Reliable federated learning method and system based on AC-GAN and dynamic probability scheduling

Abstract

The invention belongs to the technical field of wireless communication, and discloses a reliable federated learning method and system based on AC-GAN and dynamic probability scheduling, which can resist negative influence of label noise, non-independent identically distributed data and poisoning attacks on global model performance, and comprises two stages: a data preprocessing stage, a dynamic probability scheduling stage and a dynamic probability scheduling stage. A server trains an AC-GAN model on a small reference data set and deploys the AC-GAN model at a user side, and a user implements data cleaning and data enhancement by means of the AC-GAN model, so that noise tag correction is realized while the influence of non-independent identically distributed data is relieved; in the attack detection stage, the parameter server constructs a user selection probability model, and malicious users are filtered based on a dynamic probability scheduling strategy. In order to prevent leakage of user model information and speculation of user information by a semi-honest server, model parameters are encrypted by using a CKKS homomorphic encryption scheme, and a double-server architecture is adopted. According to the method, the influence of label noise, non-independent identically distributed data and poisoning attacks on the performance of the model is overcome.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication technology, and in particular relates to a reliable federated learning method and system based on AC-GAN and dynamic probability scheduling. Background Art

[0002] In recent years, the development of artificial intelligence has had a profound impact in many fields, and it has become one of the key driving forces for today's scientific and technological progress. Video recommendation algorithms, smart wearable devices, AI large models, medical image recognition technology, etc. are widely used in real-world fields, changing people's lifestyles, improving production efficiency and quality of life. In the development and application of artificial intelligence, data plays a vital role. The development of artificial intelligence requires the support of a large amount of data, and the application of artificial intelligence will lead to the generation of a large amount of data. With the emergence of a large amount of data, how to obtain and use this large amount of data has become a bottleneck in the development of artificial intelligence. Machine learning, as an artificial intelligence technology, can show good learning ability for a large amount of data by collecting data and training high-performance machine learning models. However, as time goes by, people realize that data does not always exist in a unified and complete form that can be analyzed on a large scale. On the contrary, data often presents small-scale, fragmented characteristics and is scattered in different places. For example, with the continuous emergence of terminal devices, the data generated by these devices are stored in scattered places, and it is difficult for institutions to collect these fragmented data at will. At the same time, users' concerns about privacy and data security make it more difficult to aggregate all data together and process them. In this context, a distributed machine learning solution to deal with decentralized data and privacy security issues has emerged - federated learning. Federated learning allows multiple users to perform local training on their own terminal devices and collaboratively build a high-performance machine learning model, avoiding the collection of all users' private data on a central server, reducing the overhead of data collection and data transmission, and meeting users' requirements for privacy and data confidentiality. The training process of federated learning is based on the idea of ​​distributed learning, generally including a central server and multiple participants. The training process usually includes the following steps: i) Before the training starts, the central server initializes a shared global model and sends it to all users participating in the federated learning training; ii) Users use local data to train the global model, and send local model parameters to the server after the training is completed; iii) The server aggregates the model parameters of all users, forms an update of the global model and sends it to all users. Repeat steps iii) and iii) until the model converges or reaches the preset stop condition. In this way, federated learning becomes a key technology for building machine learning models that are highly sensitive to data privacy.

[0003] However, in real-world research and applications, user data often have different levels of noise labels. Noisy data introduces additional errors, and the global model may tend to learn these noises rather than true patterns or laws. Since the characteristics of the data are not controlled by the outside world, the data of different users may be statistically heterogeneous. Moreover, there may be malicious users among the participants in federated learning, who perform poisoning attacks and Byzantine attacks by submitting customized local updates. In addition, although federated learning has unique advantages in protecting data privacy, its own security and privacy issues have attracted widespread attention. For example, semi-honest servers may actively infer the sensitive data of participants, causing user privacy leakage.

[0004] Federated learning applied in wireless network environment is also called wireless federated learning. Due to the constraints of wireless network, wireless federated learning will face new challenges. First, wireless network may be affected by interference, signal attenuation or unstable connection, which will cause data packet loss or delay in the transmission process of model update, thus affecting the accuracy and convergence speed of the model; second, the wireless network environment is an open space, and all modulated information of wireless communication information communicators is exposed in the public space. Therefore, data transmission in wireless network is more vulnerable to eavesdropping or tampering than data transmission in wired network. The above problems will adversely affect the federated learning process, resulting in reduced accuracy of the global model or even failure to converge. Therefore, taking appropriate measures in wireless federated learning to enhance data quality control, effectively deal with heterogeneous data problems and improve system security is of great value to enhance the reliability of federated learning, ensure data security and improve the convergence speed of the model.

[0005] Through the above analysis, the problems and defects of the prior art are as follows:

[0006] (1) In real-world research and applications, user data often has different levels of noise labels. Noisy data introduces additional errors, and the global model may tend to learn these noises rather than the true patterns or laws.

[0007] (2) Since the characteristics of data are not controlled by the outside world, the data of different users may be statistically heterogeneous. Moreover, there may be malicious users among the participants of federated learning who can perform poisoning attacks and Byzantine attacks by submitting customized local updates.

[0008] (3) Although federated learning has unique advantages in protecting data privacy, its own security and privacy issues have attracted widespread attention. For example, semi-honest servers may actively infer the sensitive data of participants, causing user privacy leakage. Summary of the invention

[0009] In view of the problems existing in the prior art, the present invention provides a reliable federated learning method and system based on AC-GAN and dynamic probability scheduling.

[0010] The present invention is implemented as follows: a reliable federated learning method based on AC-GAN and dynamic probability scheduling includes:

[0011] Step 1: The server trains the AC-GAN model on a small benchmark dataset and deploys it on each user device;

[0012] Step 2: The user cleans the local data set using the AC-GAN model to form a clean data set;

[0013] Step 3: The user performs data enhancement on the local data set using the AC-GAN model to form an enhanced data set;

[0014] Step 4: The coordination server constructs a user selection probability table. The user uses local data for training to form a local model and sends the encrypted local model parameters to the coordination server.

[0015] Step 5: The coordination server performs model noise addition on the encrypted model and sends it to the parameter server. The parameter server performs malicious attack detection and model aggregation to form a noisy global model, and re-encrypts the global model and sends it to the coordination server.

[0016] Step 6: Coordinate the server to perform model denoising and model distribution;

[0017] Step 7: After a certain number of communication rounds, the coordination server constructs a malicious user set and a benign user set based on the user selection probability table.

[0018] Furthermore, the reliable federated learning method based on AC-GAN and dynamic probability scheduling specifically includes:

[0019] Step 1: Key generation. The key generation center KGC first generates a unique key pair for the parameter server PS and all federated learning UEs. After generating the key pair, KGC distributes the private key to each owner through a secure communication channel, and the public key is made public. After the key pair is generated and distributed, all parties in the system can start encrypted communication and data exchange. KGC is also responsible for key maintenance and update. During the key usage cycle, if key leakage or other security risks are found, KGC needs to generate a new key pair in time, and redistribute and make the process public.

