Linux kernel vulnerability mining method based on diversity guidance
By adopting a diversity-guided kernel vulnerability discovery method, the problem of locality of exploration caused by insufficient seed diversity is solved, the code coverage and vulnerability discovery efficiency of Linux kernel fuzzing are improved, and more efficient kernel vulnerability discovery is achieved.
Patent Information
- Application Number
- CN202512003270.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-29
- Publication Date
- 2026-04-24
AI Technical Summary
Existing Linux kernel fuzzing tools suffer from insufficient seed diversity, leading to the problem of locality of exploration during testing. This limits the upper limit of code coverage growth and reduces the probability of discovering deep-seated vulnerabilities.
By constructing a kernel vulnerability mining method based on diversity guidance, clustering PoC seeds using the Louvain community segmentation algorithm, introducing the community popularity rate (CPR) index, designing a two-layer multi-armed slot machine scheduling framework and a CPR-guided seed mutation strategy, prioritizing mutation and expansion of highly diverse seeds, and constructing a diversity-aware hierarchical seed scheduling framework.
It significantly improved code coverage and vulnerability discovery capabilities. Experimental results showed that coverage increased by 17.4%, the number of vulnerabilities discovered increased by an average of 9.1 times, and 31 new vulnerabilities were successfully discovered, 11 of which have been confirmed by the Linux kernel community.
Smart Images

Figure CN121919879A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer software testing technology, specifically relating to a method for discovering vulnerabilities in the Linux operating system kernel. Background Technology
[0002] The Linux kernel is the most widely deployed operating system kernel in modern computing infrastructure, supporting a wide range of critical systems, from cloud servers to industrial control systems and autonomous driving systems. As the cornerstone of the entire system, any vulnerability in the kernel can not only lead to kernel crashes or privilege escalation attacks, but also cascade effects on all dependent applications, potentially causing irreversible damage. For example, resource management vulnerabilities in the kernel can cause performance degradation or make user applications unusable; memory out-of-bounds access or privilege bypass vulnerabilities can be exploited to leak sensitive data from servers; and in high-reliability fields such as autonomous driving and smart manufacturing, kernel vulnerabilities can pose a direct threat to human safety.
[0003] Fuzzing is a mainstream automated vulnerability discovery technique, and Syzkaller, open-sourced by Google, is widely recognized as the most successful Linux kernel fuzzing tool. Syzkaller employs a coverage-driven strategy, exploring kernel code by generating, scheduling, and mutating system call sequences to find potential vulnerabilities.
[0004] However, existing kernel fuzzing tools, such as Syzkaller, still face bottlenecks in practical applications: insufficient seed diversity. This problem stems from three key stages in the testing process: initial seed generation, seed scheduling, and seed mutation. In the early stages of fuzzing, the generated seeds can cover different code regions; however, as the test progresses, the seed corpus gradually converges, and newly generated seeds become highly similar to existing seeds, causing the fuzzer to repeatedly explore known code paths. This phenomenon is known as "exploratory locality," ultimately limiting the upper limit of code coverage growth and reducing the probability of discovering deep-seated vulnerabilities. Summary of the Invention
[0005] The purpose of this invention is to solve the testing bottleneck problem caused by insufficient seed diversity in existing kernel fuzzing techniques, and to provide a method for discovering Linux operating system kernel vulnerabilities based on diverse boot methods.
[0006] This invention proposes a Linux operating system kernel vulnerability mining method based on diversity guidance. First, it represents the collected Proof-of-Concept (PoC) tests (test cases that can trigger historical vulnerabilities) using a custom Abstract Syntax Tree (AST), and clusters them based on the Louvain community discovery algorithm to construct an initial diversity seed library, dividing the seeds into several functionally distinct communities. To quantify seed diversity, a Community Popularity Rate (CPR) metric is introduced. Based on CPR, a two-layer multi-armed bandit scheduling framework (using the Upper Confidence Bound, UCB strategy) is designed to efficiently allocate mutation resources between and within communities. Furthermore, a CPR-guided seed mutation strategy is proposed to prioritize the rapid mutation and expansion of highly diverse seeds, thereby improving the coverage speed and efficiency of vulnerability triggering. The specific steps are as follows:
[0007] (I) Construction and division of diversified initial corpora;
[0008] This invention abandons the traditional method of randomly generating initial seeds, and instead uses proof-of-concept (PoC) seeds that have historically triggered vulnerabilities as high-quality initial seeds. By calculating the semantic similarity between these seeds and using the Louvain community partitioning algorithm, they are clustered into multiple "seed communities," laying the foundation for subsequent diversity quantification and guidance.
