Adversary-aware authentication method in federated learning based on adaptive steganographic watermarking

Abstract

The application discloses an adversary-aware authentication method in federated learning based on adaptive steganographic watermarking, and the method is as follows: a device constructs its own identity information and encrypts and transmits the identity information to an authentication server; the authentication server processes the request of the external device and ensures that each device joining the system has a controlled validity period and clear identity information; the device re-applies to the authentication server according to the validity period, and the server regularly cleans up invalid internal devices; the authentication server registers the device ID of the legal device and groups the legal device, simultaneously selects a group leader device for each group and formulates a rotation rule; the authentication server generates corresponding watermark trigger data sets for each group and sends the watermark trigger data sets to all legal devices together with a global model and an authentication data set; the legal device trains the global model and learns a trigger strategy; and the authentication server performs security authentication on the legal device. The application improves the controllability, accuracy and concealment of the authentication process, and realizes safe, efficient and concealed authentication.

Description

TECHNICAL FIELD

[0001] The present application relates to network information security, distributed optimization and covert communication, and in particular to an adversary-aware authentication method in federated learning based on adaptive steganographic watermarking. BACKGROUND

[0002] Security authentication is a core technology to protect data integrity and user privacy, and is usually used to verify the identity and authority of users or devices, and is also a major difficulty in the field of network security. Among them, identity authentication is a key link of security authentication. At present, a large number of authentication schemes have been proposed, most of which use blockchain, key, certificate, physical factor and dual-factor authentication to achieve identity verification, but these methods have limitations. The blockchain scheme is easy to be cracked by joint attackers, and the verification data of the key and certificate authentication method is usually stored in plaintext form. Once the data of the physical factor authentication is leaked, all related security authentication will fail. The dual-factor authentication usually relies on explicit identity credentials, and cannot guarantee that the same user is always correctly identified in the two authentication processes. In other words, these two authentication methods do not have concealment.

[0003] Federated learning is a frontier technology in the field of data privacy and distributed optimization, and its core mechanism is to integrate the computing resources and data advantages of distributed nodes through the coordination and management of a central server to jointly train a shared model. This innovative mode not only supports multi-party collaboration modeling, but also effectively protects user data privacy. Therefore, based on its unique characteristics, federated learning is particularly suitable for large-scale security authentication scenarios.

[0004] Digital watermarking is an information security technology mainly used to identify the ownership and distribution of multimedia information on the Internet. This technology needs to meet several key requirements, including concealment, resistance to destruction, embedding capacity and time efficiency. Similarly, it can also be used as an identity verification method. If the watermark identifier can be extracted from the device and correctly verified, it can be determined that the device exists in the authentication system.

[0005] Steganography is a technology for covert communication, and its core is to hide information. In steganography, image steganography refers to embedding secret information in an image without significantly changing its appearance. This technology is often referred to as high-capacity steganography. The simplest implementation is the least significant bit replacement method, which embeds secret information in the least significant bit of each pixel, and this method does not significantly affect the visual effect of the image. In addition, there are methods based on statistical models, which use the statistical properties of images to hide data in areas that are not easily detected, such as Gaussian embedding models, statistical models of discrete cosine transform coefficients, etc. In recent years, generative image steganography has attracted attention due to its security and batch processing capabilities, for example, a generative adversarial network can be used to implement image steganography technology. SUMMARY

[0006] The purpose of the present application is to solve the problem of easy analysis of authentication features in the current network environment, to provide an adversary-aware authentication method in federated learning based on adaptive steganographic watermarking, to provide a more reliable authentication mechanism to ensure the legitimacy of users and devices, and to prevent unauthorized access.

