Risk analysis in software design

Traditional terminology

Example risk-analysis methodologies for software usually fall into two basic categories: commercial (including Microsoft’s STRIDE, Sun’s ACSM/SAR, Insight’s CRAMM, and Synopsys’ SQM) and standards-based (from the National Institute of Standards and Technology’s ASSET or the Software Engineering Institute’s OCTAVE). An in-depth analysis of all existing methodologies is beyond our scope, but we’ll look at basic approaches, common features, strengths, weaknesses, and relative advantages and disadvantages.

As a corpus, “traditional” methodologies are varied and view risk from different perspectives. Examples of basic approaches include

• financial loss methodologies that seek to provide a loss figure to balance against the cost of implementing various controls;
• mathematically derived “risk ratings” that equate risk with arbitrary ratings for threat, probability, and impact; and
• qualitative assessment techniques that base risk assessment on anecdotal or knowledge-driven factors.

Each basic approach has its distinctly different merits, but they almost all share some valuable concepts that should be considered in any risk analysis. We can capture these commonalities in a set of basic definitions:

• The asset, or object of the protection efforts, can be a system component, data, or even a complete system.
• Risk, the probability that an asset will suffer an event of a given negative impact, is determined from various factors: the ease of executing an attack, the attacker’s motivation and resources, a system’s existing vulnerabilities, and the cost or impact in a particular business context.
• The threat, or danger source, is invariably the danger a malicious agent poses and that agent’s motivations (financial gain, prestige, and so on). Threats manifest themselves as direct attacks on system security.
• A vulnerability is a defect or weakness in system security procedure, design, implementation, or internal control that an attacker can compromise. It can exist in one or more of the components making up a system, even if those components aren’t necessarily involved with security functionality. A given system’s vulnerability data are usually compiled from a combination of OS- and application-level vulnerability test results, code reviews, and higher-level architectural reviews. Software vulnerabilities come in two basic flavors: flaws (design-level problems) or bugs (implementation-level problems). Automated scanners tend to focus on bugs, since human expertise is required for uncovering flaws.
• Countermeasures or safeguards are the management, operational, and technical controls prescribed for an information system that, taken together, adequately protect the system’s confidentiality, integrity, and availability as well as its information. For every risk, a designer can put controls in place that either prevent or (at a minimum) detect the risk when it triggers.
• The impact on the organization, were the risk to be realized, can be monetary or tied to reputation, or it might result in the breach of a law, regulation, or contract. Without a quantification of impact, technical vulnerability is hard to handle—especially when it comes to mitigation activities.
• Probability is the likelihood that a given event will be triggered. It is often expressed as a percentile, although in most cases, probability calculation is extremely rough.

Although they start with these basic definitions, risk methodologies usually diverge on how to arrive at specific values. Many methods calculate a nominal value for an information asset, for example, and attempt to determine risk as a function of loss and event probability. Others rely on checklists of threats and vulnerabilities to determine a basic risk measurement

Example of a risk calculation

One classic risk-analysis method expresses risk as a financial loss, or annualized loss expectancy, based on the following equation:

ALE = SLE × ARO,

where SLE is the single loss expectancy, and ARO is the annualized rate of occurrence (or the predicted frequency of a loss event happening).

Let’s consider an Internet-based equities trading application with a vulnerability that could result in unauthorized access (the implication being that unauthorized stock trades can be made). Assume a risk analysis determines that middle- and back-office procedures will catch and negate any malicious transaction such that the loss associated with the event is simply the cost of backing out of the trade. We’ll assign a cost of $150 for any such event, so SLE=$150. With an ARO of just 100 such events per year, the cost to the company (or ALE) will be $15,000. The resulting dollar figure provides no more than a rough yardstick, albeit a useful one, for determining whether to invest in fixing the vulnerability. Of course, for our fictional equities trading company, a $15,000 annual loss might not be worth getting out of bed for (typically, a proprietary trading company’s intraday market risk dwarfs such an annual loss figure).

Other methods take a more qualitative route. In the case of a Web server providing a company’s face to the world, the Web site’s defacement might be difficult to quantify as a financial loss (although some studies indicate a link simply between security events and negative stock-price movements2). In cases in which “intangible assets” are involved (such as reputation), qualitative risk assessment might be a more appropriate way to capture the loss.

Regardless of the technique used, most practitioners advocate a return on investment study to determine whether a given countermeasure is cost-effective for achieving the desired security goal. Adding applied cryptography to an application server via native APIs without the aid of dedicated hardware acceleration might be cheap in the short term, for example, but if it results in a significant loss in transaction volume throughput, a better ROI might come from investing up front in crypto acceleration hardware. Interested organizations should adopt the risk-calculation methodology that best reflects their needs.

Figure A. Synopsys’ risk-management framework. Many aspects of frameworks such as this can be automated—for example, risk storage, business risk to technical risk mapping, and the display of status over time.

Knowledge requirement

Design-level analysis is knowledge intensive. Microsoft’s STRIDE model, for example, involves the understanding and application of several threat categories during analysis.3 Similarly, Synopsys’ SQM approach uses attack patterns4 and exploit graphs to understand attack resistance, knowledge of design principles for ambiguity analysis,5 and knowledge regarding commonly used frameworks (.NET and J2EE being two examples) and software components.