[0009] (ii) Hierarchical seed scheduling based on diversity perception;
[0010] This study introduces a novel diversity metric—Community Popularity Rate (CPR)—and designs a CPR-aware hierarchical multi-armed bandit (MAB) scheduling algorithm. This algorithm uses code coverage and CPR as a combined reward signal, prioritizing the scheduling of seed communities and seeds within those communities with high diversity potential, thereby balancing the "utilization" of high-coverage paths with the "exploration" of rare paths.
[0011] (iii) Seed variation guided by diversity;
[0012] This study designs a CPR-guided adaptive mutation strategy. For seeds belonging to high-CPR communities, a higher mutation frequency is dynamically assigned to quickly fill the sparse semantic space; otherwise, conventional mutation is used. This strategy aims to efficiently break through the boundaries of existing communities and even generate entirely new seed communities.
[0013] Furthermore:
[0014] The specific process for constructing and dividing the diverse initial corpus described in step (I) is as follows:
[0015] 1. PoC-based corpus construction: Collecting and organizing PoC seeds corresponding to confirmed and patched vulnerabilities from publicly available vulnerability reporting platforms. These PoCs themselves contain valuable semantic information that triggers complex kernel behaviors, serving as a high-quality source of initial seeds.
[0016] 2. Similarity Calculation Based on Abstract Syntax Tree (AST): To accurately quantify the semantic associations between seeds and thus provide an objective basis for subsequent clustering, this invention needs to calculate the similarity between seeds. This invention achieves this by parsing each seed into a customized Abstract Syntax Tree (AST) and calculating the normalized tree edit distance between two ASTs. The specific calculation method is defined by the following formula (1):
[0017] (1);
[0018] in, and This represents the AST structure of each of the two seeds. This is the general Transform into The minimum number of edit operations required, i.e., the tree edit distance, These represent the total number of nodes in the two trees, respectively. The result of this formula is a normalized value between 0 and 1, i.e., the similarity score. This score will serve as the edge weights in the subsequent steps of constructing the weighted graph, directly determining the seed community partitioning result. A high score indicates that the two seeds are semantically highly related and are more likely to be assigned to the same community.
[0019] The hierarchical seed scheduling based on diversity awareness described in step (II) specifically includes:
[0020] This paper addresses the locality problem caused by traditional scheduling strategies by modeling seed scheduling as a hierarchical multi-armed slot machine (MAB) problem. The core of the MAB framework is to strike a balance between "exploration" and "exploitation," with one representative solution being the upper confidence bound (UCB) algorithm. The UCB algorithm aims to dynamically balance "exploitation" of known efficient seeds and communities with "exploration" of underexplored seeds and communities to avoid prematurely falling into local optima. The UCB decision is based on a score calculated using the following formula (2):
[0021] (2);
[0022] In the t-th round of decision-making, i represents a candidate arm, that is, a community or a seed; This is the average return of arm i that was previously selected; this is the "utilization term". It is the total number of times arm i has been selected previously. This is the "exploration item," which provides a bonus to the arm that is selected less frequently. In each round of scheduling, the system selects the arm with the highest UCB score, thus achieving a dynamic balance between utilization and exploration. The specific process is as follows:
[0023] 1. Definitions of Community Density and Community Popularity (CPR):
[0024] To establish a dynamic indicator that can measure the diversity potential of a community, the relative size of the community, i.e., community density, is first calculated. Community density is defined by formula (3):
[0025] (3);
[0026] in, It is a community The number of seeds contained at time t. Based on the seed density within the community, this invention further proposes a core diversity indicator—Community Popularity (CPR), defined by formula (4):
[0027] (4);
[0028] in, It is a community The total number of times a community is scheduled before time t. The CPR metric combines the frequency of attention a community receives with its size. A high CPR indicates that a community is currently underexplored relative to its size and has greater potential for diversity. This metric is a core input for subsequent community-level scheduling decisions.