[0007] The technical solution adopted by the present application is:

[0008] The adversary-aware authentication method in federated learning based on adaptive steganographic watermarking includes a training phase and an authentication phase, and the specific steps are as follows:

[0009] Training phase:

[0010] Step 1, the device constructs its own identity information A I and encrypts and transmits to the authentication server A s ; wherein the identity information A I of the legitimate device that has been approved by the system administrator to join the network environment includes a unique authentication identifier A id ;

[0011] Step 2, the authentication server A s handles the request of external devices in the current network environment, ensures that each device joining has a controlled validity period A p and clear identity information A I ;

[0012] Step 3, the authentication server A s registers the device ID for the legitimate device, and groups the legitimate devices at the same time;

[0013] Step 4, the authentication server generates corresponding watermark trigger data sets for each group and allocates authentication data sets required for subsequent authentication, and sends the global model and authentication data sets to all legitimate devices at the same time;

[0014] Step 5, select group leader devices for each group based on the rotation rule;

[0015] Step 6, the legitimate device performs federated training, trains the global model and learns the trigger strategy;

[0016] Authentication phase:

[0017] Step 7, the authentication server performs security authentication on the legitimate device.

[0018] Further, the specific steps of step 1 are:

[0019] Step 1-1, first classify the device, and assign a unique authentication identifier A idIllegal devices that try to access the network without the confirmation of the administrator will not have A id ).

[0020] Step 1-2, when the device joins the network environment, send a join request to the authentication server A s ; the authentication server A s receives the request and requires the device to build and upload complete identity information A I , identity information A I including the manufacturing information of the device, the validity period of the identity information A p , the assigned authentication identifier A id and the associated attributes;

[0021] Step 1-3, identity information A I is encrypted using the public key of the authentication server A s before transmission, and the encryption process is as follows: ; where M represents the encrypted information, represents the encryption using the public key of the authentication server A s .

[0022] Further, the specific method of step 2 is: decrypt and verify whether the identity information A I submitted by the device is valid (the information field is complete and A id valid); if so, set the authentication status of the device to valid in the authentication system, and configure the validity period of the device information A p ; the validity period of the device information A p cannot be indefinite, and a threshold can be set to limit the tasks performed by the network each time A p ; otherwise, the verification fails (i.e. A I missing necessary fields or missing legal A id ), the system will discard or block the corresponding identity information A I , and the corresponding device will be listed in the blacklist to prohibit its access to the network environment.

[0023] Further, in step 2, the device re-applies to the authentication server according to the validity period, and the server regularly cleans up invalid devices inside, and the specific method is as follows:

[0024] When the validity period of the device information A p is about to expire, the authentication server A s warns the corresponding device; when the device wants to continue to exist in the environment, it re-initiates the request before the expiration date;

[0025] When the validity period of the device information A p expires, the authentication server A sSet the authentication status of the corresponding device to invalid, and add it to the list of dormant devices;

[0026] When the device cannot actively exit the network due to deactivation or forgetfulness, etc., resulting in long-term existence and resource occupation, the authentication server A s corresponding to the expiration of the validity period A p periodically after the end of the cleaning.

[0027] Further, the specific steps of step 3 are:

[0028] Step 3-1, the authentication server A s obtains the authentication status of the device;

[0029] Step 3-2, the authentication server A s groups each legitimate device and assigns a unique device ID, and the ID assigned to each device consists of two parts, the first part is the group number G num , and the second part is a variable-length random number;

[0030] Further, in step 3, the authentication server ensures that the number of members between groups is balanced based on the grouping probability formula, and the grouping probability formula is:

[0031] , , where

[0032] wherein, and are positive numbers to avoid division by zero; denotes the group number of the i-th group; denotes the group number of the j-th group; denotes the number of members of the group number , and denotes the number of members of the group number ; denotes the watermark trigger data set, denotes the number of data types, n denotes the number of groups, device denotes the device of the i-th group.

[0033] Specifically, the device grouping is based on two dimensions, the number of categories in the data set and the number of members currently assigned to each group, in order to ensure the balanced distribution of the number of members in each group.

[0034] Further, the specific steps of step 4 are:

[0035] Step 4-1, assign an independent watermark trigger data set to each group for subsequent training,

[0036] The selected data set of each group includes a digital image data set and a non-digital image data set, and the construction process of the watermark trigger data set is to select a digital image data set corresponding to the group number G num The corresponding image sample , and sequentially embed the image samples into the non-digital image samples corresponding to the group classification through image steganography, and the embedding position is non-continuous, and the calculation formula is as follows,

[0037]

[0038]

[0039] wherein, is the watermark embedding position, is the interval between the watermark embedding positions, is the new watermark embedding position for calculation, and U is an integer set from 1 to 5, represents a watermark embedding function, represents the data of the watermark embedding position of the non-watermark trigger data set represents the group number G num The number of samples in the corresponding digital image data set represents the group number G num corresponding to the number of samples in the digital image data set;

[0040] Step 4-2, distribute the global model, the watermark trigger data set and the authentication data set to all legal devices.