[0020] Step 2: Initialize the coordination server CS. CS first initializes the model parameters related to the federated learning task and trains an AC-GAN model related to the task. CS is responsible for maintaining a high-quality small benchmark dataset, which is used to train the AC-GAN model.

[0021] Step 3: Data cleaning DC;

[0022] Includes: noise identification, quality assessment, noise correction and post-cleaning verification;

[0023] Step 4: Data enhancement (DA);

[0024] The specific process is as follows: the UE first counts the distribution of each category and the number of samples in the local data set to evaluate the diversity of the data set; the UE determines the enhancement threshold based on its own computing power and storage space, and uses the generator of the AC-GAN model to generate new data samples for each category in the data set; the purpose of this step is to balance the category distribution and improve the diversity of the data set; then, the newly generated data samples are sampled and fused with the original data samples to form a data set after DA; finally, the UE evaluates the diversity of the data set again;

[0025] Step 5: UE local training;

[0026] In this stage, each UE optimizes the initial global model using its own dataset;

[0027] Step 6: CKKS encryption;

[0028] When the UE sends the encryption model to the CS, the CS performs noise processing on the received encrypted data;

[0029] Step 7: Attack detection;

[0030] PS and CS jointly maintain a user selection probability table to characterize the credibility of federated learning UEs. In each round of model update, PS first decrypts the model parameters of each UE, and then calculates and analyzes the similarity between the model parameters submitted by each UE and the current global model to identify UEs that may have submitted malicious updates.

[0031] Step 8: Model aggregation;

[0032] A new round of global model is formed by performing weighted aggregation operations on the model parameters submitted by the UEs identified as benign in the current round. During this aggregation process, the contribution of each benign UE is weighted according to the ratio of the similarity between its model parameters and the global model. After the global model is formed, the PS selects the UEs that will receive the model update with reference to the user selection probability table, that is, selects all UEs with a selection probability greater than zero. Subsequently, the PS encrypts the global model using the public keys of each UE to ensure that only the target UE can decrypt and use these updates. Finally, the encrypted global model is distributed to each UE by the CS.

[0033] Furthermore, the data cleaning method in step three specifically includes:

[0034] The server will train an AC-GAN on a high-quality and small-scale dataset. After the network is trained, it will be deployed to all UEs to assist in data cleaning;

[0035] After the UE obtains the AC-GAN model, for each local data sample x i , use the discriminator E of AC-GAN to calculate its difference from the true label y i The loss value between is as follows:

[0036]

[0037] Among them, D k represents all data samples of user k, Represents a collection of users;

[0038] The loss values ​​of all samples will be input as features into a Gaussian Mixture Model (GMM); GMM is a probabilistic model that assumes that all data points are a mixture of a finite number of Gaussian distributions. The purpose of this step is to identify the potential distribution of loss values, each distribution represents a group or category in the data, and these categories correspond to different noise levels and data quality; the GMM soft-classifies the data samples based on the loss values ​​and generates soft classification labels As shown below:

[0039]

[0040] Among them, components represents the number of Gaussian distributions, and components = 2 is intended to capture the loss value distribution of noisy samples and non-noise samples. The number corresponds to the value of components;

[0041] Next, calculate the means of the two Gaussian distributions and select the distribution with the largest mean as the label of the noise data to form the noise sample set index noise , and then determine the proportion of the number of noise samples in the total data volume, that is, the noise level level k ; determine the final relabeled data samples through a preset confidence threshold Trust and a cleaning threshold c; finally, use the discriminator E to relabel these data samples, that is, relabel the samples in the noise data set using the prediction results of the discriminator; this step is based on the assumption that the training of the AC-GAN discriminator on a high-quality data set enables it to effectively identify and correct noise labels; the relabeling method is as follows:

[0042]

[0043] Furthermore, the data augmentation method in step four specifically includes:

[0044] The server will train an AC-GAN on a high-quality and small-scale data set. After the network is trained, it will be deployed to all UEs to assist in data augmentation;

[0045] For each First, count the data volume v of each category i in the local data set D k ; according to the preset augmentation threshold r(0 < r ≤ 1), determine the target lower limit low of the data volume of each category

[0046] ; as follows: k As follows:

[0047]

[0048] Specifically, first find the data volume of the category with the largest number in the local data set D k , and then multiply it by the augmentation threshold to serve as the lower limit of the data volume that each category needs to reach;

[0049] After determining the data volume lower limit low k , perform data augmentation on the categories with data volume lower than D k to increase the data volume under the corresponding categories and ensure the balance and diversity of the data set; this process mainly involves the following three steps:

[0050] (1) Generate a random noise vector z: Select an appropriate dimension and randomly draw a vector from a predefined distribution (high-dimensional normal distribution); this noise vector z will be used as the input of the data augmentation process to generate diverse but consistent new data points with the existing data;

[0051] (2) Generate new data: Input the random noise vector z and the specified category label i into the trained ACGAN generator G; the generator G will generate new data samples corresponding to the category based on the input noise vector and category label.

[0052] (3) Add new data to the dataset: Generate new data samples And its corresponding category label i is added to the data set To form a more balanced and diverse dataset

[0053] Furthermore, the CKKS homomorphic encryption scheme of step six specifically includes:

[0054] In the implementation process of federated learning, UE uses global model parameters to perform model training on its own local data set; in each communication round, after UEi completes local training based on the global model parameters, it will form the corresponding local model Subsequently, for secure transmission of model updates, UEi will use the public key pk p The local model is encrypted and scheduled by CS; CS and randomly selected UE subgroup S t Establish communication connections with users in the group and request them to transmit encrypted model updates

[0055] CS calculates the SINR of each user and compares it with the preset signal-to-noise ratio threshold SINR0; UEs with SINR less than the threshold are considered to have their model updates not correctly received by CS and are removed from subgroup S t Then, CS adds noise to the user model; Specifically, CS selects d random integers τ j (j∈[1,d]), add noise to the encryption model of each UE to form a concealed model And send to PS.

[0056] Furthermore, the user scheduling probability dynamic update criteria of step seven specifically include:

[0057] In the initial state, the scheduling probability of all users is expressed as follows:

[0058]

[0059] Where K represents the number of all local users, and the scheduling probability of all users is the same before training begins;

[0060] A dynamic adaptive similarity threshold Sim is introduced; when the cosine similarity between the model update submitted by the UE and the global model exceeds the threshold; specifically, is the similarity between the currently scheduled UE subgroup and the global model calculated by the server in the tth round of communication; the server normalizes the similarities greater than or equal to 0 and takes the normalized mean as the Sim of the current round; the normalization operation is expressed as follows:

[0061]

[0062] Where S' represents the new set obtained by removing all values ​​less than 0 from the set S, min(S') represents the minimum similarity in S', and max(S') represents the maximum similarity in S';

[0063] Based on the similarity threshold Sim, the scheduled users in each round of communication can be divided into two different groups:

[0064]

[0065] in, represents the benign UE set in the tth round of communication, which includes UEs whose local model has a similarity with the global model greater than or equal to Sim; represents the set of malicious UEs in the tth round of communication, which includes UEs whose local model and global model similarity are both less than Sim;

[0066] The adjustment of user selection probability is only performed within the subgroup S scheduled in the current communication round, while the selection probability of unscheduled UEs remains unchanged; in the tth round of communication, set is the size of the benign user set, is the size of the malicious user set, and N′+M′=K1, then the updated user selection probability can be expressed as:

[0067]

[0068] Among them, P t represents the selection probability of all users in the tth round of communication, Defines a benign user set Changes in user selection probability; User i in is the selection probability of user i in the t-1th round of communication, Indicates the increase in the probability of user selection, and x is an adjustable parameter to control the range of probability change; Malicious user set The change in the user's choice probability is User j in is the selection probability of user j in the t-1th round of communication, Indicates the reduction in the user's selection probability; The unscheduled user k is specified, and its selection probability remains unchanged.