[0029] 2. Community-level scheduling:
[0030] When choosing which community to explore, a comprehensive reward function is needed to evaluate the community's value. This function should consider both coverage gains and diversity potential. (Community single-round scheduling reward) Calculated using formula (5):
[0031] (5);
[0032] Among them, the first item The second item is the average new code coverage brought by seeds within the community. This is the community's CPR value, determined by weighting factors. Adjust its impact. Average reward of the community. This is obtained by accumulating the historical single-round rewards and taking the average, as shown in formula (6):
[0033] (6);
[0034] in, yes The selected set of rounds. The calculated average reward. This will be used as the "utilization term" in formula (2). Substituting these values into the final community-level UCB score calculation formula (7), the optimal community for this round is selected:
[0035] (7)
[0036] 3. Seed-level scheduling:
[0037] After determining the optimal community for this round of scheduling, the most efficient seed within that community needs to be selected for mutation to ensure that its derived seeds have the ability to discover more new code coverage paths. The single-round reward for the seed... Defined as the average increase in new coverage per unit time resulting from its mutated offspring, it is calculated by formula (8):
[0038] (8);
[0039] in, From seeds The set of all offspring seeds obtained from the mutation. It's new coverage. It is the execution time. This is the crash identifier during this round of testing. Similarly, the average reward of the seed... It is obtained by accumulating the historical single-round rewards and taking the average, as shown in formula (9):
[0040] (9);
[0041] The average reward This will be used as the "utilization term" in formula (2). Substituting this into the final seed-level UCB score calculation formula (10), the best seed for this round is selected within the community for mutation.
[0042] (10);
[0043] Figure 1 The hierarchical decision-making process within the diversity-aware seed scheduling module is demonstrated. First, the optimal community for this round is selected. Then, the optimal seed within that community is selected based on the UCB score to complete the scheduling. Subsequently, the seed is executed, and the reward information for that seed is recorded and updated.
[0044] The specific process for the diversity-guided seed mutation described in step (iii) is as follows:
[0045] 1. CPR-guided adaptive mutation
[0046] In Syzkaller, low-frequency, single-generation regular mutation and high-frequency, batch-generation Smash mutation are considered as two optional strategies. Before mutating a selected seed, the CPR value of its community is first evaluated to rank it among all communities. If the ranking is high, such as in the top two, the community is considered to have extremely high diversity potential, and a high-intensity Smash mutation strategy is applied; otherwise, a regular mutation strategy is used.
[0047] 2. Determining the affiliation of new seed communities
[0048] A seed generated through mutation is considered valuable when it triggers new code overwriting or causes a system crash. In this case, the AST similarity between the new seed and all existing communities is calculated. If its maximum similarity with a community exceeds a preset threshold, the new seed is added to that community, effectively filling that community. If its similarity with all communities is below the threshold, the seed is considered to have explored a completely new semantic region, and a new community is created for it.
[0049] Figure 2 It details the diversity-guided seed mutation module and the subsequent corpus update logic, demonstrating how to select mutation strategies based on CPR ranking and how to handle valuable new seeds.
[0050] Compared with the prior art, the present invention has three main advantages.
[0051] 1. Significantly Improved Code Coverage and Vulnerability Discovery Capabilities: Experimental results show that compared with state-of-the-art kernel fuzzing tools such as Syzkaller, this method improves code coverage by an average of 17.4% and the number of vulnerabilities discovered by an average of 9.1 times. During a two-month test, this method successfully discovered 31 new Linux kernel vulnerabilities, 11 of which have been officially confirmed by the Linux kernel community.