[0041] Further, in step 5, the authentication server A s randomly selects a device as the group leader device in each group; when a device has served as a group leader device in a task, the priority of the corresponding device to be selected as the group leader device in the next task execution is reduced; the rotation rule is that the system re-groups all devices at a random time point, and avoids re-allocating the same group of devices to the original group, and after completing the grouping adjustment, the auxiliary device ID, the device identifier and the authentication data set in the group are regenerated and re-allocated.

[0042] Further, the specific steps of step 6 are:

[0043] Step 6-1, after the legal device receives the content sent by the authentication server A s , enters the training stage, divides the watermark trigger data set into two parts locally, including watermark data and non-watermark data, both of which are used for federated learning training;

[0044] ​​Step 6-2, training is divided into two stages, the first stage, the device uses each local watermark trigger data set to train the local model and upload the local model update to the authentication server A s Participate in global model aggregation;

[0045] Step 6-3, the second stage, the device extracts the watermark from the specified location of the watermark data and identifies it; when the watermark is correctly identified, the device modifies the label of the specified location of the watermark data to its own authentication identifier A id .

[0046] Further, the specific steps of step 7 are:

[0047] Step 7-1, the authentication server A s Randomly select an image and embed the corresponding specified index; the specified index is a set of numerical sequences, indicating the authentication server A s Select an image combination that can constitute a special meaning; combine the image with the specified index with the authentication server A s Generate a digital signature combined with the authentication data A d , wherein the digital signature is the authentication server A s Generate a set of ciphertexts that verify the authentication data A d whether it comes from the authentication server A s , and finally the authentication data A d is sent to the group leader device G leader ;

[0048] Step 7-2, the group leader device G leader After receiving the authentication data A d , verify whether the data source of the digital signature is from the authentication server A s ; if so, determine that the verification is passed and execute step 7-3; otherwise, determine that the data source is unreliable, discard the authentication data and no longer continue, and the group leader device G leader will generate a response A r and then upload it to the authentication server A s and end the current authentication;

[0049] Step 7-3, the group leader device G leader forward the authentication data A d to the group members, and the group members extract the embedded index from the image;

[0050] Step 7-4, the group members generate a response A r after uploading the corresponding data generated by the authentication data set to the authentication server A s , the response A r contains the authentication identifier A id, a regular image inference result, a timestamp, and a value obtained by encrypting the timestamp by a device ID.

[0051] The present application adopts the above technical scheme, which can effectively guarantee the security authentication of devices in a network environment. The scheme combines federated learning with watermark triggering mechanism to realize access control, identity authentication and continuous state monitoring of devices in the network, which includes a training phase and an authentication phase. The training phase is subdivided into three sub-phases: device joining request processing, authentication server grouping configuration, and legal device training.

[0052] The present application involves three types of participants: authentication servers, regular devices, and group leader devices. Specifically, devices are divided into groups, with each group designated a group leader device. The authentication server is responsible for identity verification of all devices, while the group leader device is responsible for verifying the reliability of the authentication data source and distributing authentication data to other members within the group. In addition, different access permissions can be assigned to each group according to the grouping strategy, thereby building a complete security authentication mechanism.