[0069] Another object of the present invention is to provide a reliable federated learning system based on AC-GAN and dynamic probability scheduling, including:

[0070] The key generation module is used by the key generation center to generate unique key pairs for the parameter server and all federated learning UEs;

[0071] The server initialization module is used to initialize the model parameters related to the federated learning task and train an AC-GAN model related to the task;

[0072] Data cleaning module, used by UE to identify and correct noise in local data sets using AC-GAN model;

[0073] The data enhancement module is used by the UE to perform data augmentation on the local data set using the AC-GAN model;

[0074] UE local training module, used for UE to optimize the initial global model using local data sets;

[0075] A CKKS encryption module, used to encrypt the local model parameters of the UE using the CKKS homomorphic encryption scheme;

[0076] Attack detection module, used to distinguish malicious UEs from benign UEs in each communication round according to certain rules;

[0077] The model aggregation module is used to perform a formaldehyde aggregation operation on the model parameters submitted by the benign UE identified in the current round according to the analysis results of the attack detection module, so as to form a new round of global model.

[0078] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the reliable federated learning method based on AC-GAN and dynamic probability scheduling.

[0079] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute a reliable federated learning method based on AC-GAN and dynamic probability scheduling.

[0080] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0081] First, in wireless reliable federated learning application scenarios where noise labels, non-independent and identically distributed data and poisoning attacks coexist, it is crucial to improve the quality of user local data sets, reduce the impact of non-independent and identically distributed data, and resist poisoning attacks to protect the privacy of the federated training network, and improve the availability of low-quality data to expand the scale of the federated training network data set. In order to correctly achieve data quality improvement, alleviate the impact of non-independent and identically distributed data and resist poisoning attacks in such scenarios, the present invention discloses a reliable federated learning method based on AC-GAN and dynamic probability scheduling, which trains an AC-GAN model on a small benchmark data set through a server and deploys it on the UE side to achieve data cleaning and data enhancement operations; encrypts user-side model parameter information to prevent user information leakage, and measures the cosine similarity between the user-side model and the global model and implements an adaptive scheduling strategy to construct a scheduling probability table for each user in the federated training network, thereby achieving the distinction between malicious users and benign users.

[0082] By encrypting the user-side model in wireless reliable federated learning application scenarios with coexistence of noise labels, non-independent and identically distributed data, and poisoning attacks, the present invention can effectively prevent user model information leakage and various poisoning attacks, such as data poisoning attacks, model poisoning attacks, and backdoor attacks, thereby improving the stability and model performance of the federated training network. The present invention adopts a dual-server architecture, which can effectively prevent a single server from decrypting and cracking the user model, thereby inferring user-related information and realizing user privacy protection.

[0083] In summary, the present invention proposes an effective solution for the scenario where label noise, non-independent and identically distributed data and poisoning attacks coexist, which can significantly improve the security and reliability of the federated training network and optimize the system performance, which will have a positive impact on the application fields related to wireless security federated learning.

[0084] The present invention overcomes the influence of label noise, non-independent and identically distributed data and poisoning attacks on model performance, while preventing user-side model leakage, reasonably and effectively scheduling benign users to participate in federated training, avoiding the problem of channel resource waste in wireless networks, greatly improving the efficiency of federated learning, and optimizing the training process of the global model.

[0085] The technical solution to be protected by the present invention provides a solution based on CKKS homomorphic encryption technology for the problem of reliable federated learning in scenarios where label noise, non-independent and identically distributed data, and poisoning attacks coexist. The present invention has the following technical effects and advantages:

[0086] Stronger universality: The present invention considers data preprocessing and malicious user identification in scenarios where label noise, non-independent and identically distributed data, and poisoning attacks coexist. The data cleaning operation can be used to achieve the transformation of low-quality benign users into high-quality benign users. The data enhancement operation can be used to expand the user's local data set and reduce the negative impact of non-independent and identically distributed data. The dynamic probability scheduling strategy can be used to filter out malicious users. This allows the present invention to cope with complex scenarios where label noise, non-independent and identically distributed data, and poisoning attacks coexist, as well as simple scenarios where only malicious users or label noise and non-independent and identically distributed data exist, and has stronger universality.

[0087] Improve network security and reliability: The present invention adopts CKKS homomorphic encryption technology to effectively prevent the leakage of user-side models. At the same time, it adopts a dual-server architecture, which not only avoids the disadvantages of a single server (such as single point failure), but also prevents the server from speculating user-related information, thereby improving the security and reliability of the federated training network.

[0088] Second, as auxiliary evidence of the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0089] (1) The expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0090] The technical solution of the present invention provides a solution based on AC-GAN and dynamic probability scheduling for the problem of reliable federated learning in scenarios where label noise, non-IID data and poisoning attacks coexist. After the technical solution of the present invention is converted into a specific product, the expected benefits and commercial value are mainly reflected in the following aspects:

[0091] Enhanced communication security and reliability: This invention reduces the interference of label noise, non-independent and identically distributed data, and poisoning attacks on the overall federated learning process through data preprocessing methods (data cleaning and data enhancement) and dynamic probability scheduling strategies, thereby improving the accuracy and reliability of the model.

[0092] Promoting the development of related industries: The technical solution of the present invention is innovative in the field of federated learning and network security, and can be widely used in a variety of federated learning scenarios, such as medical and health care systems, network security systems, transportation systems, etc., which will help promote the development of related industries, attract investment and cooperation opportunities, and create more commercial value for enterprises.

[0093] (2) The technical solution of the present invention fills the technical gap in the industry at home and abroad:

[0094] The technical solution of the present invention provides a solution based on AC-GAN and dynamic probability scheduling for the problem of reliable federated learning in scenarios where label noise, non-IID data and poisoning attacks coexist, effectively filling the technical gap in the industry at home and abroad, which is specifically reflected in the following aspects:

[0095] Reliable federated learning in scenarios where label noise, non-IID data, and poisoning attacks coexist: In wireless networks, the presence of label noise, non-IID data, and poisoning attacks poses challenges to the performance and security of federated learning. By adopting a data preprocessing method based on AC-GAN and combining it with a dynamic probabilistic scheduling strategy, this invention fills the technical gap in this specific scenario.

