**2. Related Work**

At present, researchers have applied fuzzy testing to different types of vulnerability mining. Lin et al. [9] proposed a priority-based path searching method (PBPS) to utilize the capability of concolic execution better. Peng et al. [10] proposed Angora, a new mutationbased fuzzer, and proved that Angora has better performance than other fuzzing tools. Wang et al. [8] used a neural network to guide the sample generation of fuzzy testing

and proposed a solution called NeuFuzz. NeuFuzz has a very significant performance in the vulnerability mining of the critical execution path of the program. Zhang et al. [11] summarized existing fuzzy testing technologies and use case generation technologies of fuzzy testing. Zhang el al. [12] proposed an algorithm of sensitive region prediction based on a neural network and improved the detection efficiency and detection depth through the incremental learning method of sensitive areas. Combining fuzzy testing and symbolic execution, Xie [13] proposed a hybrid testing method based on a branch coverage called AFLeer. Xu et al. [14] applied a recurrent neural network to fuzzy testing sample generation. Luca et al. [2] designed a novel concolic executor to improve the efficiency of concolic execution and investigate whether techniques borrowed from the fuzzing domain can be used to solve the symbolic query problem. Stefan [15] proposed the notion of coverage-guided tracing to improve the efficiency of code coverage guided fuzzy testing. Yang et al. [16] proposed a novel programmable fuzzy testing framework. Developers only need to write a small number of fuzzy testing guidance programs to implement customized fuzzy testing. Patrice et al. [17] proposed learn&fuzz, which used a learned input probability distribution to intelligently guide fuzzing inputs. Li et al. [18] proposed symfuzz, which is a method combining directed fuzzy testing technology with selective symbolic execution technology and can realize vulnerability detection under complex path conditions. Liang et al. [19] proposed a machine-learning-based framework to improve the quality of seed inputs for fuzzing programs. Zou et al. [1] described the development from traditional automation to intelligent vulnerability mining in software vulnerability mining. This paper also pointed out that the application of traditional machine learning technology in the vulnerability mining field still has limitations. Ma et al. [20] proposed the optimization strategy of sample set reduction in the fuzzy process, including approximation algorithm. Cornelius [21] proposed IJON, an annotation mechanism that a human analyst can used to guide the fuzzer.

In the experimental part, this paper compares the simplification and efficiency of sample set between heuristic genetic algorithm and approximation algorithm. He et al. [22] proposed a tool called VCCFinder to find potential vulnerabilities. Nick et al. [23] used mined vulnerabilities by utilizing a code attribute graph for fuzzy testing. She et al. [24] proposed a novel program smoothing technique using a surrogate neural network models to achieve higher edge coverage and improve the ability of finding new bugs. Chen et al. [25] proposed POLYGLOT, a genetic fuzzing framework that generates high-quality test cases for exploring processors of different programing languages. Huang et al. [4] proposed PANGOLIN, an approach based on polyhedral path abstraction, which preserves the exploration state in the concolic execution stage and allows more effective mutation and constraint solving over existing techniques. Zhang et al. [26] proposed a novel incremental and stochastic rewriting technique STOCHFUZZ that piggy-backs on the fuzzing procedure. Liang et al. [3] presented DeepFuzzer, an enhanced greybox fuzzer with qualified seed generation, balanced seed selection and hybrid seed mutation. Chen et al. [7] proposed an ensemble fuzzing method, EnFuzz. Enfuzz contains many different heuristic genetic algorithms and achieves a better performance in terms of path coverage, branch coverage and bug discovery. The idea of ensemble is also similar to the CVDF DYNAMIC proposed in this paper. Yue et al. [27] presented a variant of the adversarial multi-armed bandit model for modeling AFL's power schedule process named EcoFuzz, which can effectively regulate seed energy in fuzzy testing. Zong et al. [28] proposed FuzzGuard, a deep-learning-based approach to predict the reachability of inputs and further improve the performance of DGF. Gan [5] proposed a data flow sensitive fuzzing solution GREYONE, which can further improve the performance of data flow analysis, and the experiments show that GREYONE has better performance than the existing fuzzy testing tools such as AFL. Sebastian et al. [29] proposed ParmeSan, a sanitizer guided fuzzing method to solve the low-bug coverage problem. Oleksii et al. [30] proposed specfuzz, which is a novel fuzzy testing method, which can be used to detect speculative execution vulnerabilities including spectre and outof-order execution vulnerabilities. Compared with the traditional static analysis method,

specfuzz has further improved the analysis accuracy. Lee et al. [31] proposed a constraintguided directed greybox fuzzing method, which aims to satisfy a sequence of constraints rather than merely reaching a set of target sites. Christopher et al. [32] proposed a brandnew token-level fuzzing method. Different from the fuzzy method based on data flow or seed energy regulation, token-level fuzzing applies mutations at the token level instead of applying mutations either at the byte level or at the grammar level. The authors found many unknown bugs through the token-level fuzzing method on popular javascript engines. In recent years, the safety of deep learning technology has also attracted the attention of scholars. It is possible for attackers to deduce the sensitive training data of engineering through the unsafe deep learning model. Ximeng Liu et al. [33] briefly introduced four different types of attacks in deep learning, reviewed and summarized the security defense measures of deep learning attack methods and further discussed the remaining challenges and privacy issues of deep learning security. MB mollah et al. [34] proposed an efficient data-sharing scheme, which allows smart devices to share secure data with others at the edge of cloud-assisted IOT.