[0053] The present application adopts the above technical scheme, which has the following beneficial effects: 1. Innovation of federated learning authentication paradigm. By seamlessly integrating adaptive steganographic watermark technology with federated learning, this framework creates a new paradigm for covert and robust device authentication, effectively addressing the inherent flaws of current network security frameworks, such as insufficient authentication concealment and vulnerability to attacks. 2. Use of dynamic watermark to achieve tamper-proof authentication. This innovative mechanism achieves a seamless integration of explicit interaction and implicit authentication, hiding the traces of the authentication process while ensuring verifiable trust, thereby significantly improving the security posture of the system. 3. Implementation of implicit authentication in explicit federated interaction. This strategy not only eliminates detectable authentication patterns but also maintains system stability, providing users with a seamless and secure authentication experience without sacrificing performance. 4. Use of a strict verification mechanism to achieve a perfect balance between efficiency and security. Compared to existing top-notch solutions, this scheme reduces computational overhead by more than 50% and communication costs by more than 30%. In addition, rigorous formal security proofs and comprehensive informal analysis also confirm the security features of LiveMark, ensuring that it meets the highest security standards and reliability requirements. BRIEF DESCRIPTION OF DRAWINGS

[0054] The present application will be further described in detail below in conjunction with the drawings and specific embodiments;

[0055] Figure 1 Flowchart of the present application for implementing adversary-aware authentication in federated learning based on adaptive steganographic watermark technology;

[0056] Figure 2 Structure diagram of the present application for implementing adversary-aware authentication in federated learning based on adaptive steganographic watermark technology. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0058] In the modern network environment, various network applications and services make today's networks more complex than traditional networks. Basic network security protection alone is insufficient to completely prevent malicious attacks, especially when insider attackers exist. These attackers not only possess user device registration information but also know the security parameters stored on servers, making it easier to infiltrate the environment and launch attacks from within. Therefore, more reliable mechanisms are needed to ensure the legitimacy of user devices and prevent unauthorized access.

[0059] This invention proposes a security authentication scheme called LiveMark, the main framework of which is as follows: Figure 2 As shown, the process includes a training phase and an authentication phase. The training phase is further divided into three sub-phases: device join request processing, authentication server group configuration, and legitimate device training. In the first training phase, a new device initiates a join request to the authentication server. Upon approval, the server assigns an authentication identifier and constructs a complete identity based on it. This information is encrypted using the server's public key and stored by the server. In the next phase, the authentication server groups all devices and assigns device IDs. Then, it selects a device as the group leader for each group and allocates a global model, a watermark-triggered dataset, and an authentication dataset to all legitimate devices. In the third training phase, devices first train locally using the watermark-triggered dataset and then upload the federated learning parameters to the authentication server to participate in global model aggregation. Next, the device attempts to recognize watermarked data; if successful, it modifies the data label to its own authentication identifier. In the final authentication phase, the server selects an image from the watermark-triggered dataset, embeds a specified index, and then digitally signs the image, forming authentication data which is sent to the group leader devices. Upon receiving the authentication data, the group leader devices verify the digital signature to ensure the reliability of the message source. If the source is reliable, the data is forwarded to group members. Each member extracts the embedding index from the image and then retrieves the corresponding data from the authentication dataset based on that index. The generated response is then uploaded to the server, containing the authentication identifier, the general image inference result, a timestamp, and a value obtained by encrypting the timestamp using the device ID. If the group leader device determines during the verification process that the data source is unreliable, it discards the authentication data, stops the process, and the group leader device then uploads the generated response to the server.

[0060] like Figure 1Or 2, the application discloses an adversary-aware authentication method in federated learning based on adaptive steganographic watermarking, which includes a training phase and an authentication phase, and specific steps are as follows:

[0061] Training phase:

[0062] Step 1, the device constructs its own identity information A I and encrypts and transmits to the authentication server A s ; wherein the identity information A I of the legal device approved by the system administrator to join the network environment includes a unique authentication identifier A id ;

[0063] Step 2, the authentication server A s handles the request of the external device under the current network environment, ensures that each device joined has a controlled validity period A p and clear identity information A I ;

[0064] Step 3, the authentication server A s registers the device ID of the legal device, and groups the legal devices at the same time;

[0065] Step 4, the authentication server generates corresponding watermark trigger data set for each group and allocates authentication data set required for subsequent authentication, and sends the global model and the authentication data set to all legal devices at the same time;

[0066] Step 5, select group leader device for each group based on round rule;

[0067] Step 6, the legal device performs federated training, trains the global model and learns the trigger strategy;

[0068] Authentication phase:

[0069] Step 7, the authentication server performs security authentication on the legal device.