[0096] User-side model similarity as a measure of positive and negative behaviors: Compared with traditional measurement indicators such as accuracy, recall rate and F1 score, the present invention uses the cosine similarity between the user-side model and the global model as a measure of user positive and negative behaviors. For the situation where the accuracy of the global model is low in the early stage of training, this method can more intuitively, reasonably and effectively judge the quality of the user-side model.

[0097] (3) Whether the technical solution of the present invention solves the technical problems that people have been eager to solve but have not been able to solve successfully:

[0098] The technical solution of the present invention effectively overcomes the long-standing technical problem of how to correctly improve the quality of local data and identify malicious users in scenarios where label noise, non-independent and identically distributed data, and poisoning attacks coexist. This innovative solution will strongly promote the progress of federated learning reliability issues, thereby improving the performance and security of wireless federated learning networks. This is specifically reflected in the following aspects:

[0099] Wireless federated learning in scenarios where label noise, non-independent and identically distributed data, and poisoning attacks coexist: In reliable federated learning scenarios where label noise, non-independent and identically distributed data, and poisoning attacks coexist, how to correctly screen out truly malicious users while ensuring that more benign users participate in federated training has always been a technical challenge. This invention proposes a dynamic probabilistic scheduling strategy based on the characteristics of scenarios where label noise, non-independent and identically distributed data, and poisoning attacks coexist, and adopts the CKKS homomorphic encryption scheme to effectively solve this problem.

[0100] Data preprocessing method based on AC-GAN: The existence of label noise and non-independent and identically distributed data will lead to low accuracy of the model trained by the user device on the local data set. When there are too many such user devices, the performance of the global model will be reduced or even diverged. Therefore, in order to solve this problem, the present invention proposes a data preprocessing method based on AC-GAN. The user device uses the trained AC-GAN model to perform data cleaning and data enhancement on the local data set, which can effectively improve the data quality and alleviate the impact of non-independent and identically distributed data.

[0101] Dynamic probability scheduling strategy based on user selection probability: In the existing malicious user identification schemes that use cosine similarity as a measure of user positive and negative behaviors, most adopt the strategy of "one-time identification, direct characterization", which may cause some benign users to be mistakenly classified as malicious users, and some malicious users to be mistakenly classified as benign users. The present invention makes an attempt in this regard and proposes a dynamic probability scheduling strategy based on user selection probability, which effectively solves the problem of misidentification of malicious users.

[0102] (4) The technical solution of the present invention overcomes technical prejudice:

[0103] The technical solution of the present invention successfully overcomes multiple technical biases and provides an innovative solution for reliable federated learning in scenarios where label noise, non-IID data, and poisoning attacks coexist. This will help promote the development of reliable federated learning and enhance the security and reliability of wireless federated learning. This is specifically reflected in the following aspects:

[0104] Breaking through the limitations of traditional user positive and negative behavior measurement indicators: Traditional user positive and negative behavior measurement indicators (such as accuracy and F1 score) cannot reflect the continuity of user behavior in scenarios where label noise, non-independent and identically distributed data, and poisoning attacks coexist. The present invention overcomes this technical bias and uses the cosine similarity between the user-side model and the previous round of global model as a measurement indicator, providing a more effective solution for user behavior analysis in scenarios where label noise, non-independent and identically distributed data, and poisoning attacks coexist.

[0105] Data preprocessing method for label noise and non-independent and identically distributed data: In the prior art, there is a lack of unified processing for label noise and non-independent and identically distributed data. The present invention overcomes this technical bias and uses the AC-GAN model to perform data cleaning and data enhancement on the local data set of the user device, thereby achieving quality improvement and data enhancement of the user data set.

[0106] Adaptive scheduling strategy for malicious users: In the prior art, there is a lack of adaptive scheduling strategies for scenarios where label noise, non-independent and identically distributed data, and poisoning attacks coexist. The present invention overcomes this technical bias and dynamically adjusts the scheduling probability value of each user according to the cosine similarity of each user-side model in the network, thereby realizing malicious user identification in scenarios where label noise, non-independent and identically distributed data, and poisoning attacks coexist.

[0107] Third, the technical problems solved by the present invention are:

[0108] 1) Data quality issues: In federated learning, due to the diverse data sources and collection methods of each user device, the data quality is uneven, and there are problems such as noise and label errors. These problems directly affect the accuracy and generalization ability of model training.

[0109] 2) Data imbalance problem: The datasets of user devices often have the problem of category imbalance, that is, the number of samples in some categories is much larger than that in other categories, which will cause the model to favor the category with a large number of samples, affecting the overall performance.

[0110] 3) Security issues: In federated learning, there is a risk of data leakage and malicious attacks during data transmission and aggregation. In particular, malicious users may destroy the global model by submitting erroneous or harmful model updates.

[0111] 4) Communication efficiency and scheduling issues: Traditional federated learning methods often adopt random or fixed scheduling strategies when scheduling users to participate in model training. They are unable to make dynamic adjustments based on the actual contribution and credibility of user data, resulting in low communication efficiency and limited global model performance.

[0112] ####Significant technological advancements:

[0113] 1) Improve data quality: By introducing the AC-GAN model for data cleaning and enhancement, the present invention can automatically identify and correct noise and erroneous labels in the data, and perform targeted data enhancement according to the diversity of the data set, significantly improving data quality, thereby improving the accuracy and robustness of model training.

[0114] 2) Solve the problem of data imbalance: The data enhancement module can generate new data samples in a targeted manner according to the category distribution of the user's local data set, balance the number of samples in each category, thereby solving the problem of data imbalance and improving the generalization ability of the model in different categories.

[0115] 3) Enhanced security: The CKKS homomorphic encryption scheme is used to encrypt and transmit model parameters to ensure the security of data transmission. At the same time, the attack detection module identifies and excludes model updates by malicious users, effectively resists malicious attacks, and ensures the stability and reliability of the global model.

[0116] 4) Improve communication efficiency and scheduling effect: Through the dynamic probability scheduling strategy, the user scheduling probability is dynamically adjusted according to the actual contribution and credibility of the user data, and users with high contribution and good credibility are prioritized to participate in model training, which significantly improves communication efficiency and the convergence speed of the global model. In addition, this strategy can also automatically identify and exclude malicious users to further ensure the performance of the global model.

[0117] In summary, the present invention has achieved significant technological progress in many aspects by introducing the AC-GAN model for data cleaning and enhancement, adopting the CKKS homomorphic encryption scheme to ensure data security, and combining the dynamic probabilistic scheduling strategy to improve communication efficiency and scheduling effect, and has effectively solved the problems of data quality, data security, communication efficiency and scheduling strategy in federated learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0118] Figure 1 It is a flow chart of a reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by an embodiment of the present invention.

[0119] Figure 2 It is a scenario diagram of a reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by an embodiment of the present invention;

[0120] Figure 3 It is a structural block diagram of a reliable federated learning system based on AC-GAN and dynamic probability scheduling provided by an embodiment of the present invention.

[0121] Figure 4 It is a system flow chart of a reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by an embodiment of the present invention.