[0052] 2. Effectively solves the problem of seed locality: This invention uses a three-pronged strategy of diverse initial corpus construction, diversity-aware seed scheduling, and diversity-guided seed mutation to ensure that the fuzzing process can continuously explore new, under-covered kernel functional regions, fundamentally solving the problem of seeds converging to local optima.
[0053] 3. A systematic diversity optimization framework is provided: This invention decomposes the seed lifecycle into three stages for the first time and designs corresponding diversity enhancement strategies for each stage. This framework, which combines seed clustering into communities, quantized scheduling, and adaptive mutation, provides new ideas for the future development of fuzz testing technology and can be extended to scenarios such as intelligent vehicles, industrial control, and cloud operating systems. Attached Figure Description
[0054] Figure 1 This is a hierarchical seed scheduling framework based on diversity perception.
[0055] Figure 2 This is a diversity-guided seed variation framework.
[0056] Figure 3 This is a schematic diagram of the method execution flow. Detailed Implementation
[0057] The overall execution flow of the Linux kernel vulnerability discovery method proposed in this invention is based on "seed diversity guidance". Through a four-stage closed-loop framework of "preparation-initialization-iterative fuzzing-termination", it combines AST similarity clustering, CPR index quantification, hierarchical UCB scheduling and adaptive mutation depth to continuously balance the "exploration" and "exploitation" of fuzzing, break through the exploration locality bottleneck of traditional tools, and improve code coverage and vulnerability discovery efficiency. The key points include: 1. By considering the low value problem caused by the randomness of traditional initial seeds, an initial corpus is constructed using historical real vulnerability PoC seeds, and seed communities are divided based on AST similarity and the Louvain algorithm, laying the foundation for diversity quantification and scheduling; 2. A community-seed two-level UCB scheduling strategy is designed based on the hierarchical multi-armed slot machine (MAB) framework, and reward signals are calculated by combining code coverage and CPR indicators, prioritizing communities and seeds with high diversity potential, and dynamically balancing path exploration and utilization; 3. A CPR-guided adaptive mutation mechanism is designed, selecting high-frequency Smash mutation or regular mutation based on the CPR ranking of the seed's community, and determining the community affiliation of new seeds through AST similarity, continuously expanding the test coverage and updating the corpus status to form an iterative optimization closed loop.
[0058] The overall execution flow of the method of this invention is as follows: Figure 3 As shown, it includes the following stages.
[0059] 1. Preparation phase: Collect a set of PoC seeds that have historically triggered kernel vulnerabilities as the initial seed corpus.
[0060] 2. Initialization Phase: For each seed in the corpus, its system call sequence is parsed into an AST. Based on the AST, the similarity between all seed pairs is calculated, and a similarity matrix is constructed. The Louvain algorithm is used to partition the similarity matrix into communities, obtaining the initial seed community set.
[0061] 3. Iterative Fuzz Testing Phase: Enter the main loop and continue executing until the preset time budget or number of iterations is reached. Each iteration performs the following operations:
[0062] 3.1 Community Selection: Calculate the average reward of each community based on its current status, and then calculate the UCB score for each community. Select the community with the highest UCB score as the optimal community for this round.
[0063] 3.2 Seed Selection: Within the selected optimal community, the average reward is calculated based on the historical performance of each seed, and then its UCB score is calculated. The seed with the highest UCB score within the community is selected as the mutation target for this iteration.
[0064] 3.3 Adaptive Mutation: Obtain the CPR value of the community to which the mutated object belongs in this iteration, and calculate its ranking among all communities to determine the diversity potential of the seed. If the seed's CPR ranking is higher than a preset threshold... r If the seed is positive, perform a high-intensity Smash mutation on it to generate multiple offspring seeds at once, maximizing its diversity potential; otherwise, perform a regular mutation to generate a single offspring seed.
[0065] 3.4 Execution and Feedback: Execute one or more new seeds generated by the mutation process, monitor their execution, and collect code coverage information and any crash reports.