[0070] Further, the specific steps of step 1 are as follows:

[0071] Step 1-1, first, classify the device, and assign a unique authentication identifier A id to the legal device approved by the system administrator to join the network environment (illegal devices that try to access the network without the confirmation of the administrator have no A id );

[0072] Step 1-2, when the device joins the network environment, send a joining request to the authentication server A s ; the authentication server A s receives the request and requires the device to construct and upload complete identity information A IIdentity information A I including manufacturing information of the device, identity information, validity period A p , assigned authentication identifier A id , and associated attributes.

[0073] Step 1-3, identity information A I encrypted using the public key of the authentication server A s before transmission, and the encryption process is as follows: ; where M represents the encrypted information, encrypted using the public key of the authentication server A s .

[0074] Further, the specific method of step 2 is to decrypt and verify whether the identity information A I submitted by the device is valid (the information field is complete and A id valid); if so, set the authentication status of the device to valid in the authentication system, and configure the validity period A p of the device information; the validity period A p of the device information cannot be indefinite, and a threshold can be set to limit the tasks performed by the network each time A p ; otherwise, the verification fails (i.e., the submitted A I is missing necessary fields or lacks a legal A id ), the system will discard or block the corresponding identity information A I , and the corresponding device will be added to the blacklist to prohibit its access to the network environment.

[0075] Further, in step 2, the device re-applies to the authentication server according to the validity period, and the server regularly cleans up invalid devices inside, and the specific method is as follows:

[0076] When the validity period A p of the device information is about to expire, the authentication server A s warns the corresponding device; when the device wants to continue to exist in the environment, it re-initiates the request before the expiration date;

[0077] When the validity period A p of the device information expires, the authentication server A s sets the authentication status of the corresponding device to invalid and adds it to the list of dormant devices;

[0078] When the device (due to deactivation or forgetfulness, etc.) cannot actively exit the network, leading to long-term existence and resource occupation, the authentication server A s periodically cleans up the corresponding devices after the expiration of the validity period A p .

[0079] Further, the specific steps of step 3 are:

[0080] Step 3-1, authentication server A s Obtain the authentication status of the device;

[0081] Step 3-2, authentication server A s Group each legal device and assign a unique device ID, and the ID assigned to each device consists of two parts, the first part is the group number G num , and the second part is a variable-length random number;

[0082] Further, in step 3, the authentication server ensures the balanced number of members between groups based on the grouping probability formula, which is:

[0083] , , where

[0084] wherein, and are positive numbers to avoid division by zero; denotes the group number of the i-th group; denotes the group number of the j-th group; denotes the number of members of group number , and denotes the number of members of group number ; denotes the watermark trigger data set, denotes the number of data types, n denotes the number of groups, and device denotes the i-th group of devices.

[0085] Specifically, device grouping is based on two dimensions, the number of categories in the data set and the current number of members assigned to each group, in order to ensure the balanced distribution of the number of members in each group.

[0086] Further, the specific steps of step 4 are:

[0087] Step 4-1, assign an independent watermark trigger data set to each group for subsequent training, and the data set selected for each group includes a digital image data set and a non-digital image data set, the construction process of the watermark trigger data set is to select the image sample num corresponding to the group number G from the digital image data set , and sequentially embed the image sample into the non-digital image sample corresponding to the group category through image steganography, and the embedding position is non-continuous, the calculation formula is as follows:

[0088] ;

[0089] ;

[0090] wherein, is the watermark embedding position, is the interval between watermark embedding positions, represents calculating new watermark embedding positions, U is a set of integers from 1 to 5, represents a watermark embedding function, represents the data of the watermark embedding position of the watermark trigger data set . represents the group number G num corresponding to the number of data in the digital image data set, represents the group number G num corresponding to the sample number of the digital image data set;

[0091] Step 4-2, distribute the global model, watermark trigger data set and authentication data set to all legal devices.

