Self-adaptive Byzantine attack defense method based on multi-mechanism cooperation

Abstract

The invention discloses a self-adaptive Byzantine attack defense method based on multi-mechanism cooperation, and relates to the technical field of radio networks. The method runs in a cognitive radio network comprising a primary user, a fusion center and a plurality of secondary users, and the core lies in that a self-adaptive closed-loop defense system is constructed through the synergistic effect of four core components including a dynamic reputation updating mechanism, a soft weight mapping mechanism, a sequential fusion and early stop mechanism and a lagging state machine; therefore, high sensing precision, high decision-making efficiency and high system stability are synchronously realized in a complex attack environment.

Description

Technical Field

[0001] This invention relates to the field of radio network technology, specifically to an adaptive Byzantine attack defense method based on multi-mechanism collaboration. Background Technology

[0002] In existing technologies, common defense methods against Byzantine attacks in the field of cognitive radio networks include:

[0003] 1. Traditional Weighted Sequential Probability Ratio Detection (WSPRT) scheme:

[0004] • Solution Summary: This solution dynamically adjusts the weight of locally perceived results in the final decision-making process based on user reputation scores, directly incorporating reputation scores into the likelihood ratio calculation. It employs a sequential likelihood ratio test, processing data according to the order of user reports, and terminates the data fusion process when the cumulative likelihood ratio reaches a preset decision threshold.

[0005] 2. A fusion scheme based on Bayesian inference:

[0006] • Solution Summary: Collaborative spectrum sensing is modeled as a Bayesian estimation problem. The fusion center knows (or estimates) the report probability distribution (i.e., detection probability and false alarm probability) for each secondary user under two assumptions (presence / absence of the primary user) and uses it as a likelihood function. Combining the prior probability of the primary user's state, after receiving reports from all users, the fusion center calculates its posterior probability and selects the assumption with the larger posterior probability as the global decision.

[0007] 3. Defense solutions based on Beta reputation:

[0008] • Solution Summary: A Beta distribution is used to model and update user reputation scores. After each perception, the reputation parameter is updated based on the accuracy of the user's report. The user's reputation score is typically taken as the expectation of the Beta distribution. This reputation score can be used directly as a fusion weight or compared with a fixed threshold to determine whether a user is malicious.

[0009] 4. Machine learning-based defense solutions:

[0010] • Solution Summary: This method treats malicious user identification as a classification problem. First, a training dataset is collected, containing user reporting behavior under known spectral conditions. Then, features are used to train a classifier (such as a support vector machine, neural network, decision tree, etc.). In the online phase, the trained model classifies users based on their real-time behavioral features, identifying malicious users and excluding their reports from the fusion process.

[0011] The objective shortcomings of existing technologies are mainly reflected in the following four aspects:

[0012] (1) The reliability of decision-making is disconnected from the reputation assessment.

[0013] Disadvantages: Traditional schemes (such as traditional WSPRT and Beta reputation) do not consider the reliability of the current global decision itself when updating reputation or assigning weights. Regardless of whether the decision is made under high confidence or low confidence (such as when the channel is poor), the same set of standards is used for rewards and penalties.

[0014] (2) The inherent contradiction between perception efficiency and decision-making accuracy

[0015] Disadvantages: Traditional WSPRT merges data in a fixed order, which can easily lead to prioritizing low-quality users and delaying correct decisions. Bayesian methods, on the other hand, typically require waiting for all reports and cannot terminate early. Neither method can intelligently minimize fusion overhead while strictly guaranteeing accuracy.

[0016] (3) Insufficient system self-adaptation and state stability

[0017] Disadvantages: Schemes such as Beta Reputation update reputation slowly and respond slowly to new malicious users; at the same time, classification methods based on fixed thresholds (see Beta Reputation and some simple WSPRT variants) are prone to frequent oscillations between "honest" and "malicious" user status when facing channel fluctuations, resulting in poor system stability.

[0018] (4) Model dependency and practical bottleneck

[0019] Disadvantages: Bayesian methods heavily rely on accurate prior probabilities and attack models, while machine learning methods require large amounts of labeled data for training and have weak generalization ability. Both methods experience significant performance degradation when faced with model mismatch or novel attacks, and lack robustness in unknown dynamic environments.

[0020] Therefore, a new solution is needed to address the above problems. Summary of the Invention

[0021] The purpose of this invention is to provide an adaptive Byzantine attack defense method based on multi-mechanism collaboration to solve the technical problems mentioned in the background art.

