameters. (Cloud server) Initialize the global model parameters for federated learning It is then broadcast to all hospital edge computing devices. The cloud server sets the learning rate. , number of communication rounds Filtering and retaining numbers Frequency domain low-frequency cutoff threshold Simultaneously initialize audit parameters: hospital edge computing device mutation threshold. Drift smoothing coefficient Minimum standard deviation tolerance and drift panic coefficient Initialize the system drift reference. .

[0019] Step S2: Based on the global model, the client uses the local medical image database to train and obtain local gradient updates and extract frequency domain features.

[0020] Specifically, after receiving the global model, the edge computing devices in each hospital train the local gradient updates based on the local medical image dataset. The local gradient updates are then flattened in one dimension, and a one-dimensional discrete cosine transform (1D-DCT) is performed on the flattened gradients. Based on the energy concentration characteristics of the discrete cosine transform, low-frequency coefficients representing the global directionality and main semantics of the model update are truncated and retained. These frequency domain coefficients are then packaged and uploaded to the cloud server in preparation for subsequent detection.

[0021] Hospital edge computing equipment Receive global model parameters Then, using the locally stored patient medical image dataset conduct Local training iterations, updating private model parameters The specific update method is as follows:

[0022] in, These are the private model parameters for the i-th hospital edge device after the e-th local iteration; These are the private model parameters for the i-th hospital edge device after the (e-1)-th local iteration; The learning rate for local training can be set to 0.01 in specific implementation scenarios; The gradient of the target loss function calculated for the i-th hospital edge device based on the local medical image dataset.

[0023] After training is complete, calculate the full local gradient update tensor:

[0024] in, Let be the local complete gradient update amount calculated by client i in round t; For client i through The parameters of the local private model after local training.

[0025] Flatten it into a one-dimensional vector ,in, Update the nth element of the one-dimensional vector for the local gradient of client i. The frequency domain coefficients are extracted by performing a one-dimensional discrete cosine transform (1D-DCT) on the vector.

[0026] in, The k-th frequency domain coefficient after one-dimensional discrete cosine transform; orthogonal normalization coefficient. Defined as:

[0027] Hospital edge computing devices based on thresholds In this implementation example, it can be set to 0.25, that is, the first 25% of the frequency domain coefficients of each client are classified as low frequency, and the last 75% of the frequency domain coefficients are classified as high frequency. The lossless low-frequency feature sequence representing the global optimization direction is extracted. Subsequently, the hospital's edge computing device packages the low-frequency feature sequence and the high-frequency feature sequence and uploads them to the cloud server. The low-frequency feature sequence is used by the cloud server to generate fingerprints and perform coarse structural screening, while the high-frequency feature sequence works in conjunction with the low-frequency sequence to perform aggregation.

[0028] This embodiment utilizes one-dimensional DCT to map the high-dimensional gradient update vector locally on the client from the spatial domain to the frequency domain. Leveraging its energy concentration characteristic, the core information of the model update is concentrated in low-frequency coefficients, while detailed information is distributed in high-frequency coefficients, providing a foundation for subsequent frequency domain structure screening and anomaly detection. Simultaneously, by utilizing the energy compression characteristics of 1D-DCT, the complex model auditing and screening process is concentrated in the extremely low-dimensional frequency domain fingerprint space, significantly reducing the communication and computational load between edge devices and the central server.

[0029] Step S3: After standardizing the frequency domain features, the cloud server generates an adaptive perceptual hash (APHash) fingerprint using a four-level quantization mechanism, selects a global center fingerprint using the median principle, and filters the fingerprints of each client based on the global center fingerprint to obtain a candidate running set.

[0030] Step S301: The central medical cloud server performs Z-score standardization and APHash fingerprint generation for low-frequency features.

[0031] Specifically, after receiving information uploaded by the hospital's edge computing devices, the central medical cloud server extracts low-frequency coefficients, performs Z-score standardization, and generates an adaptive perceptual hash (APHash). In detail, the server calculates the group mean and standard deviation across each frequency domain dimension, standardizing the coefficients for each device to eliminate "scale bias" caused by the sheer volume of data. Then, a four-level quantization mechanism is used to map continuous Z-scores to discrete structural fingerprints, precisely characterizing the gradient's structural state.