A central activity in design-level risk analysis is to build up a consistent view of the target system at a reasonably high level. The idea is to see the forest, not get lost in the trees. The most appropriate level for this description is the typical “white board” view of boxes and arrows describing the interaction of various critical design components. The nature of software systems leads many developers and analysts to assume (incorrectly) that a code-level description of software is sufficient for spotting design problems. Although this might occasionally be true, it does not generally hold. Extreme programming’s claim that “the code is the design” represents one radical end of this approach. Without a white-board level of description, an architectural risk analysis is likely to overlook important risks related to flaws.

Risk analysis and requirements

Previous articles in this series consider security requirements definitions and discuss abuse cases as a method for generating requirements. In the purest sense, risk analysis begins at this point: design requirements should take into account the risks you’re trying to counter. Let’s look at three approaches to interjecting a risk-based philosophy into the requirements phase (note that the requirements systems based on UML tend to focus more attention on security functionality than they do on misuse and abuse cases):

• SecureUML (www.informatik.uni-freiburg.de/~tolo/pubs/secuml_uml2002.pdf) is a methodology for modeling access control policies and their integration into model-driven software development.
• UMLsec (http://www4.in.tum.de/~umlsec/) is an extension to UML that enables the modeling of security-related features such as confidentiality and access control.
• Guttorm Sindre and Andreas Opdahl6 attempt to model abuse cases as a way of understanding how applications might respond to threats in a less controllable environment; they describe functions that the system should not allow.

A key variable in the risk equation is impact. Business impacts generally boil down into three broad categories:

• federal or state laws and regulations (including the Gramm-Leach-Bliley Act, HIPAA, and the much cited California Senate Bill 1386);
• financial or commercial considerations (such as revenue protection, control over high-value intellectual property, and preservation of brand and reputation); and
• contractual considerations (including service-level agreements and avoidance of liability).

The first step to risk analysis at the requirements stage is to break down requirements into three simple categories: must haves, important to haves, and nice but unnecessary. Unless you’re running an illegal operation, you should always class laws and regulations into the first category—these requirements should be instantly mandatory and not subject to further risk analysis (although an ROI study can help you select the most cost-effective mitigations). If the law requires you to protect private information, for example, this requirement is compulsory and should not be subject to a risk-based decision. Why? Because the government has the power to put you out of business, which is the mother of all risks (if you want to test government regulators on this one, go right ahead).

You’re then left with risk impacts—the ones that have as variables potential impact and probability—that must be managed in other ways. Examples of mitigations range from technical protections and controls, to business decisions for living with the risk. At the initial requirements definition stage, you might be able to make some assumptions regarding which controls are necessary.

Evenly applying these simple ideas will put you ahead of most application developers. As you move toward the design and build stages, risk analysis should begin to test your first assumptions from the requirements stage by testing the threats and vulnerabilities inherent in the design.

Limitations

Traditional risk-analysis output is difficult to apply directly to modern software design. Even assuming a high level of confidence in the ability to predict the dollar loss for a given event and performing Monte Carlo distribution analysis of prior events to derive a statistically sound probability distribution for future events, there’s still a large gap between an ALE’s raw dollar figure (as discussed earlier) and a detailed software security mitigation definition.

A more worrying concern is that traditional risk-analysis techniques do not necessarily provide an easy guide (not to mention an exhaustive list) of all potential vulnerabilities and threats to consider at a component/environment level. This is why a large knowledge base and lots of experience is invaluable. The thorny knowledge problem arises in part because modern applications, including Web services applications, are designed to span multiple boundaries of trust. The vulnerability of—and threat to any given component varies with the platform on which that component exists (think C# on a Windows .NET server versus J2EE on Tomcat/Apache/Linux) and the environment in which it lives (think secure DMZ versus directly exposed LAN). However, few traditional methodologies adequately address the contextual variability of risk given changes in the core environment. This is a fatal flaw when considering highly distributed applications or Web services.

In modern frameworks such as .NET and J2EE, security methods exist at almost every layer, yet too many applications today rely on a “reactive” protection infrastructure that only provides protection at the network transport layer. This is too often summed up by saying, “We’re secure because we use SSL and implement firewalls,” which opens the door to all sorts of problems such as those engendered by port 80 attacks, SQL injection, class spoofing, and method overwriting (to name just a few).

One approach to overcoming these problems is to start looking at software risk analysis on a component-by-component, tier-by-tier, and environment-by-environment level and then apply the principles of measuring threats, vulnerabilities, and impacts at each level.

Acknowledgments

We thank John Steven and Stan Wisseman for their insightful comments on early drafts of this work. We also thank Bruce Phillips of Fidelity National Financial.

References

1. G. McGraw, “Software Security,” IEEE Security & Privacy, vol. 2, no. 2, 2004, pp. 80–83.
2. H. Cavusoglu, B. Mishra, and S. Raghunathan, The Effect of Internet Security Breach Announcements on Market Value of Breached Firms and Internet Security Developers, tech. report, Univ. of Texas at Dallas, School of Management, Feb. 2002; www.ut dallas.edu/~huseyin/breach.pdf.
3. M. Howard and D. LaBlanc, Writing Secure Code, 2nd ed., Microsoft Press, 2003.
4. G. Hoglund and G. McGraw, Exploiting Software, Addison-Wesley, 2004.
5. J. Viega and G. McGraw, Building Secure Software: How to Avoid Security Problems the Right Way, Addison-Wesley, 2001.
6. G. Sindre and A.L. Opdahl, “Eliciting Security Requirements by Misuse Cases,” Proc. 37th Technology of Object-Oriented Languages and Systems (TOOLS-37), IEEE CS Press, 2000.