[0022] To achieve the above objectives, the present invention provides the following technical solution: an adaptive Byzantine attack defense method based on multi-mechanism collaboration, comprising at least the following steps:

[0023] S1: Start running, boot the system;

[0024] S2: Perform local spectrum sensing. All secondary users in the cognitive radio network independently perform energy detection on the primary user's spectrum, generate local binary sensing results, where 1 indicates spectrum occupancy and 0 indicates spectrum idleness, and report the results to the fusion center.

[0025] S3: Set up a dynamic reputation update mechanism. The fusion center conducts dynamic reputation assessments for each user based on the global decision results of the previous cycle.

[0026] S4: Set up a soft weight mapping mechanism. The fusion center will transform the updated user reputation value into a basic weight for data fusion through normalization and non-linear S-shaped function mapping. The basic weight will then be decayed according to the user's status label, and finally output the final fusion weight after minimum weight constraint and mean normalization processing.

[0027] S5: Using a hysteresis state machine, the fusion center updates the user's hysteresis counter based on the user's reputation value and a dynamic threshold calculated based on the overall reputation statistical characteristics. Only when the user's reputation value is continuously higher than the upper threshold or lower than the lower threshold for a preset number of confirmation frames will the transition of its state (malicious, suspicious, honest) be triggered, thereby ensuring the stability of user classification. The hysteresis state machine is used to maintain stability.

[0028] S6: Introducing sequential fusion and early stopping mechanisms, which are the core of improving decision-making efficiency;

[0029] S7: After merging each user's report, determine whether the current accumulated amount of evidence exceeds the preset decision threshold. At the same time, based on the weight of the remaining users and the upper limit of single-sample evidence, estimate the potential range of the remaining evidence and determine whether it is possible to stop early.

[0030] S8: Output global decision. When the judgment of S7 is yes, the fusion center immediately stops fusion and outputs the final global binary decision (whether the main user exists).

[0031] S9: The system determines whether to start the next sensing cycle;

[0032] S10: System terminated.

[0033] Furthermore, the dynamic reputation update mechanism includes at least the following steps:

[0034] First, a forgetting factor is applied to decay historical reputation;

[0035] Then calculate the confidence level of the global decision in the previous frame;

[0036] Finally, based on confidence gating, the consistency between user reports and global decisions is judged, and user reputation values ​​are updated with an asymmetric adaptive step size, while users who make consecutive mistakes are punished more severely.

[0037] The dynamic reputation update mechanism is the foundation for the system to achieve self-adaptive capabilities. Its inputs are historical reputation, current user reports, and global decisions, and its output is the updated reputation value.

[0038] Furthermore, the application of the forgetting factor to decay historical reputation includes at least the following steps:

[0039] First, the reputation value from the previous moment is decayed to prevent historical data from having too much influence.

[0040] (1)

[0041] in, For users At any moment Reputation value Forgetting factor;

[0042] The calculation of the confidence level of the global decision in the previous frame includes at least the following steps:

[0043] The confidence level of the decision is calculated based on the cumulative likelihood ratio, and the reliability of the current global decision is assessed to determine whether to update the confidence.

[0044] (2)

[0045] in, For the current cumulative likelihood ratio, As the decision threshold, The closer the value is to 1, the more reliable the decision.

[0046] The asymmetric adaptive step-size update of user reputation value includes at least the following steps:

[0047] when Above the threshold Update according to the following rules:

[0048] Reward, report correct:

[0049] (3)

[0050] Punishment, report error:

[0051] (4)

[0052] in, and They are positive and negative step sizes, respectively, and With continuous error counter Positive correlation enables severe penalties for persistent malicious behavior;

[0053] A reputation update is triggered only when the decision confidence level exceeds a certain threshold.

[0054] Enhance the credibility of users whose reports align with the final decisions;

[0055] For users who report inconsistent information, their reputation will be lowered, and the reduction will increase significantly with the number of consecutive errors made by the user.

[0056] Furthermore, the soft weight mapping mechanism smoothly and non-linearly maps reputation values ​​to fusion weights;

[0057] The soft weight mapping mechanism includes at least the following steps:

[0058] First, normalization is performed, standardizing the reputation scores of all users to a standard range:

[0059] (5)

[0060] in It is a small positive number used to prevent division by zero errors;

[0061] Then, a nonlinear mapping is performed. Through an S-shaped nonlinear function, the normalized reputation value is smoothly mapped to a basic weight. The S-shaped nonlinear function shows that the weight of users with medium reputation values ​​changes sensitively, while the weight of users with extremely high or low reputation values ​​changes slowly.