[0032] The cloud server receives data packets uploaded by edge computing devices from various hospitals and extracts low-frequency features. It then calculates the group mean for all selected hospital edge computing devices in each frequency domain dimension k. with standard deviation :

[0033]

[0034] in, To prevent division by zero of constants; The number of edge computing devices in the selected hospital. Perform Z-score normalization to align the feature distribution:

[0035] In this embodiment, after the server receives the low-frequency coefficients uploaded by each client, it performs Z-score standardization on each frequency domain dimension to eliminate the difference in gradient magnitude caused by the heterogeneity of local data among different clients, so that the subsequent structural fingerprint generation and similarity comparison have a fair statistical basis.

[0036] Subsequently, an adaptive perceptual hash (APHash) is generated using a four-level quantization mechanism to map continuous features to discrete states:

[0037] in, This is the hash fingerprint of client i in the frequency domain dimension k. This allows us to obtain the hospital edge computing device... structural fingerprint , Let i be the APHash structure fingerprint vector of client i.

[0038] Step S302: Global center selection and coarse screening of Top-K candidate set.

[0039] Specifically, based on APHash fingerprints, the server elects a global central hash and calculates the similarity score of each hospital edge computing device. The Top-K algorithm is used to initially select a set of candidate devices whose structure is closest to a benign consensus.

[0040] The server uses the median principle to select a global central fingerprint that represents a positive consensus. :

[0041] in, The total number of clients participating in federated learning.

[0042] Calculate the sum of absolute deviations between the fingerprints of each hospital's edge computing device and the global center, and normalize them into a similarity score. (Maximum deviation in one dimension is 3):

[0043] in, The structural similarity score is given to client i.

[0044] Based on similarity Sort in descending order and select the first few. The hospital's edge computing devices constitute a candidate operation set. K is an adjustable hyperparameter, which is set to 70% of the number of hospital edge computing devices selected in this round of the experiment.

[0045] This embodiment utilizes Top-K sorting to select the top K clients whose structures are closest to a benign consensus from all participating clients to form a candidate set, thereby enabling the rapid removal of malicious nodes that significantly deviate from the consensus.

[0046] Step S4: The server calculates the global model displacement for this round of trial operation based on the candidate running set, and performs hospital edge computing device mutation detection, group outlier detection, and EMA-based global drift panic detection in combination with the historical similarity distribution of edge computing devices in each hospital.

[0047] In this embodiment, the server does not directly merge the candidate set during the audit evaluation, but instead performs trial run aggregation and conducts multi-dimensional in-depth audit.

[0048] The cloud server stores a current global model security snapshot in memory. To execute the trial run aggregation, the server first processes the candidate run set. The complete frequency domain coefficients uploaded by edge computing devices in various hospitals are subjected to inverse discrete cosine transform (IDCT) to restore them to gradient vectors in the spatial domain. :

[0049] The restored one-dimensional vector is reconstructed into a gradient tensor consistent with the model structure. Then, weighted aggregation is performed to obtain a temporary global model. Calculate the parameter drift for this round of trial operation:

[0050] in, For the first The model drift generated by the aggregation during round-trip testing; This is the index for the current communication round.

[0051] And perform a multi-dimensional audit before formal aggregation, including: (1) Hospital edge computing device mutation detection: The server continuously tracks and records the similarity statistical distribution (historical mean and variance) of each client in historical rounds, and calculates the Z-score offset of the current round's similarity relative to the historical distribution. Once it falls below the set mutation threshold, it is determined that the node has undergone an abnormal mutation, which aims to accurately capture hidden malicious nodes that disguise themselves and lurk in the early stage and suddenly launch attacks in the later stage.