[0122] Figure 5 This is a curve chart showing how the accuracy of the global model of the reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by an embodiment of the present invention changes with the number of communication rounds under the MNIST data set.

[0123] Figure 6 This is a curve chart showing how the accuracy of the global model of the reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by an embodiment of the present invention changes with the number of communication rounds under the Fashion-MNIST dataset.

[0124] Figure 7 This is a curve chart showing how the accuracy of the global model of the reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by an embodiment of the present invention changes with the number of communication rounds under the CIFAR10 dataset.

[0125] Figure 8This is a curve diagram of selection probability changes of different users in a reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0126] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0127] like Figure 1 As shown, a reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by an embodiment of the present invention includes the following steps:

[0128] S101, the server trains the AC-GAN model on a small benchmark dataset and deploys it to each user device;

[0129] S102, the user cleans the local data set using the AC-GAN model to form a clean data set;

[0130] S103, the user performs data enhancement on the local data set by using the AC-GAN model to form an enhanced data set;

[0131] S104, the coordination server builds a user selection probability table, the user uses local data for training to form a local model, and sends the encrypted local model parameters to the coordination server;

[0132] S105, the coordination server performs model noise addition on the encrypted model and sends it to the parameter server, the parameter server performs malicious attack detection and model aggregation to form a noisy global model, re-encrypts the global model and sends it to the coordination server;

[0133] S106, coordinating the server to perform model denoising and model distribution;

[0134] S107, after a certain number of communication rounds, the coordination server constructs a malicious user set and a benign user set according to the user selection probability table.

[0135] like Figure 2 As shown, the scenario diagram of the reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by the embodiment of the present invention includes the following subjects:

[0136] User Equipment (UE): UE is responsible for receiving global model parameters from CS, training local models using local data sets, encrypting model parameters after training, and sending them to CS via uplink;

[0137] Coordination Server (CS): CS is responsible for receiving the encryption model parameters uploaded by UE, masking the encryption model parameters, and then sending them to PS. It is also responsible for the distribution process of the global model.

[0138] Parameter Server (PS): The main task of PS is to complete model parameter aggregation;

[0139] Key Generation Center (KGC): KGC is responsible for generating key pairs for PS and all UEs.

[0140] like Figure 3 As shown, the reliable federated learning method system based on AC-GAN and dynamic probability scheduling provided by the present invention includes:

[0141] Key generation module,The key generation center generates unique key pairs for the parameter server and all federated learning UEs.

[0142] The server initialization module initializes the model parameters related to the federated learning task and trains an AC-GAN model related to the task.

[0143] In the data cleaning module, the UE uses the AC-GAN model to identify and correct noise in the local data set.

[0144] In the data enhancement module, the UE uses the AC-GAN model to perform data augmentation on the local dataset.

[0145] UE local training module: UE uses local data sets to optimize the initial global model.

[0146] The CKKS encryption module uses the CKKS homomorphic encryption scheme to encrypt the UE's local model parameters.

[0147] The attack detection module distinguishes malicious UEs from benign UEs in each communication round according to certain rules.

[0148] The model aggregation module performs a formaldehyde aggregation operation on the model parameters submitted by the benign UE identified in the current round according to the analysis results of the attack detection module to form a new round of global model.

[0149] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0150] In the actual wireless communication network scenario with label noise, non-independent and identically distributed data and poisoning attacks, the present invention proposes a reliable federated learning method based on AC-GAN and dynamic probability scheduling: the server trains an AC-GAN model on a small benchmark data set and deploys it on the user side. The user uses the AC-GAN model to implement data cleaning and data enhancement, thereby achieving noise label correction while alleviating the impact of non-independent and identically distributed data; the parameter server constructs a user selection probability model and filters malicious users based on a dynamic probability scheduling strategy. In order to prevent user model information leakage and semi-honest servers from inferring user information, the model parameters are encrypted using the CKKS homomorphic encryption scheme, and a dual-server architecture is adopted. The present invention overcomes the impact of label noise, non-independent and identically distributed data and poisoning attacks on model performance.

[0151] like Figure 4 As shown, the reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by the embodiment of the present invention specifically includes the following steps:

[0152] In the first step, the coordination server initializes the parameters related to the federated learning task, trains the task-related AC-GAN model on a small benchmark dataset, and distributes it to all user devices;

[0153] In the second step, each user device uses the AC-GAN model to perform data preprocessing and local model training;

[0154] In the third step, each user encrypts the local model parameters and sends them to the coordination server;

[0155] Step 4: The coordination server adds noise to the received user model and sends it to the parameter server;

[0156] In the fifth step, the parameter server performs malicious attack detection and aggregates the models of benign users;

[0157] Step 6: The parameter server re-encrypts the global model and sends it to the coordination server;

[0158] In the seventh step, the coordination server performs denoising on the global model and distributes it to each user device.

[0159] Furthermore, the reliable federated learning method based on AC-GAN and dynamic probability scheduling specifically includes:

[0160] Step 1: Key generation. The Key Generation Center (KGC) first generates a unique key pair for the parameter server (PS) and all federated learning UEs. After generating the key pair, KGC distributes the private key to each owner through a secure communication channel, and the public key is made public. After the key pair is generated and distributed, all parties in the system can start encrypted communication and data exchange. KGC is also responsible for the maintenance and update of the key. During the key usage cycle, if a key leak or other security risks are found, KGC needs to generate a new key pair in time, and redistribute and make the process public.

[0161] Step 2: Initialize the coordination server (CS). CS first initializes the model parameters related to the federated learning task and trains an AC-GAN model related to the task. CS is responsible for maintaining a high-quality small benchmark dataset, which is used to train the AC-GAN model. During this training process, AC-GAN learns how to generate new and similar data samples based on the benchmark dataset and can control the categories of generated samples. The trained AC-GAN model can provide data enhancement functions and assist in data cleaning. After training, the AC-GAN model and the model parameters related to the federated learning task are distributed to all UEs.

[0162] Step 3: Data Cleaning (DC). The main task of DC is to enhance the learning efficiency and accuracy of the model by improving data quality. This step includes: noise identification, quality assessment, noise correction and post-cleaning verification. Specifically, relying on the guidance of the AC-GAN model, the UE first calculates the loss value of the local data sample on the AC-GAN discriminator, converts the loss value matrix into a probability distribution, and inputs it into a Gaussian mixture model to distinguish potential noise samples from clean samples; then, the UE estimates the proportion of noise samples, and then quantitatively evaluates the quality of the local data set; the UE determines the threshold of data cleaning based on its own computing resources and the advice of the CS, and uses the AC-GAN discriminator to perform the noise correction step for noise samples with high confidence; then, the UE can choose to re-evaluate the quality to ensure the effectiveness of the cleaning process. The cleaned data set will enter the data enhancement stage, that is, the data set output by DC is used as the input data set of the data enhancement stage.