[0062] (6)

[0063] in, The slope factor controls the "softness" of the mapping;

[0064] Finally, state decay is performed. Based on the user's current state label (such as "malicious", "suspicious", "honest") given by the hysteresis state machine, the basic weights are decayed by different proportions.

[0065] (7)

[0066] in, The state labels output by the hysteresis state machine are (2: honest, 1: suspicious, 0: malicious). This represents the corresponding attenuation coefficient.

[0067] Furthermore, S5 includes at least the following steps:

[0068] First, dynamic thresholds are set. Based on the average and dispersion of all users' reputation scores, an upper and lower threshold for a category are dynamically calculated:

[0069] (8)

[0070] (9)

[0071] in, and The mean and standard deviation of the reputation score. It is a scaling factor that controls the deviation of the threshold from the average value;

[0072] Then, hysteresis counting is performed, maintaining two counters for each user: one to record the number of frames in which their reputation value is consistently above the upper threshold, and the other to record the number of frames in which their reputation value is consistently below the lower threshold.

[0073] When the user's credit score is consistently higher than or below When this occurs, the corresponding "rise counter" or "fall counter" increases;

[0074] Finally, the status transition is performed. The user's classification status (such as "malicious", "suspicious", "honest") will not change immediately due to a one-time fluctuation in their reputation value. Only when the corresponding counter confirms that the user's reputation value has continuously deviated from the threshold for a preset confirmation period will the system perform a status upgrade or downgrade.

[0075] Furthermore, S6 includes at least the following steps:

[0076] The fusion center first sorts all user reports in descending order of weight based on the weights calculated in step S4.

[0077] Within each perception cycle, the fusion center, based on calculated weights, assigns users according to their weights. Arrange in descending order to obtain the sequence ;

[0078] Then, the weighted evidence value of each report is calculated sequentially in this order and added to the global accumulated evidence volume for sequential fusion. Following this order, each user's report is converted into weighted evidence and accumulated to the global evidence volume.

[0079] Finally, an early termination decision is made; after each user's report is integrated, it is determined whether the accumulated amount of evidence has exceeded the decision threshold.

[0080] When the threshold is reached, a global decision is made immediately and the fusion is terminated;

[0081] If the threshold is not reached, the maximum possible range of evidence that the remaining unfused reports can provide will be dynamically estimated.

[0082] If this range indicates that even merging all remaining users would not change the current preliminary conclusions (e.g., never reaching the threshold for the opposite state), the system will stop merging early and make a final decision based on the current trend of evidence.

[0083] Compared with the prior art, the beneficial effects of the present invention are:

[0084] This invention operates in a cognitive radio network that includes a primary user, a fusion center, and multiple secondary users. Its core lies in the synergistic effect of four core components: a dynamic reputation update mechanism, a soft weight mapping mechanism, a sequential fusion and early stopping mechanism, and a hysteresis state machine. This constructs an adaptive closed-loop defense system, thereby simultaneously achieving high perception accuracy, high decision-making efficiency, and high system stability in complex attack environments. Attached Figure Description

[0085] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0086] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0087] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0088] Please see Figure 1 An adaptive Byzantine attack defense method based on multi-mechanism collaboration includes at least the following steps:

[0089] S1: Start running, boot the system;

[0090] S2: Perform local spectrum sensing. All secondary users in the cognitive radio network independently perform energy detection on the primary user's spectrum, generate local binary sensing results, where 1 indicates spectrum occupancy and 0 indicates spectrum idleness, and report the results to the fusion center.

[0091] S3: Set up a dynamic reputation update mechanism. The fusion center conducts dynamic reputation assessments for each user based on the global decision results of the previous cycle.

[0092] S4: Set up a soft weight mapping mechanism. The fusion center will transform the updated user reputation value into a basic weight for data fusion through normalization and non-linear S-shaped function mapping. The basic weight will then be decayed according to the user's status label, and finally output the final fusion weight after minimum weight constraint and mean normalization processing.