[0052] Specifically, targeting For each hospital edge computing device, the server calculates the offset of its current similarity relative to its historical behavior distribution to accurately capture malicious nodes that suddenly launch attacks after disguising themselves. Let the hospital edge computing devices... The similarity in this round is The average historical similarity is The historical standard deviation is The system first calculates the similarity of this round. -score offset:

[0053] If the offset satisfies If the node is identified as having undergone an abnormal mutation, it is marked as a suspected mutation node. For example, a preset mutation determination threshold. To significantly reduce server memory overhead in long-term federated learning communication, this system does not store complete historical similarity sequences. Instead, it employs a streaming online algorithm (Welford algorithm) to... The space complexity is dynamically maintained to preserve the aforementioned historical statistics. Assume the hospital's edge computing device... The total number of rounds participated in is Then the above mean With the second central moment The iterative update formula is:

[0054]

[0055]

[0056] in, This is the difference between the current round's similarity and the average historical similarity before the update; The historical similarity mean of the i-th hospital edge computing device before the update; The updated historical similarity mean of the i-th hospital edge computing device; The second-order central moment of similarity for the i-th hospital edge computing device after the update; Let be the second-order central moment of the similarity of the i-th hospital edge computing device before the update.

[0057] The historical standard deviation used for mutation detection can then be calculated: ,when Effective at time To prevent division by zero of constants.

[0058] (2) Outlier detection in a group: Calculation Mean similarity within with respect to the true standard deviation .

[0059] Introducing minimum tolerance Obtain the effective standard deviation In this example, the minimum tolerance is set to 0.02. If If it is, then it is marked as a suspected outlier node.

[0060] (3) Global panic detection: If and This triggers a system-level model drift panic alarm, in which... The drift panic coefficient; This is the moving average of the system-level model drift exponential after the (t-1)th round update.

[0061] Existing federated learning defense architectures generally lack system-level fault tolerance and correction capabilities. In traditional aggregation processes, once a covert malicious update bypasses the pre-processing statistical screening and is directly aggregated into the global model by the server, it will cause severe and irreversible contamination of the global model, seriously affecting the accuracy of medical auxiliary diagnosis. The server implements a trial run mechanism before executing formal aggregation to ensure the security and controllability of the aggregation process. First, before executing any aggregation operation, the server forcibly saves a complete snapshot of the current global model state as a baseline for safe rollback. Subsequently, the server only uses the candidate set coarsely screened in the frequency domain to perform trial aggregation: the complete frequency domain coefficients uploaded by the candidate clients are subjected to inverse discrete cosine transform (IDCT) to recover the spatial domain gradient vectors, and weighted aggregation is performed according to the local sample size of each client to obtain a temporary global model. The model parameter drift (flattened L2 norm) of this trial run is calculated to provide quantitative evidence for subsequent multi-dimensional in-depth auditing.

[0062] Step S5: The cloud server executes decisions and performs rollback and re-aggregation based on the audit results. If the audit is passed, the aggregation is confirmed to be effective, and the global model and EMA drift baseline are updated. If an abnormal mutation node is captured or a global drift alarm is triggered, the rollback mechanism is activated to restore the model snapshot state before the trial run. After removing malicious nodes, the re-aggregation operation is performed to prevent malicious poisoning attacks.

[0063] The central medical cloud server executes a poisoning prevention and fault-tolerance decision based on the results of step S4: (1) Confirm Submission: If the audit finds no suspicious nodes and no abnormal alarms are triggered, the results of this trial run are confirmed to be safe, and the server executes... And update the system drift baseline:

[0064] in, This is the drift smoothing coefficient, typically set to 0.1.

[0065] (2) Rollback and Reassembly: If an abnormal mutation node is detected or a panic alarm is triggered, the server will be immediately abandoned. Roll back the model state to .from The system precisely removes all flagged suspected hospital edge computing devices, creating a clean set. . use The frequency domain characteristics of the edge computing devices in the clean hospital are recovered into the spatial domain gradient tensor through the inverse discrete cosine transform (IDCT). Re-execute the formal weighted aggregation:

[0066] in, This represents the amount of data samples used locally on the client side for training.

[0067] And updated based on the actual displacement generated after re-aggregation. The process continues in the next round of federated communication until the preset convergence condition is met.