[0163] Step 4: Data Augment (DA). DA is committed to improving the generalization ability of the model and improving the adaptability and accuracy of the model in a diverse environment by expanding the data set. This step covers: diversity assessment, sample generation, sample fusion, and post-augmentation verification. The specific process is as follows: the UE first counts the distribution of each category and the number of samples in the local data set to evaluate the diversity of the data set; the UE determines the enhancement threshold based on its own computing power and storage space, and uses the generator of the AC-GAN model to generate new data samples for each category in the data set. The purpose of this step is to balance the category distribution and improve the diversity of the data set; then, the newly generated data samples are sampled and fused with the original data samples to form a data set after DA; finally, the UE can evaluate the diversity of the data set again to ensure the effectiveness of DA. The data set output by DA will be used as the input data set for the training of the federated learning model, thereby providing a richer and more extensive feature representation to support the learning and generalization of the model in a more complex data environment.

[0164] Step 5, UE local training. In the federated learning process, UE local training is a crucial link. At this stage, each UE optimizes the initial global model using its own data set. Specifically, the UE applies various machine learning algorithms, such as stochastic gradient descent (SGD) or its variants, to adjust and optimize the parameters of the local model. This process is performed completely independently on the UE to ensure the locality of the data and avoid the leakage of sensitive data. During the optimization process, the UE adjusts the model parameters according to its own data characteristics to better adapt to the local data distribution. This step is the key to achieving personalized model optimization because it allows the model to adjust itself to match the specific needs of the user without accessing other users' data.

[0165] Step six, CKKS encryption. In order to ensure the security and privacy of the model parameters generated by the UE during the communication process, the CKKS homomorphic encryption scheme is used to encrypt the local model parameters of the UE. The encryption operation is performed using the public key provided by the PS, ensuring that only the PS has the ability to decrypt and access the encrypted data. In order to further enhance the confidentiality of the data, when the UE sends the encrypted model to the CS, the CS performs noise processing on the received encrypted data. This measure is intended to prevent potential man-in-the-middle attacks and ensure that even when the data is decrypted on the PS side, the plaintext information of the original model cannot be directly obtained. In addition, before the global model is sent by the CS to each UE, the encrypted global model is denoised to ensure that the UE can obtain an accurate global model and ensure the correctness of the model update.

[0166] Step seven, attack detection. The attack detection step is performed on the PS side, and the core goal is to protect the global model from malicious UEs by identifying and screening out potential malicious updates. Specifically, the PS and CS jointly maintain a user selection probability table to characterize the credibility of the federated learning UE. During each round of model update, the PS first decrypts the model parameters of each UE, and then identifies the UE that may have submitted malicious updates by calculating and analyzing the similarity between the model parameters submitted by each UE and the current global model. The similarities of all UEs in the current round are normalized and their mean is calculated as the similarity threshold. The malicious UEs and benign UEs in the current communication round are distinguished based on this threshold, and then the user selection probability table is adjusted according to the corresponding update mechanism. This dynamic adjustment mechanism not only increases the probability of benign UEs being selected in subsequent model updates, but also reduces the impact of malicious UEs.

[0167] Step eight, model aggregation. The aggregation step is performed on the PS side, aiming to integrate the analysis results of the attack detection step, and perform weighted aggregation operations on the model parameters submitted by the benign UEs identified in the current round, so as to form a new round of global models. In this aggregation process, the contribution of each benign UE is weighted according to the ratio of the similarity between its model parameters and the global model. After the global model is formed, the PS refers to the user selection probability table to select the UE that will receive the model update, that is, select all UEs with a selection probability greater than zero. Subsequently, the PS uses the public key of each UE to encrypt the global model to ensure that only the target UE can decrypt and use these updates. Finally, the encrypted global model is distributed to each UE by the CS.

[0168] In order to prove the creativity and technical value of the technical solution of the present invention, this section provides application examples of the claimed technical solution on specific products or related technologies.

[0169] Example 1: Medical and health care system

[0170] In the medical and health care system, privacy protection is crucial for the high sensitivity of medical data and system security. The technical solution of the present invention can filter the data of malicious users to ensure that malicious data will not appear in the medical and health care system, thereby improving the security of the system; using the data cleaning method to filter out low-quality data from the data set is helpful to improve the data quality of the medical and health care system; the medical data set can be expanded through the data enhancement method, which helps the medical and health care system to learn more medical data.

[0171] Example 2: Transportation System

[0172] In the transportation system, identifying and filtering malicious information and noise in traffic data plays a vital role in understanding traffic conditions and behaviors. Through the technical solution of the present invention, the performance and accuracy of the transportation federated learning system can be improved by cleaning low-quality data, thereby improving the universality of the transportation federated learning system by expanding a small number of category data, and providing better preparation for future intelligent transportation.

[0173] Example 3: Network Security System

[0174] In network security systems, identifying malicious users' behavior patterns and filtering low-quality data are crucial to maintaining network systems and protecting confidential data. The technical solution of the present invention can not only prevent privacy leakage, but also make the network security of all federation participants stronger and maintain the healthy development of the Internet ecosystem.

[0175] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. A person of ordinary skill in the art will appreciate that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0176] 1. Simulation conditions:

[0177] The simulation experiment of the present invention is carried out on the Windows platform, and the main configuration is: CPU is Intel (R) Core (TM) i9-10900K, 3.70GHz; memory is 8G; operating system is Windows 10; simulation software environment is Python 3.8 + PyTorch.

[0178] by Figure 4 The scenario diagram of the wireless federated learning system under the dual-server architecture is a schematic diagram of the simulation scenario used in the simulation experiment of the present invention. The diagram includes a wireless federated learning system composed of ordinary user devices, malicious user devices, low-quality benign user devices, servers, key generation centers and cloud platforms.

[0179] 2. Simulation content and result analysis:

[0180] This simulation experiment compares the method of the present invention with the existing federated average aggregation strategy and the identification mechanism based on the coordinate level median. Figure 4 In the simulation scenario, according to different simulation parameters, the secure wireless federated learning method in the malicious and low-quality user scenarios of the present invention is used for simulation. The simulation experiment is carried out within a circular range with a radius of 300m, and a federated learning process in a typical cell is selected as the research object. Unless otherwise specified, the simulation process parameters are as follows: centered on the location of the server, the number of users K=100 evenly distributed in the cell, the number of orthogonal sub-channels between the terminal user and the server antenna N=10, the path loss index γ=3.8, the terminal user learning rate η=0.01, the SINR decoding threshold is 5dB, and the proportion of malicious users in the cell is ρ m =30%, the ratio of noisy users in the cell to the total number of users ρ = 0.5, the lower limit of the number of noise samples in the local data set of noisy users to the total number of samples θ = 0.5, the binomial distribution probability coefficient p = 0.5, and the Dirichlet distribution concentration coefficient α = 0.1. Among them, p and α are used to construct non-independent and identically distributed data sets. In addition, in the simulation, the simulation data sets used are MNIST data set, Fashion-MNIST data set and CIFAR10 data set. In order to highlight the superiority of the proposed method, this study considers the following four scenarios:

[0181] (1) Scenario 1: UE does not perform data preprocessing and PS does not run the attack detection module

[0182] (2) Scenario 2: UE performs data preprocessing and PS does not run the attack detection module

[0183] (3) Scenario 3: UE does not perform data preprocessing, PS runs attack detection module

[0184] (4) Scenario 4: Some UEs do not perform data preprocessing, some UEs perform data preprocessing, and the PS runs the attack detection module

[0185] (5) Scenario 5: UE performs data preprocessing and PS runs attack detection module

[0186] For scenario 4, the proportion of UEs that perform data preprocessing to the total noise UEs is set to π = 50%. Through numerical analysis and simulation, we can get the following Figures 5 to 8 result.