[0093] S5: Using a hysteresis state machine, the fusion center updates the user's hysteresis counter based on the user's reputation value and a dynamic threshold calculated based on the overall reputation statistical characteristics. Only when the user's reputation value is continuously higher than the upper threshold or lower than the lower threshold for a preset number of confirmation frames will the transition of its state (malicious, suspicious, honest) be triggered, thereby ensuring the stability of user classification. The hysteresis state machine is used to maintain stability.

[0094] S6: Introducing sequential fusion and early stopping mechanisms, which are the core of improving decision-making efficiency;

[0095] S7: After merging each user's report, determine whether the current accumulated amount of evidence exceeds the preset decision threshold. At the same time, based on the weight of the remaining users and the upper limit of single-sample evidence, estimate the potential range of the remaining evidence and determine whether it is possible to stop early.

[0096] S8: Output global decision. When the judgment of S7 is yes, the fusion center immediately stops fusion and outputs the final global binary decision (whether the main user exists).

[0097] S9: The system determines whether to start the next sensing cycle. In most application scenarios, collaborative spectrum sensing is a continuous process, so it usually returns "Yes," returning the process to step S2 to begin a new round of sensing, reputation updating, and fusion decision-making, forming an adaptive closed-loop system. When the system is powered off or actively stopped, select "No," and the process ends.

[0098] S10: System terminated.

[0099] A dynamic reputation update mechanism includes at least the following steps:

[0100] First, a forgetting factor is applied to decay historical reputation;

[0101] Then calculate the confidence level of the global decision in the previous frame;

[0102] Finally, based on confidence gating, the consistency between user reports and global decisions is judged, and user reputation values ​​are updated with an asymmetric adaptive step size, while users who make consecutive mistakes are punished more severely.

[0103] The dynamic reputation update mechanism is the foundation for the system to achieve self-adaptive capabilities. Its inputs are historical reputation, current user reports, and global decisions, and its output is the updated reputation value.

[0104] Compared to existing technologies: Traditional reputation mechanisms update without considering the reliability of the decision itself and punish both accidental errors and malicious attacks indiscriminately. This mechanism avoids erroneous updates under uncertain decisions through confidence gating and accurately distinguishes between accidental errors and malicious attacks through asymmetric step size and continuous error counting.

[0105] Applying a forgetting factor to decay historical reputation includes at least the following steps:

[0106] First, the reputation value from the previous moment is decayed to prevent historical data from having too much influence.

[0107] (1)

[0108] in, For users At any moment Reputation value Forgetting factor;

[0109] Calculating the confidence level of the global decision in the previous frame includes at least the following steps:

[0110] The confidence level of the decision is calculated based on the cumulative likelihood ratio, and the reliability of the current global decision is assessed to determine whether to update the confidence.

[0111] (2)

[0112] in, For the current cumulative likelihood ratio, As the decision threshold, The closer the value is to 1, the more reliable the decision.

[0113] Asymmetric adaptive step-size update of user reputation values ​​includes at least the following steps:

[0114] when Above the threshold Update according to the following rules:

[0115] Reward, report correct:

[0116] (3)

[0117] Punishment, report error:

[0118] (4)

[0119] in, and They are positive and negative step sizes, respectively, and With continuous error counter Positive correlation enables severe penalties for persistent malicious behavior;

[0120] A reputation update is triggered only when the decision confidence level exceeds a certain threshold.

[0121] Enhance the credibility of users whose reports align with the final decisions;

[0122] For users who report inconsistent information, their reputation will be lowered, and the reduction will increase significantly with the number of consecutive errors made by the user.

[0123] The soft weight mapping mechanism smoothly and non-linearly maps reputation values ​​to fusion weights;

[0124] The soft weight mapping mechanism includes at least the following steps:

[0125] First, normalization is performed, standardizing the reputation scores of all users to a standard range:

[0126] (5)

[0127] in It is a small positive number used to prevent division by zero errors;

[0128] Then, a nonlinear mapping is performed. Through an S-shaped nonlinear function, the normalized reputation value is smoothly mapped to a basic weight. The S-shaped nonlinear function shows that the weight of users with medium reputation values ​​changes sensitively, while the weight of users with extremely high or low reputation values ​​changes slowly.

[0129] (6)

[0130] in, The slope factor controls the "softness" of the mapping;

[0131] Finally, state decay is performed. Based on the user's current state label (such as "malicious", "suspicious", "honest") given by the hysteresis state machine, the basic weights are decayed by different proportions.