[0066] 3.5 Corpus Update: For any new seed that has been discovered or triggered a crash, calculate the AST similarity between the seed and all existing communities. If the maximum similarity is greater than the threshold, the corpus is updated accordingly. c If the maximum similarity is less than the threshold, then add the seed to the most similar community. c Then, using this seed as the initial member, a new community is created.
[0067] 3.6 Status Update: Update the statistics of all affected communities and seeds, such as scheduling times, density, coverage gain, etc., to provide a basis for the next round of scheduling decisions.
[0068] 4. Termination Phase: The loop ends when the fuzzing reaches the set termination conditions. All unique crash reports discovered during the testing process are compiled and output; these reports represent potential kernel vulnerabilities for subsequent manual analysis and verification.
[0069] Experimental verification
[0070] We compared SyzDiversity with four state-of-the-art kernel fuzzing tools using two sets of seed corpora and two versions of the Linux kernel.
[0071] (1) Selection of state-of-the-art tools: Specifically, four kernel fuzzing tools were selected, namely Syzkaller[1] (the de facto standard and the basis for most subsequent work), Healer[2] (which generates semantically meaningful seeds through system call dependency analysis, serving as a baseline for effective comparison with the initial seeds built in SyzDiversity phases 1-2), SyzVegas[3] (which also uses a MAB scheduler similar to SyzDiversity) and SyzGPT[4] (the latest kernel fuzzing tool).
[0072] (2) Initial Seed Settings: By default, state-of-the-art fuzzing tools such as Syzkaller, SyzVegas, and SyzGPT do not include initial seeds. To more accurately evaluate the impact of different initial seeds on fuzzing effectiveness, as well as the overall effectiveness of SyzDiversity, we used two seed corpora of the same size. The first is 22,506 initial seeds generated by Healer (see Table 1), and the second is 22,506 initial seeds collected and segmented by SyzDiversity (see Table 2).
[0073] (3) Experimental environment configuration: We used the long-standing Linux kernel, namely Linux v5.15, and the latest Linux kernel (i.e., Linux v6.14), which are widely deployed around the world and have been rigorously tested by many previous works. To mitigate the effects of inherent randomness, each experiment was repeated three times, and each experiment had a time budget of 72 hours. All kernels were compiled under the same configuration with common features enabled (e.g., KCOV[4] and KASAN[5]). All experiments were conducted on a machine running Ubuntu 20.04 with 48 CPUs and 64GB of memory, equipped with an AMD EPYC 9754128-core processor.
[0074] (4) Experimental results:
[0075] A comprehensive evaluation was conducted on two different initial seeds (detailed experimental data are shown in Tables 1 and 2 below), using basic block coverage and the number of bugs (including the number of known bugs and unknown new bugs) as the main evaluation metrics. The results show that in both sets of experiments, SyzDiversity improved coverage by an average of 17.4% and increased the number of detected bugs by an average of 9.1 times, significantly outperforming state-of-the-art fuzzing tools.
[0076] Table 1: Experimental Results of the Effectiveness of Healer Initial Seed Fuzzy Test
[0077] .
[0078] Table 2: Experimental Results of the Effectiveness of the Initial Seed Fuzz Test in SyzDiversity
[0079] .
Claims
1. A method for discovering Linux kernel vulnerabilities based on diverse boot methods, characterized in that, First, the collected PoCs are represented using a custom Abstract Syntax Tree (AST) and clustered based on the Louvain community discovery algorithm to construct an initial diversity seed library, dividing the seeds into several functionally distinct communities. To quantify seed diversity, a Community Popularity Rate (CPR) metric is introduced. Based on CPR, a two-layer multi-armed slot machine (MAB) scheduling framework is designed for efficient allocation of mutation resources between and within communities. A CPR-guided seed mutation strategy is adopted to prioritize the rapid mutation and expansion of highly diverse seeds, thereby improving the coverage speed and efficiency of vulnerability triggering. The specific steps are as follows: (I) Construction and division of diversified initial corpora; We used proof-of-concept (PoC) seeds that had historically triggered vulnerabilities as high-quality initial seeds; by calculating the semantic similarity between these seeds and using the Louvain community partitioning algorithm, we clustered them into multiple "seed communities," laying the foundation for subsequent diversity quantification and guidance. (ii) Hierarchical seed scheduling based on diversity perception; Among them, a new diversity metric, Community Popularity (CPR), is introduced, and a CPR-aware hierarchical multi-armed slot machine (MAB) scheduling algorithm is designed. This algorithm uses code coverage and CPR as a comprehensive reward signal, and prioritizes scheduling seed communities and seeds within those communities with high diversity potential, thereby balancing the "utilization" of high-coverage paths and the "exploration" of rare paths. (iii) Seed variation guided by diversity; Among them, a CPR-guided adaptive mutation strategy is designed; for seeds belonging to high CPR value communities, a higher mutation frequency is dynamically allocated to quickly fill the sparse semantic space; otherwise, regular mutation is used, aiming to efficiently break through the boundaries of existing communities and even generate entirely new seed communities.