[0187] Figure 5This is a curve chart showing the accuracy of the global model of the reliable federated learning method based on AC-GAN and dynamic probability scheduling under the MNIST data set as a function of the number of communication rounds provided by an embodiment of the present invention. As can be seen from the figure, the model accuracy of scenario 5 shows a continuous upward trend and is maintained at a relatively high level. On the contrary, the model accuracy of scenario 2 fluctuates greatly between 50% and 95%, showing a characteristic of being unable to converge. This is because the PS does not perform malicious attack detection, which causes malicious UEs to interfere with the convergence of the global model. The model accuracy of scenarios 1 and 3 is relatively low, maintaining at only about 20%. This shows that the UE model trained on a data set that has not been preprocessed performs poorly, emphasizing the key role of data preprocessing in improving model performance. In scenario 4, since 50% of the UEs correctly performed the data preprocessing steps, the accuracy of the model was able to converge stably and remain at a high level.

[0188] Figure 6 This is a graph showing the accuracy of the global model of the reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by the embodiment of the present invention in the Fashion-MNIST dataset as a function of the number of communication rounds. The law presented by this method in this dataset is similar to that in the MNIST dataset, and will not be repeated here.

[0189] Figure 7 This is a graph showing the accuracy of the global model of the reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by the embodiment of the present invention in the CIFAR10 dataset as a function of the number of communication rounds. The law presented by this method in this dataset is similar to that in the MNIST dataset, and will not be repeated here.

[0190] Figure 8 is a curve diagram of the selection probability changes of different users in the reliable federated learning method based on AC-GAN and dynamic probability scheduling provided by an embodiment of the present invention. In the study of scenario 4, Figure 8The dynamic changes of user selection probabilities of noisy UEs that perform data preprocessing normally (hereinafter referred to as preprocessed UEs), noisy UEs that do not perform data preprocessing (hereinafter referred to as non-preprocessed UEs), and malicious UEs in the federated learning process are shown. It can be observed from the figure that the selection probability of preprocessed UEs gradually increases with the increase of communication rounds and gradually stabilizes at a higher level (about 4%). This change reflects that the method proposed in this study tends to select UEs that provide high-quality data. Specifically, the data of preprocessed UEs has been cleaned and enhanced, so that it surpasses other UEs in quality and shows a higher similarity with the global model parameters, so it is more likely to be selected to promote the optimization of the global model. This also reveals the positive role of data preprocessing in improving the quality of UE local models, and the importance of consistency between high-quality models and global models in increasing the probability of UE selection. For malicious UEs, their selection probability shows a downward trend throughout the communication process and eventually drops to zero. This phenomenon shows that the method proposed in this study can identify and exclude malicious UEs that carry out malicious attacks to a certain extent. For non-preprocessed UEs, their selection probability is slightly higher than that of malicious UEs, but the overall trend is consistent with that of malicious UEs. This trend reflects the data sensitivity of the proposed method to noisy labels and non-IID features, indicating that the similarity between these UE local model parameters and the global model parameters is low, reflecting the direct consequence of poor data quality. This difference causes the proposed method to reduce their selection probability.

[0191] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A reliable federated learning method based on AC-GAN and dynamic probability scheduling, characterized in that: The following steps are involved: Step 1: The server trains the AC-GAN model on a small benchmark dataset and deploys it on each user device; Step 2: The user cleans the local data set using the AC-GAN model to form a clean data set; Step 3: The user performs data enhancement on the local data set using the AC-GAN model to form an enhanced data set; Step 4: The coordination server constructs a user selection probability table. The user uses local data for training to form a local model and sends the encrypted local model parameters to the coordination server. Step 5: The coordination server performs model noise addition on the encrypted model and sends it to the parameter server. The parameter server performs malicious attack detection and model aggregation to form a noisy global model, and re-encrypts the global model and sends it to the coordination server. Step 6: Coordinate the server to perform model denoising and model distribution; Step 7: After a certain number of communication rounds, the coordination server constructs a malicious user set and a benign user set based on the user selection probability table.

2. The reliable federated learning method based on AC-GAN and dynamic probability scheduling as claimed in claim 1, characterized in that: The reliable federated learning method based on AC-GAN and dynamic probability scheduling specifically includes: Step 1: Key generation. The key generation center KGC first generates a unique key pair for the parameter server PS and all federated learning UEs. After generating the key pair, KGC distributes the private key to each owner through a secure communication channel, and the public key is made public. After the key pair is generated and distributed, all parties in the system can start encrypted communication and data exchange. KGC is also responsible for key maintenance and update. During the key usage cycle, if key leakage or other security risks are found, KGC needs to generate a new key pair in time, and redistribute and make the process public. Step 2: Initialize the coordination server CS. CS first initializes the model parameters related to the federated learning task and trains an AC-GAN model related to the task. CS is responsible for maintaining a high-quality small benchmark dataset, which is used to train the AC-GAN model. Step 3: Data cleaning DC; Includes: noise identification, quality assessment, noise correction and post-cleaning verification; Step 4: Data enhancement (DA); The specific process is as follows: the UE first counts the distribution of each category and the number of samples in the local data set to evaluate the diversity of the data set; the UE determines the enhancement threshold based on its own computing power and storage space, and uses the generator of the AC-GAN model to generate new data samples for each category in the data set; the purpose of this step is to balance the category distribution and improve the diversity of the data set; then, the newly generated data samples are sampled and fused with the original data samples to form a data set after DA; finally, the UE evaluates the diversity of the data set again; Step 5: UE local training; In this stage, each UE optimizes the initial global model using its own dataset; Step 6: CKKS encryption; When the UE sends the encryption model to the CS, the CS performs noise processing on the received encrypted data; Step 7: Attack detection; PS and CS jointly maintain a user selection probability table to characterize the credibility of federated learning UEs. In each round of model update, PS first decrypts the model parameters of each UE, and then calculates and analyzes the similarity between the model parameters submitted by each UE and the current global model to identify UEs that may have submitted malicious updates. Step 8: Model aggregation; A new round of global model is formed by performing weighted aggregation operations on the model parameters submitted by the UEs identified as benign in the current round. During this aggregation process, the contribution of each benign UE is weighted according to the ratio of the similarity between its model parameters and the global model. After the global model is formed, the PS selects the UEs that will receive the model update with reference to the user selection probability table, that is, selects all UEs with a selection probability greater than zero. Subsequently, the PS encrypts the global model using the public keys of each UE to ensure that only the target UE can decrypt and use these updates. Finally, the encrypted global model is distributed to each UE by the CS.