[0132] (7)

[0133] in, The state labels output by the hysteresis state machine are (2: honest, 1: suspicious, 0: malicious). This represents the corresponding attenuation coefficient.

[0134] Compared with existing technologies: Existing solutions mostly use linear mapping or hard decision-making, which can cause small fluctuations in user reputation to lead to drastic changes in their weights. This mechanism avoids abrupt decision changes through nonlinear smooth mapping and achieves smooth suppression of the weights of bad users through state-dependent decay, rather than abruptly cutting them off, thus enhancing the robustness of the system.

[0135] S5 includes at least the following steps:

[0136] First, dynamic thresholds are set. Based on the average and dispersion of all users' reputation scores, an upper and lower threshold for a category are dynamically calculated:

[0137] (8)

[0138] (9)

[0139] in, and The mean and standard deviation of the reputation score. It is a scaling factor that controls the deviation of the threshold from the average value;

[0140] Then, hysteresis counting is performed, maintaining two counters for each user: one to record the number of frames in which their reputation value is consistently above the upper threshold, and the other to record the number of frames in which their reputation value is consistently below the lower threshold.

[0141] When the user's credit score is consistently higher than or below When this occurs, the corresponding "rise counter" or "fall counter" increases;

[0142] Finally, the status transition is performed. The user's classification status (such as "malicious", "suspicious", "honest") will not change immediately due to a one-time fluctuation in their reputation value. Only when the corresponding counter confirms that the user's reputation value has continuously deviated from the threshold for a preset confirmation period will the system perform a status upgrade or downgrade.

[0143] Compared to existing technologies, traditional methods use fixed thresholds to classify users, which can easily lead to frequent oscillations between "honest" and "malicious" user states when channel conditions fluctuate drastically. This mechanism adapts to changes in the overall network performance through dynamic thresholds and effectively filters out temporary fluctuations through a hysteresis acknowledgment mechanism, making user state determination more stable and reliable, and fundamentally improving system stability.

[0144] S6 includes at least the following steps:

[0145] The fusion center first sorts all user reports in descending order of weight based on the weights calculated in step S4.

[0146] Within each perception cycle, the fusion center, based on calculated weights, assigns users according to their weights. Arrange in descending order to obtain the sequence ;

[0147] Then, the weighted evidence value of each report is calculated sequentially in this order and added to the global accumulated evidence volume for sequential fusion. Following this order, each user's report is converted into weighted evidence and accumulated to the global evidence volume.

[0148] Finally, an early termination decision is made; after each user's report is integrated, it is determined whether the accumulated amount of evidence has exceeded the decision threshold.

[0149] When the threshold is reached, a global decision is made immediately and the fusion is terminated;

[0150] If the threshold is not reached, the maximum possible range of evidence that the remaining unfused reports can provide will be dynamically estimated.

[0151] If this range indicates that even merging all remaining users would not change the current preliminary conclusions (e.g., never reaching the threshold for the opposite state), the system will stop merging early and make a final decision based on the current trend of evidence.

[0152] Compared to existing technologies, traditional WSPRT only fuses evidence in a fixed order and lacks early stopping analysis. This mechanism prioritizes the fusion of high-quality evidence through reputation ranking and minimizes the number of samples by using accessibility analysis for early stopping decisions, thus greatly improving efficiency while strictly ensuring performance.

[0153] The working principle is as follows:

[0154] 1. A dynamic reputation update method based on confidence gating and asymmetric penalty

[0155] Operational process: First, calculate the confidence level of the current global decision; second, compare this confidence level with a preset threshold, and only when the confidence level is high enough can the subsequent reputation update operation be performed; then, for users whose reports are consistent with the decision, give positive rewards of varying degrees according to their current reputation level; for users whose reports are inconsistent with the decision, implement asymmetric and differentiated penalties based on their history of consecutive errors—that is, impose a much greater reputation penalty on users with more consecutive errors than on users with occasional errors.

[0156] The essence of what needs to be protected: This is the core of the entire defense system, solving the critical problem that traditional solutions cannot distinguish between occasional errors and persistent malicious behavior. It protects the gating principle of "updating reputation only when the decision is reliable," and the specific means of distinguishing between accidental errors and malicious behavior—"implementing differentiated penalties based on the persistence of errors." This is the core of achieving high precision.