[0092] Further, in step 5, the authentication server A s randomly selects a device as the group leader device in each group; when a device has served as a group leader device in a task, the priority of the corresponding device to be selected as the group leader device in the next task is reduced; the rotation rule is that the system re-groups all devices at a random time point, and avoids re-allocating the same group of devices to the original group, and after completing the grouping adjustment, the auxiliary device ID, device identifier and authentication data set in the group are regenerated and re-allocated.

[0093] Further, the specific steps of step 6 are:

[0094] Step 6-1, after the legal device receives the content sent by the authentication server A s , it enters the training stage, and divides the watermark trigger data set into two parts locally, including watermark data and non-watermark data, both of which are used for federated learning training;

[0095] Step 6-2, the training is divided into two stages, in the first stage, the device uses the local watermark trigger data set of each group to train the local model and uploads the local model update to the authentication server A s participate in global model aggregation;

[0096] Step 6-3, in the second stage, the device extracts the watermark from the specified position of the watermark data and identifies it; when the watermark is correctly identified, the device modifies the label of the specified position of the watermark data to its own authentication identifier A id .

[0097] Further, the specific steps of step 7 are:

[0098] Step 7-1, authentication server A s Randomly select an image and embed the corresponding specified index; the specified index is a sequence of numbers, indicating the authentication server A s Select an image combination that can constitute a special meaning; the image with the specified index and the authentication server A s The generated digital signature is combined to form authentication data A d , where the digital signature is the authentication server A s A set of ciphertexts generated by the digital signature authentication data A d Whether it comes from the authentication server A s , and finally the authentication data A d Sent to the group leader device G leader ;

[0099] Step 7-2, group leader device G leader After receiving the authentication data A d , verify whether the data source of the digital signature is from the authentication server A s ; if so, determine that the verification is passed and execute step 7-3; otherwise, determine that the data source is unreliable, discard the authentication data and stop, and the group leader device G leader The generated response A r Subsequently uploaded to the authentication server A s And end the current authentication;

[0100] Step 7-3, group leader device G leader Forward the authentication data A d To the group members, and the group members extract the embedded index from the image;

[0101] Step 7-4, the group members generate the response A r After uploading the corresponding data from the authentication data set according to the index to the authentication server A s , the response A r Contains authentication identifier A id , regular image inference result, timestamp and the value obtained by encrypting the timestamp with the device ID.

[0102] The present application proposes a secure authentication scheme named LiveMark, which effectively solves the problem of easy analysis of authentication features in current network environment through federated learning and steganographic watermarking technology.

[0103] In the present application, watermark information is embedded in the specified authentication trigger data as a trigger feature. The device to be authenticated generates a specific authentication identifier based on this trigger feature and hides it in the regular inference result, making the output result appear normal on the surface. In addition, this scheme introduces group management, leader coordination and information rotation mechanism, which ensures the security of authentication while minimizing the association between authentication results and devices. These mechanisms also enhance the controllability and concealment of the authentication process. Experimental results show that the proposed scheme can effectively achieve secure, efficient and covert authentication.

[0104] The present application adopts the above technical scheme, which can effectively guarantee the secure authentication of devices in a network environment. The scheme combines federated learning with watermark trigger mechanism to realize access control, identity authentication and continuous state monitoring of devices in the network, which includes a training phase and an authentication phase. The training phase is subdivided into three sub-phases: device joining request processing, authentication server grouping configuration, and legal device training.

[0105] The present application involves three types of participants: authentication servers, regular devices and group leader devices. Specifically, devices are divided into several groups, each with a designated group leader device. The authentication server is responsible for identity verification of all devices, while the group leader device is responsible for verifying the reliability of the authentication data source and distributing the authentication data to other members in the group. In addition, different access permissions can be assigned to each group according to the grouping strategy, thereby building a complete secure authentication mechanism.