3. The reliable federated learning method based on AC-GAN and dynamic probability scheduling as claimed in claim 2, characterized in that: The data cleaning method in step 3 specifically includes: The server will train an AC-GAN on a high-quality and small-scale dataset. After the network is trained, it will be deployed to all UEs to assist in data cleaning; After the UE obtains the AC-GAN model, for each local data sample x i , use the discriminator E of AC-GAN to calculate the loss value between it and the true label yi, as shown below: Among them, D k represents all data samples of user k, Represents a collection of users; The loss values ​​of all samples will be input as features into a Gaussian Mixture Model (GMM); GMM is a probabilistic model that assumes that all data points are a mixture of a finite number of Gaussian distributions. The purpose of this step is to identify the potential distribution of loss values, each distribution represents a group or category in the data, and these categories correspond to different noise levels and data quality; the GMM soft-classifies the data samples based on the loss values ​​and generates soft classification labels As shown below: Among them, components represents the number of Gaussian distributions, and components = 2 is intended to capture the loss value distribution of noisy samples and non-noise samples. The number corresponds to the value of components; Next, calculate the mean of the two Gaussian distributions and select the distribution with the largest mean as the label of the noise data to form the noise sample set index noise , and then determine the ratio of the number of noise samples to the total data volume, that is, the noise level level k ; The final relabeled data samples are determined by the preset confidence threshold Trust and cleaning threshold c; Finally, the discriminator E is used to relabel these data samples, that is, the samples in the noise data set are relabeled using the discriminator's prediction results; This step is based on the assumption that the training of the AC-GAN discriminator on a high-quality data set enables it to effectively identify and correct noise labels; The relabeling method is as follows: Furthermore, the data enhancement method in step 4 specifically includes: The server will train an AC-GAN on a high-quality and small-scale dataset. After the network is trained, it will be deployed to all UEs to assist in data augmentation. For each First, count the local data set D k The amount of data v for each category i in According to the preset enhancement threshold r (0 < r ≤ 1), determine the target lower limit value low of the data volume of each category k ; as follows: Specifically, first in the local dataset D k Find the amount of data that the category with the largest number has, and then multiply it by the enhancement threshold as the lower limit of the amount of data that each category needs to reach; Determine the lower limit of data volume k Afterwards, for data volumes less than D k Data enhancement is performed on the categories to increase the amount of data in the corresponding categories and ensure the balance and diversity of the data set; this process mainly involves the following three steps: (1) Generate a random noise vector z: Select an appropriate dimension and randomly extract a vector from a predefined distribution (high-dimensional normal distribution); this noise vector z will be used as input to the data augmentation process to generate new data points that are diverse but consistent with the existing data; (2) Generate new data: Input the random noise vector z and the specified category label i into the trained ACGAN generator G; the generator G will generate new data samples corresponding to the category based on the input noise vector and category label. (3) Add new data to the dataset: Generate new data samples And its corresponding category label i is added to the data set To form a more balanced and diverse dataset 4. The reliable federated learning method based on AC-GAN and dynamic probability scheduling as claimed in claim 2, characterized in that: The CKKS homomorphic encryption scheme of step 6 specifically includes: In the implementation process of federated learning, UEs use global model parameters to perform model training on their respective local data sets; in each communication round, after UE i completes local training based on the global model parameters, it will form the corresponding local model Subsequently, for secure transmission of model updates, UE i will use the public key p of the PS k p encrypts the local model and is scheduled by CS; CS and the randomly selected UE subgroup S t Establish communication connections with users in the group and request them to transmit encrypted model updates CS calculates the SINR of each user and compares it with the preset signal-to-noise ratio threshold SINR0; UEs with SINR less than the threshold are considered to have their model updates not correctly received by CS and are removed from subgroup S t Then, CS adds noise to the user model; Specifically, CS selects d random integers τ j (j∈[1, d]), add noise to the encryption model of each UE to form a concealed model And send to PS.

5. The reliable federated learning method based on AC-GAN and dynamic probability scheduling as claimed in claim 2, characterized in that: The user scheduling probability dynamic update criteria of step 7 specifically include: In the initial state, the scheduling probability of all users is expressed as follows: Where K represents the number of all local users, and the scheduling probability of all users is the same before training begins; A dynamic adaptive similarity threshold Sim is introduced; when the cosine similarity between the model update submitted by the UE and the global model exceeds the threshold; specifically, is the similarity between the currently scheduled UE subgroup and the global model calculated by the server in the tth round of communication; the server normalizes the similarities greater than or equal to 0 and takes the normalized mean as the Sim of the current round; the normalization operation is expressed as follows: Where S′ represents the new set obtained by removing all values ​​less than 0 from the set S, min(S′) represents the minimum similarity in S′, and max(S′) represents the maximum similarity in S′; Based on the similarity threshold Sim, the scheduled users in each round of communication can be divided into two different groups: in, represents the benign UE set in the tth round of communication, which includes UEs whose local model has a similarity with the global model greater than or equal to Sim; represents the set of malicious UEs in the tth round of communication, which includes UEs whose local model and global model similarity are both less than Sim; The adjustment of user selection probability is only performed within the subgroup S scheduled in the current communication round, while the selection probability of unscheduled UEs remains unchanged; in the tth round of communication, set is the size of the benign user set, is the size of the malicious user set, and N′+M′=K1, then the updated user selection probability can be expressed as: Among them, P t represents the selection probability of all users in the tth round of communication, Defines a benign user set Changes in user selection probability; User i in is the selection probability of user i in the t-1th round of communication, Indicates the increase in the probability of user selection, and x is an adjustable parameter to control the range of probability change; Malicious user set The change in the user's choice probability is User j in is the selection probability of user j in the t-1th round of communication, Indicates the reduction in the user's selection probability; The unscheduled user k is specified, and its selection probability remains unchanged.

6. A reliable federated learning system based on AC-GAN and dynamic probability scheduling according to any one of claims 1 to 5, characterized in that: The reliable federated learning system based on AC-GAN and dynamic probability scheduling includes: The key generation module is used by the key generation center to generate unique key pairs for the parameter server and all federated learning UEs; The server initialization module is used to initialize the model parameters related to the federated learning task and train an AC-GAN model related to the task; Data cleaning module, used by UE to identify and correct noise in local data sets using AC-GAN model; The data enhancement module is used by the UE to perform data augmentation on the local data set using the AC-GAN model; UE local training module, used for UE to optimize the initial global model using local data sets; A CKKS encryption module, used to encrypt the local model parameters of the UE using the CKKS homomorphic encryption scheme; Attack detection module, used to distinguish malicious UEs from benign UEs in each communication round according to certain rules; The model aggregation module is used to perform a formaldehyde aggregation operation on the model parameters submitted by the benign UE identified in the current round according to the analysis results of the attack detection module, so as to form a new round of global model.

7. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the reliable federated learning method based on AC-GAN and dynamic probability scheduling as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the reliable federated learning method based on AC-GAN and dynamic probability scheduling according to any one of claims 1 to 5.

Images