[0157] 2. A sequential early stopping decision method integrating reputation ranking and reachability analysis

[0158] Operational process: First, within each perception cycle, user reports are sorted in descending order based on their real-time fusion weights. Then, user reports are fused sequentially in this order, and accumulated evidence is updated. After fusing each report, two judgments are performed: a) Determine whether the accumulated evidence has exceeded the decision threshold; b) Dynamically estimate the range of evidence changes that all remaining unfused reports can provide, and determine whether this range is unlikely to change the current preliminary decision tendency. If either a) or b) is met, fusion is immediately stopped and the final decision is output.

[0159] The essence of what needs to be protected is the combination of "determining the fusion order based on reputation weights" and "early cessation based on the potential range of remaining evidence." This is the core of achieving high efficiency.

[0160] 3. Soft-weight mapping method combining nonlinear mapping and state decay

[0161] Operational process: First, normalize the user's reputation value; then, smoothly map the normalized reputation value to the initial fusion weights using an S-shaped nonlinear function; next, apply different proportions of state decay coefficients to the initial weights based on the user's current state (e.g., malicious, suspicious, honest) determined by the hysteresis state machine; finally, normalize all weights and apply minimum weight constraints.

[0162] The essence of the protection sought is to safeguard the combination of "using nonlinear functions to achieve a smooth mapping from reputation to weights" and "weight decay based on user state." This is crucial for achieving system robustness and avoiding abrupt decision changes.

[0163] 4. A Three-State User Classification Method Based on Dynamic Thresholds and Lag Acknowledgment

[0164] Operation process: First, dynamically calculate the upper and lower thresholds for classification based on the statistical distribution (mean, standard deviation) of all user reputation values; then, maintain two independent hysteresis counters for each user to record the number of frames in which their reputation value is continuously higher than the upper threshold or continuously lower than the lower threshold; finally, only when either hysteresis counter reaches the preset number of confirmation frames will the user's status be triggered to switch between "malicious", "suspicious", and "honest".

[0165] The essence of the protection sought is to safeguard the stable classification mechanism that relies on both "dynamic thresholds based on group reputation statistics" and "state transition confirmation via hysteresis counters." This is crucial for achieving system stability.

[0166] 5. A closed-loop, collaborative defense system comprised of the aforementioned four mechanisms.

[0167] Operational process: Integrate the methods described in technical points 1 to 4 above to form a closed-loop system that executes in a loop: the output of dynamic reputation update serves as the input of soft weight mapping; the result of weight mapping guides sequential fusion; the global decision feedback generated by fusion drives dynamic reputation update and the hysteresis state machine; the output of the state machine then affects the soft weight mapping of the next cycle.

[0168] The system architecture itself integrates the four independent technical points as a whole, working collaboratively to form a closed loop. This is the top-level design for achieving the synergistic amplification of all technical effects (high precision, high efficiency, and high stability).

[0169] In summary, the technical effects of the present invention are as follows:

[0170] 1. Improved perception accuracy and robustness

[0171] Effect: This invention can maintain a very high perception accuracy even in extreme attack environments where malicious users account for the vast majority.

[0172] Principle Analysis: Unlike traditional reputation mechanisms that blindly update all decision results, this invention innovatively introduces a decision confidence assessment. The system first determines whether the current global decision is sufficiently reliable; only under a high confidence level will the reputation update mechanism be triggered. This "confidence gating" design fundamentally avoids making incorrect reputation evaluations of users when channel conditions are poor or the decision itself is questionable. Furthermore, in terms of the penalty mechanism, this invention adopts an asymmetric strategy, distinguishing between accidental perceptual errors and systemic malicious attacks. For users who occasionally make mistakes, the system imposes a mild penalty; while for users who continuously report errors, a rapid and severe reputation decay is imposed. This refined differentiation capability is the core reason why this method can maintain an extremely low false positive rate even under a high proportion of malicious attacks.

[0173] 2. High efficiency and low cost

[0174] Effect: This invention can achieve global decision-making quickly with far fewer samples than traditional methods, greatly improving the efficiency of spectrum sensing and the real-time performance of the system.