2. The Linux kernel vulnerability discovery method according to claim 1, characterized in that, The specific process for constructing and dividing the diverse initial corpus described in step (I) is as follows: (1) Construction of a corpus based on PoC: collect and organize PoC seeds corresponding to confirmed and fixed vulnerabilities from public vulnerability reporting platforms; (2) Similarity calculation based on Abstract Syntax Tree (AST): By parsing each seed into a custom Abstract Syntax Tree (AST), and calculating the normalized tree edit distance between two ASTs, specifically defined by the following formula (1): (1); in, and Represents the AST structure of each of the two seeds. This is the general Transform into The minimum number of edit operations required, i.e., the tree edit distance, These are the total number of nodes in the two trees, respectively. The result of this formula is a normalized value between 0 and 1, which is the similarity score. This score will be used as the edge weight in the subsequent steps of constructing the weighted graph, directly determining the division result of the seed community. A high score indicates that the two seeds are highly related semantically and are more likely to be classified into the same community.
3. The Linux kernel vulnerability discovery method according to claim 2, characterized in that, The hierarchical seed scheduling based on diversity awareness described in step (II) specifically includes: The upper confidence bound (UCB) algorithm is adopted to dynamically balance the "utilization" of known efficient seeds and communities with the "exploration" of underexplored seeds and communities, so as to avoid premature local optima in testing; the UCB decision is based on the score calculated by the following formula (2): (2); In the t-th round of decision-making, i represents a candidate arm, that is, a community or a seed; It is the average return of arm i previously selected; this is the "utilization term". It is the total number of times arm i has been selected previously. It is an "exploration item" that gives a bonus to the arm that is selected less often; in each round of scheduling, the arm with the highest UCB score will be selected, thus achieving a dynamic balance between utilization and exploration.
4. The Linux kernel vulnerability discovery method according to claim 3, characterized in that, The hierarchical seed scheduling based on diversity perception described in step (II) is as follows: (1) Definition of Community Density and Community Popularity (CPR): To establish a dynamic indicator that can measure the diversity potential of a community, the relative size of the community, i.e., the community density, is first calculated. The community density is defined by formula (3): (3); in, It is a community Based on the seed density within the community, and considering the number of seeds contained at time t, a core diversity indicator—Community Popularity (CPR)—is further proposed, defined by formula (4): (4); in, It is a community The total number of times a community is scheduled before time t; the CPR metric combines the frequency of attention a community receives with its own size; this metric is the core input for subsequent community-level scheduling decisions. (2) Community-level scheduling: When choosing which community to explore, a comprehensive reward function is needed to evaluate the community's value, taking into account both coverage gains and diversity potential; the community's single-round scheduling reward. Calculated using formula (5): (5); Among them, the first item The second item is the average new code coverage brought by seeds within the community. This is the community's CPR value, determined by weighting factors. Moderating its impact; average reward of the community This is obtained by accumulating the historical single-round rewards and taking the average, as shown in formula (6): (6); in, yes The selected set of rounds; the calculated average reward This will be used as the "utilization term" in formula (2). Substituting these values into the final community-level UCB score calculation formula (7), the optimal community for this round is selected: (7); (3) Seed-level scheduling: After determining the optimal community for this round of scheduling, the most efficient seed within the community is selected for mutation to ensure that its derived seeds have the ability to discover more new code coverage paths; the single-round reward for the seed. Defined as the average increase in new coverage per unit time resulting from its mutated offspring, it is calculated by formula (8): (8); in, From seeds The set of all offspring seeds obtained from the mutation. It's new coverage. It is the execution time. This is the crash identifier during this round of testing; similarly, it's the average reward of the seed. It is obtained by accumulating the historical single-round rewards and taking the average, as shown in formula (9): (9); The average reward This will be used as the "utilization term" in formula (2). Substituting this into the final seed-level UCB score calculation formula (10), the best seed for this round is selected within the community for mutation: (10)。 