[0068] This embodiment employs the Exponential Moving Average (EMA) as the global model drift monitoring method. The EMA is a commonly used time series smoothing method, its core idea being to assign higher weights to recent data while retaining historical trend information. This scheme uses the EMA to maintain a baseline for the drift of model parameters across generations. In each round, the model displacement generated during the trial run is compared with the EMA baseline to detect whether abnormal drift has occurred in the global model. This allows the defense threshold to adaptively adjust as the training phase progresses, eliminating the need for frequent manual parameter tuning and achieving automated and intelligent security collaboration.

[0069] By addressing the limitations of existing technologies, such as high computational overhead when dealing with high-dimensional parameters, susceptibility to covert attacks in single-layer static defenses, and lack of aggregation and fault tolerance mechanisms, this invention employs Federated Learning as its core architecture. Federated Learning is a distributed machine learning paradigm designed to solve the data silo problem. Participating clients, while ensuring their original local data remains within their domain, utilize their local datasets to calculate model gradient updates or parameter updates. Only the updated parameters are uploaded to the central server, which aggregates and updates the global model before distributing it. This process ensures user data privacy while enabling joint training of multiple models. This invention significantly reduces computational load, provides historical dynamic tracking and multi-dimensional behavior auditing for clients, and offers automatic model rollback capabilities in the face of sudden attacks. Furthermore, it effectively resists covert attacks through historical behavior tracking and endows the federated learning system with powerful aggregation fault tolerance and automatic correction capabilities, completely preventing irreversible damage to the global model from malicious attacks and achieving high-precision, robust convergence of the global model.

[0070] Example 2 In one or more embodiments, a robust federated learning system based on frequency domain filtering and model rollback is disclosed, specifically including: The initialization module is configured to: initialize the global model and parameters of the central medical cloud server, and broadcast the model parameters to each hospital client. The gradient frequency domain lossless mapping module is configured such that: the client, based on the global model, uses a local medical image database to train and obtain local gradient updates and extract frequency domain features. The frequency domain structure coarse screening module is configured as follows: after the cloud server standardizes the frequency domain features, it generates an adaptive perceptual hash fingerprint using a four-level quantization mechanism, selects a global center fingerprint using the median principle, and filters the fingerprints of each client based on the global center fingerprint to obtain a candidate running set. The trial run aggregation module is configured to: the cloud server calculates the global model displacement for this round of trial run based on the candidate run set, and performs audit evaluation of the client by combining the historical similarity distribution of each client; The decision-making and rollback reassembly module is configured such that the cloud server performs decision-making and rollback reassembly based on the audit assessment results.

[0071] Example 3 This embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps of the robust federated learning method based on frequency domain filtering and model rollback described above.

[0072] Example 4 This embodiment provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the robust federated learning method based on frequency domain filtering and model rollback described above.

[0073] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0074] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0075] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0076] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A robust federated learning method based on frequency domain selection and model rollback, characterized in that, include: The central medical cloud server initializes the global model and parameters, and broadcasts the model parameters to each hospital client. Based on the global model, the client uses a local medical image database to train and obtain local gradient updates and extract frequency domain features. After standardizing the frequency domain features, the cloud server uses a four-level quantization mechanism to generate an adaptive perceptual hash fingerprint, selects a global central fingerprint using the median principle, and filters the fingerprints of each client based on the global central fingerprint to obtain a candidate running set. The cloud server calculates the global model displacement for this round of trial operation based on the candidate running set, and performs audit evaluation of the client by combining the historical similarity distribution of each client; The cloud server makes decisions and rollbacks based on the audit assessment results.

2. The robust federated learning method based on frequency domain selection and model rollback as described in claim 1, characterized in that, The extraction of frequency domain features specifically involves: flattening the local gradient update into a one-dimensional vector, and performing a one-dimensional discrete cosine transform (1D-DCT) on the one-dimensional vector to extract frequency domain coefficients. in, The k-th frequency domain coefficient after the one-dimensional discrete cosine transform; These are the orthogonal normalization coefficients; Update the nth element of the one-dimensional vector for the local gradient of client i; This represents the total dimension of the global medical imaging diagnostic model parameters.