[0175] Principle Analysis: The efficiency improvement of this invention stems from two key designs. First, it's a descending order of fusion based on reputation weight. In each round of perception, the fusion center no longer processes reports in a fixed order, but prioritizes evidence provided by users with high reputation and high weight. This ensures that the most likely correct information is used first, thereby rapidly improving the quality of accumulated evidence and accelerating the decision-making process. Second, and more innovative, is the early stopping mechanism based on evidence potential prediction. After fusing each report, the system estimates in real time the upper and lower boundaries of the amount of evidence that all remaining reports can provide. If the calculation finds that even fusing all remaining reports will absolutely not change the current decision bias (e.g., it will never be possible to flip from "judged as idle" to "judged as occupied"), then fusion will immediately stop, and the current decision will be output directly. This mechanism mathematically ensures that the reliability of the decision is not sacrificed, while boldly skipping a large number of redundant or low-value reports, achieving a leap in efficiency.

[0176] 3. Enhanced system stability

[0177] Effects: This invention can effectively avoid frequent fluctuations in user status (honest / suspicious / malicious), ensuring stable system operation and insensitivity to temporary interference.

[0178] Principle Analysis: Traditional methods use fixed thresholds to classify users. When channel conditions fluctuate, causing user reputation values ​​to jitter around the threshold, their status repeatedly swings between "honest" and "malicious," severely compromising system stability. This invention solves this problem in two ways. First, it employs a dynamic threshold, which is not a fixed value but dynamically calculated based on the average level and dispersion of the reputation of all users. This allows the system to adapt to changes in the overall network reputation, avoiding large-scale misjudgments caused by temporary widespread channel fading. Second, it introduces a hysteresis state machine, which requires a user's reputation value to be consistently better or worse than the threshold for a certain period of time (i.e., a consecutive number of sensing frames) before their state changes. This "delayed confirmation" mechanism acts like a filter, effectively smoothing reputation fluctuations caused by instantaneous deep fading or noise spikes, ensuring the final stability of user state determination, thereby improving the robustness of the entire system.

[0179] 4. Achieve a unified balance between sensing efficiency and accuracy.

[0180] Results: This invention achieves optimal sensing efficiency while ensuring sensing reliability, thus achieving a unified balance between sensing accuracy and efficiency.

[0181] Principle Analysis: The four mechanisms form a tightly coupled, self-optimizing closed-loop system. Reputation updates provide input for weight mapping, weight mapping guides sequential fusion, the fusion result drives reputation updates and state machine evolution, and the state machine, in turn, influences weight mapping. This collaboration prioritizes the integration of high-reliability reports through precise credibility assessment and terminates the fusion process promptly when confidence conditions are met.

[0182] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An adaptive Byzantine attack defense method based on multi-mechanism collaboration, characterized in that: At least the following steps are included: S1: Start running, boot the system; S2: Perform local spectrum sensing. All secondary users in the cognitive radio network independently perform energy detection on the primary user's spectrum, generate local binary sensing results, where 1 indicates spectrum occupancy and 0 indicates spectrum idleness, and report the results to the fusion center. S3: Set up a dynamic reputation update mechanism. The fusion center conducts dynamic reputation assessments for each user based on the global decision results of the previous cycle. S4: Set up a soft weight mapping mechanism. The fusion center will transform the updated user reputation value into a basic weight for data fusion through normalization and non-linear S-shaped function mapping. The basic weight will then be decayed according to the user's status label, and finally output the final fusion weight after minimum weight constraint and mean normalization processing. S5: Using a hysteresis state machine, the fusion center updates the user's hysteresis counter based on the user's reputation value and a dynamic threshold calculated based on the statistical characteristics of the entire reputation. The user's state transition is only triggered when the user's reputation value is continuously higher than the upper threshold or lower than the lower threshold for a preset number of confirmation frames. This ensures the stability of user classification. The hysteresis state machine is used to maintain stability. S6: Introducing sequential fusion and early stopping mechanisms, which are the core of improving decision-making efficiency; S7: After merging each user's report, determine whether the current accumulated amount of evidence exceeds the preset decision threshold. At the same time, based on the weight of the remaining users and the upper limit of single-sample evidence, estimate the potential range of the remaining evidence and determine whether it is possible to stop early. S8: Output global decision. When the judgment of S7 is yes, the fusion center immediately stops fusion and outputs the final global binary decision. S9: The system determines whether to start the next sensing cycle; S10: System terminated.

2. The adaptive Byzantine attack defense method based on multi-mechanism collaboration according to claim 1, characterized in that: The dynamic reputation update mechanism includes at least the following steps: First, a forgetting factor is applied to decay historical reputation; Then calculate the confidence level of the global decision in the previous frame; Finally, based on confidence gating, the consistency between user reports and global decisions is judged, and user reputation values ​​are updated with an asymmetric adaptive step size, while users who make consecutive mistakes are punished more severely. The dynamic reputation update mechanism is the foundation for the system to achieve adaptive capabilities. Its inputs are historical reputation, current user reports, and global decisions, and its output is the updated reputation value.