[0106] The present application has the following beneficial effects: 1. Innovation of federated learning authentication paradigm. By seamlessly integrating adaptive steganographic watermark technology with federated learning, this framework creates a new paradigm for covert and robust device authentication, effectively addressing the inherent flaws of current network security frameworks, such as insufficient authentication concealment and vulnerability to attacks. 2. Use of dynamic watermark to achieve tamper-proof authentication. This innovative mechanism achieves a seamless integration of explicit interaction and implicit authentication, hiding the traces of the authentication process while ensuring verifiable trust, thereby significantly improving the security posture of the system. 3. Implementation of implicit authentication in explicit federated interaction. This strategy not only eliminates detectable authentication patterns but also maintains system stability, providing users with a seamless and secure authentication experience without sacrificing performance. 4. Use of strict verification mechanism to achieve perfect balance between efficiency and security. Compared to existing top-notch solutions, this scheme reduces computational overhead by more than 50% and communication cost by more than 30%. In addition, strict formal security proofs and comprehensive informal analysis also confirm the security features of LiveMark, ensuring that it meets the highest security standards and reliability requirements.

[0107] It is apparent that the described embodiments are only some — but not all — of the embodiments of the present application. The embodiments described in this application and features in the embodiments can be combined with each other in cases without conflict. The components of the embodiments of the present application, which are generally described and shown in the accompanying drawings, can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work, shall fall within the scope of protection of the present application.

Claims

1. An adversary-aware authentication method in federated learning based on adaptive steganographic watermarking, characterized in that: It includes a training phase and an authentication phase, and the specific steps are as follows: Training phase: Step 1, device builds its own identity information A I and encrypts it for transmission to the authentication server A s ; wherein the identity information of the legitimate device that has been approved by the system administrator to join the network environment A I comprising a unique authentication identifier A id ; Step 2, authentication server A s Handling requests from external devices in the current network environment, ensuring that each device that joins has a controlled validity period A p and explicit identity information A I ; Step 3, authentication server A s The legal device is registered with the device ID, and the legal device is grouped. Step 4, the authentication server generates corresponding watermark trigger data set for each group and assigns authentication data set required for subsequent authentication, and then sends the global model and authentication data set to all legal devices; the specific steps of step 4 are: Step 4-1: Assign an independent watermark-triggered dataset to each group for subsequent training; the dataset selected for each group includes both digital image datasets and non-digital image datasets. Image samples from the digital image dataset are then processed using image steganography. After being sequentially embedded into non-digital image samples corresponding to the group classification, a watermark-triggered dataset is formed. The embedding positions are non-contiguous, and the embedding expression is as follows: ； ； wherein, is a watermark embedding position, is a spacing between watermark embedding positions, is a new watermark embedding position for calculating, U is a set of integers from 1 to 5, denotes a watermark embedding function, denotes data of watermark embedding positions of the watermarkless trigger data set denotes a group number G num the number data in the corresponding digital image data set, denotes a group number G num a sample number of the corresponding digital image data set;​​ Step 4-2, distribute the global model, watermark trigger data set and authentication data set to all legal devices; Step 5, select group leader device for each group based on rotation rule; Step 6, the legal device performs federated training, trains the global model and learns the trigger strategy; the specific steps of step 6 are: Step 6-1, legitimate device receives authentication server A s After sending the content, enter the training phase, divide the watermark trigger dataset into two parts locally, containing watermark data and non-watermark data, both of which are used for federated learning training; Step 6-2, the training is divided into two stages, in the first stage, the device trains the local model using each respective local watermark trigger dataset and uploads the local model updates to the authentication server A s participate in global model aggregation; Step 6-3, second stage, the device extracts the watermark from the designated location of the watermarked data and identifies it; when the watermark is correctly identified, the device modifies the tag of the designated location of the watermarked data to its own authentication identifier A id ; Authentication phase: Step 7, the authentication server conducts security authentication on the legal device.

2. The method of claim 1, wherein the method is based on adaptive steganographic watermarking. The specific steps of step 1 are: Step 1-1, first classifying the device, assigning a unique authentication identifier to a legitimate device that has been approved by a system administrator to join the network environment A id ; Steps 1-2, when a device joins the network environment, it sends a join request to the authentication server A s The authentication server sends a request for the device to build and upload a complete identity profile A s Upon receiving the request, the device is required to build and upload a complete identity profile A I The identity profile A I includes manufacturing information of the device, validity period of the identity profile A p , assigned authentication identifier A id and associated attributes; Step 1-3, identity information A I Using authentication server before transmission A s The public key of the authentication server is used for encryption, and the encryption process is as follows: ; wherein M represents the encrypted information, represents the encryption using the public key of the authentication server A s .

