Browse Theses

No single vulnerability detection technique can identify all vulnerabilities in a software system. However, the vulnerabilities that are identified from a detection technique may be predictive of the residuals. We focus on creating and evaluating statistical models that predict the components that contain the highest risk residual vulnerabilities.

The cost to find and fix faults grows with time in the software life cycle (SLC). A challenge with our statistical models is to make the predictions available early in the SLC to afford for cost-effective fortifications. Source code static analyzers (SCSA) are available during coding phase and are also capable of detecting code-level vulnerabilities. We use the code-level vulnerabilities identified by these tools to predict the presence of additional coding vulnerabilities and vulnerabilities associated with the design and operation of the software. The goal of this research is to reduce vulnerabilities from escaping into the field by incorporating source code static analysis warnings into statistical models that predict which components are most susceptible to attack.

The independent variable for our statistical model is the count of security-related source SCSA warnings. We also include the following metrics as independent variables in our models to determine if additional metrics are required to increase the accuracy of the model: non-security SCSA warnings, code churn and size, the count of faults found manually during development, and the measure of coupling between components. The dependent variable is the count of vulnerabilities reported by testing and those found in the field.

We evaluated our model on three commercial telecommunications software systems. Two case studies were performed at an anonymous vendor and the third case study was performed at Cisco Systems. Each system is a different technology and consists of over one million source lines of C/C++ code. The results show positive and statistically significant correlations between the metrics and vulnerability counts. Additionally, the predictive models produce accurate probability rankings that indicate which components are most susceptible to attack. The models are evaluated with receiver operating characteristic curves where each case study showed over 92% of the area was under the curve. We also performed five-fold cross-validation to further demonstrate statistical confidence in the models. Based on these results we contribute the following theory:

Theory . Above a statistically determined threshold, SCSA vulnerability warnings are in the same components as vulnerabilities that are likely to be exploited.

Components that contain security-related warnings identified by SCSA are also likely to contain other exploitable vulnerabilities. Software engineers should systematically inspect and test code for other vulnerabilities when a security-related warning is present. Fortifying these vulnerabilities may facilitate other techniques to identify more undetected vulnerabilities.