A Survey of Automatic Software Vulnerability Detection, Program Repair, and Defect Prediction Techniques

Table 1

Summary of deep learning technologies on automatic vulnerability detection, automatic program patching, and automatic defect prediction.

	Method type	Advantage	Disadvantage
Automatic vulnerability detection	Code similarity-based vulnerability detection	Source code-based detects multiple clone types; binary code-based achieves higher detection accuracy	False negative rate high (source); analytical complexity (binary)
	Code pattern-based vulnerability detection	Static method achieves higher code; coverage dynamic method detects faster	Lack run-time information (static); low code coverage (dynamic)
Automatic program patching	Grammar-based program patching	Token-based method error analysis; simple text-based method generates higher quality patches	Poor patches interpretability (token); unacceptable program behavior (tex)
	Semantic-based program patching	Static method achieves good repair effect dynamic method and accurately captures program behavior	Limited type of repair (static); high cost (dynamic)
Automatic defect prediction	Within-project defect prediction	End-to-end implementation, accurate prediction of defective program modules	Poor expansion
	Crossproject defect prediction	Effectively integrates dataset resources to better promote new project development practices	Excessive code feature extraction granularity
	Just-in-time defect prediction	Earlier identifies faulty modules and fine-grained analysis, efficiently identifies the number of defects	Lack of extensive training data to train the model