3. The method of claim 1, wherein the method further comprises: The specific method of step 2 is: decrypt and verify the identity information submitted by the device A I If yes, set the authentication state of the device to valid in the authentication system, and configure the valid period of the device information A p ; otherwise , the verification fails, the system will correspond to the identity information A I Discard or stop, and the corresponding device is listed in the blacklist, prohibit its access to the network environment.

4. The adversary-aware authentication method in federated learning based on adaptive stego-watermarking according to claim 1 or 3, characterized in that: In step 2, the device re-applies to the authentication server according to the validity period, and the server regularly cleans up the internal invalid devices, and the specific method is as follows: When the validity period of the device information is about to expire, the authentication server A p warns the respective device; when the device wishes to continue to exist in the environment, then the request is reinitiated before the expiration A s warns the respective device; when the device wishes to continue to exist in the environment, then the request is reinitiated before the expiration When the validity period of the device information expires A p The authentication server A s The authentication state of the corresponding device is set to invalid, and is added to the dormant device list; When the device cannot actively exit the network, causing long-term existence and occupation of resources, the authentication server A s The effective period of the corresponding device A p Periodically clean up after expiration.

5. The method of adversary-aware authentication in federated learning based on adaptive stego-watermarking according to claim 1, characterized in that: The specific steps of step 3 are: Step 3-1, authentication server A s obtaining an authentication status of the device; Step 3-2, authentication server A s Grouping and assigning unique device to each legitimate device ID , each device assigned ID Composed of two parts, the first part is the group number G num The second part is a variable length random number.

6. The adversary-aware authentication method in federated learning based on adaptive stego-watermarking according to claim 1 or 5, characterized in that: In step 3, the authentication server ensures the balanced number of inter-group members of device grouping based on the grouping probability formula, and the grouping probability formula is: , , where wherein, and are positive numbers to avoid division by zero; denotes the group number of the i-th group; denotes the group number of the j-th group; denotes the number of members of the group number denotes the number of members of the group number denotes the set of watermark trigger data, denotes the number of data types, n denotes the number of groups, device denotes the device belonging to the i-th group.​​ 7. The method of claim 1, wherein: Authentication server in step 5 A s Randomly select a device in each group as a group leader device; When a device has served as a group leader device in a task, the priority of the corresponding device being selected as a group leader device in the next task execution is reduced; The rotation rule is: the system re-groups all devices at a random time point, and avoids re-allocating the same group of devices to the original group. After completing the grouping adjustment, the auxiliary device ID, device identifier and authentication data set in the group are regenerated and re-allocated.

8. The method of adversary-aware authentication in federated learning based on adaptive stego-watermarking according to claim 1, characterized in that: The specific steps of step 7 are: Step 7-1, authentication server A s Randomly select an image and embed the corresponding specified index; the specified index is a sequence of numbers, indicating the authentication server A s Select an image combination that can constitute a special meaning; the image with the specified index is combined with the authentication server A s The generated digital signature is combined to form authentication data A d , wherein the digital signature is the authentication server A s A set of ciphertexts are generated, which verify the authentication data through the digital signature A d Whether it comes from the authentication server A s , and finally the authentication data A d is sent to the group leader device G leader ; Step 7-2, group leader device G leader Upon receiving the authentication data A d , the group leader device verifies whether the data source of the digitally signed data is from the authentication server A s ; if yes, it determines that the verification is passed and proceeds to Step 7-3; otherwise, it determines that the data source is unreliable, directly discards the authentication data and stops further proceeding, and sends a response to the group member device G leader A r to the authentication server A s and ends the current authentication;​ Step 7-3, group leader device G leader forwarding the authentication data A d to the group members, which extract the embedding index from the image; Step 7-4, group member gets corresponding data from authentication data set according to index to generate response A r Post to authentication server A s Response A r Contains authentication identifier A id , regular image inference result, timestamp, and value obtained after encrypting timestamp by device ID ​

Images