5. The Linux kernel vulnerability discovery method according to claim 4, characterized in that, The specific process for the diversity-guided seed mutation described in step (iii) is as follows: (1) CPR-guided adaptive mutation; In Syzkaller, low-frequency, single-generation new seed regular mutation and high-frequency, batch-generation new seed Smash mutation are used as two optional strategies. Before mutating a selected seed, the CPR value of its community is first evaluated to rank among all communities. If the ranking is high, the community is considered to have extremely high diversity potential, and a high-intensity Smash mutation strategy is adopted for it. Conversely, a conventional mutation strategy is adopted; (2) Determination of affiliation to new seed communities; When a new seed generated through mutation triggers new code coverage or causes a system crash, it is considered a valuable seed. At this time, the AST similarity between the new seed and all existing communities is calculated. If its maximum similarity with a certain community exceeds a preset threshold, the new seed is added to that community to effectively fill the community. If its similarity with all communities is below the threshold, it is considered that the seed has explored a completely new semantic region, and a new community is created for it.
6. The Linux kernel vulnerability discovery method according to claim 5, characterized in that, The overall execution process includes the following four stages: (1) Preparation phase: Collect a set of PoC seeds that have historically triggered kernel vulnerabilities as the initial seed corpus; (2) Initialization phase: For each seed in the corpus, its system call sequence is parsed into an AST; The similarity between all seed pairs is calculated based on the AST, and a similarity matrix is constructed. The Louvain algorithm is used to divide the similarity matrix into communities to obtain an initial seed community set; (3) Iterative fuzz testing phase: Enter the main loop and continue to execute until the preset time budget or number of iterations is reached; Perform the following operations in each iteration: (3.1) Community selection: Calculate the average reward of the community based on the current status of all communities, and then calculate the UCB score of each community based on this. Select the community with the highest UCB score as the optimal community for this round. (3.2) Seed selection: Within the selected optimal community, calculate the average reward based on the historical performance of each seed, and then calculate its UCB score; select the seed with the highest UCB score in the community as the mutation object for this round of iteration; (3.3) Adaptive mutation: Obtain the CPR value of the community to which the mutated object belongs in the current iteration, and calculate its ranking in all communities to determine the diversity potential of the seed; if the CPR ranking of the seed is higher than the preset threshold, perform high-intensity Smash mutation on the seed to generate multiple offspring seeds at once to maximize its diversity potential. Otherwise, perform regular mutation to generate a single offspring seed; (3.4) Execution and Feedback: Execute one or more new seeds generated by the mutation process, monitor their execution, and collect code coverage information and any crash reports; (3.5) Corpus update: For any new seed that has been discovered to cover or has triggered a crash, calculate the AST similarity between the seed and all existing communities. If the maximum similarity is greater than the threshold, the seed is added to the most similar community. If the maximum similarity is less than the threshold, a new community is created with the seed as the initial member. (3.6) Status update: Update the statistics of all affected communities and seeds, including scheduling times, density, and coverage gain, to provide a basis for scheduling decisions in the next round; (4) Termination phase: When the fuzzing reaches the set termination condition, the loop ends; all unique crash reports found during the testing process are compiled and output. These reports are potential kernel vulnerabilities for subsequent manual analysis and verification.