3. The robust federated learning method based on frequency domain selection and model rollback as described in claim 1, characterized in that, The adaptive perceptual hash fingerprint is generated using a four-level quantization mechanism, mapping continuous features to discrete states: in, The hash fingerprint of client i in the frequency domain dimension k; The low-frequency characteristics of client i in the frequency domain dimension k are standardized.

4. The robust federated learning method based on frequency domain selection and model rollback as described in claim 1, characterized in that, Based on the global central fingerprint, the fingerprints of each client are filtered to obtain a candidate execution set, specifically: Calculate the sum of absolute deviations between each client fingerprint and the global center fingerprint, and normalize it into a similarity score: in, Score the structural similarity of client i; This is the low-frequency truncation threshold in the frequency domain; The hash fingerprint of client i in the frequency domain dimension k; As the global central fingerprint; Based on the similarity scores, clients are sorted in descending order, and the top few clients are selected to form a candidate running set.

5. A robust federated learning method based on frequency domain selection and model rollback as described in claim 1, characterized in that, The audit assessment of the execution client includes mutation detection, outlier detection, and global panic detection; The mutation detection involves the server continuously tracking and recording the similarity statistical distribution of each client in historical rounds, and calculating the Z-score offset of the current round's similarity relative to the historical distribution; when the offset is lower than the set mutation threshold, the node is determined to have undergone an abnormal mutation. The group outlier detection involves the server dynamically calculating the mean and standard deviation of the group similarity within the current candidate running set, and setting a dynamic outlier red line based on the principle of 3 times the standard deviation to eliminate abnormal isolated points within the candidate running set. The global panic detection calculates the total displacement of model parameters before and after aggregation, and compares it with the historical drift benchmark maintained based on exponential moving average. If the displacement in this round exceeds the drift benchmark, a system-level global drift panic alarm is triggered.

6. The robust federated learning method based on frequency domain selection and model rollback as described in claim 1, characterized in that, The cloud server executes decisions and rollbacks / reassemblies based on audit assessment results, specifically as follows: If all audits pass, the aggregation is confirmed to be effective, and the global model and EMA drift baseline are updated. If an abnormal mutation node is detected or a global drift panic alarm is triggered, the rollback mechanism will be activated to restore the model snapshot state before the trial run.

7. A robust federated learning method based on frequency domain selection and model rollback as described in claim 6, characterized in that, The rollback mechanism involves the server discarding the temporary global model and rolling back the model state to a global model security snapshot; accurately removing all marked suspected hospital edge computing devices from the candidate running set to form a clean set; using the frequency domain features of the clients in the clean set to recover the spatial domain gradient tensor through the inverse discrete cosine transform, re-executing the formal weighted aggregation, and updating the system drift benchmark based on the real displacement generated after the re-aggregation.

8. A robust federated learning system based on frequency domain selection and model rollback, characterized in that, include: The initialization module is configured to: initialize the global model and parameters of the central medical cloud server, and broadcast the model parameters to each hospital client. The gradient frequency domain lossless mapping module is configured such that: the client, based on the global model, uses a local medical image database to train and obtain local gradient updates and extract frequency domain features. The frequency domain structure coarse screening module is configured as follows: after the cloud server standardizes the frequency domain features, it generates an adaptive perceptual hash fingerprint using a four-level quantization mechanism, selects a global center fingerprint using the median principle, and filters the fingerprints of each client based on the global center fingerprint to obtain a candidate running set. The trial run aggregation module is configured to: the cloud server calculates the global model displacement for this round of trial run based on the candidate run set, and performs audit evaluation of the client by combining the historical similarity distribution of each client; The decision-making and rollback reassembly module is configured such that the cloud server performs decision-making and rollback reassembly based on the audit assessment results.

9. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the robust federated learning method based on frequency domain filtering and model rollback as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the robust federated learning method based on frequency domain filtering and model rollback as described in any one of claims 1-7.

Images