3. The adaptive Byzantine attack defense method based on multi-mechanism collaboration according to claim 2, characterized in that: The application of the forgetting factor to decay historical reputation includes at least the following steps: First, the reputation value from the previous moment is decayed to prevent historical data from having too much influence. (1) in, For users At any moment Reputation value Forgetting factor; The calculation of the confidence level of the global decision in the previous frame includes at least the following steps: The confidence level of the decision is calculated based on the cumulative likelihood ratio, and the reliability of the current global decision is assessed to determine whether to update the confidence. (2) in, For the current cumulative likelihood ratio, As the decision threshold, The closer the value is to 1, the more reliable the decision. The asymmetric adaptive step-size update of user reputation value includes at least the following steps: when Above the threshold Update according to the following rules: Reward, report correct: (3) Punishment, report error: (4) in, and They are positive and negative step sizes, respectively, and With continuous error counter Positive correlation enables severe penalties for persistent malicious behavior; A reputation update is triggered only when the decision confidence level exceeds a certain threshold. Enhance the credibility of users whose reports align with the final decisions; For users who report inconsistent information, their reputation will be lowered, and the reduction will increase significantly with the number of consecutive errors made by the user.

4. The adaptive Byzantine attack defense method based on multi-mechanism collaboration according to claim 3, characterized in that: The soft weight mapping mechanism smoothly and non-linearly maps reputation values ​​to fusion weights; The soft weight mapping mechanism includes at least the following steps: First, normalization is performed, standardizing the reputation scores of all users to a standard range: (5) in It is a small positive number used to prevent division by zero errors; Then, a nonlinear mapping is performed. Through an S-shaped nonlinear function, the normalized reputation value is smoothly mapped to a basic weight. The S-shaped nonlinear function shows that the weight of users with medium reputation values ​​changes sensitively, while the weight of users with extremely high or low reputation values ​​changes slowly. (6) in, The slope factor controls the "softness" of the mapping; Finally, state decay is performed. Based on the user's current state label given by the lag state machine, the basic weights are decayed by different proportions. (7) in, These are the state labels output by the hysteresis state machine. This represents the corresponding attenuation coefficient.

5. The adaptive Byzantine attack defense method based on multi-mechanism collaboration according to claim 1, characterized in that: The S5 includes at least the following steps: First, dynamic thresholds are set. Based on the average and dispersion of all users' reputation scores, an upper and lower threshold for a category are dynamically calculated: (8) (9) in, and The mean and standard deviation of the reputation score. It is a scaling factor that controls the deviation of the threshold from the average value; Then, hysteresis counting is performed, maintaining two counters for each user: one to record the number of frames in which their reputation value is consistently above the upper threshold, and the other to record the number of frames in which their reputation value is consistently below the lower threshold. When the user's credit score is consistently higher than or below When this occurs, the corresponding "rise counter" or "fall counter" increases; Finally, the status transition is performed. The user's classification status will not change immediately due to a one-time fluctuation in their reputation value. Only when the corresponding counter confirms that their reputation value has continuously deviated from the threshold for a preset confirmation period will the system perform a status upgrade or downgrade.

6. The adaptive Byzantine attack defense method based on multi-mechanism collaboration according to claim 1, characterized in that: S6 includes at least the following steps: The fusion center first sorts all user reports in descending order of weight based on the weights calculated in step S4. Within each perception cycle, the fusion center, based on calculated weights, assigns users according to their weights. Arrange in descending order to obtain the sequence ; Then, the weighted evidence value of each report is calculated sequentially in this order and added to the global accumulated evidence volume for sequential fusion. Following this order, each user's report is converted into weighted evidence and accumulated to the global evidence volume. Finally, an early termination decision is made; after each user's report is integrated, it is determined whether the accumulated amount of evidence has exceeded the decision threshold. When the threshold is reached, a global decision is made immediately and the fusion is terminated; If the threshold is not reached, the maximum possible range of evidence that the remaining unfused reports can provide will be dynamically estimated. If this range indicates that even merging all remaining users would not change the current preliminary conclusions, the system will stop merging early and make a final decision based on the current trend of evidence.

Images