Prioritizing Security Efforts with a Risk-Based Attack Surface Approximation

Prioritizing Security Efforts with a Risk-

Based Attack Surface Approximation

Chris Theisen

Advisor: Laurie Williams

Introduction

What is the “Attack Surface”? Quoting the Open Web Application

Security Project…

• All paths for data and commands in a software system

• The data that travels these paths

• The code that implements and protects both

Concept used for security effort prioritization.

3

Crashes represent activity that put the system under

stress.

Stack Traces tell us what happened.

Binary Function

foo!foobarDeviceQueueRequest+0x68

foo!fooDeviceSetup+0x72

foo!fooAllDone+0xA8

bar!barDeviceQueueRequest+0xB6

bar!barDeviceSetup+0x08

bar!barAllDone+0xFF

center!processAction+0x1034

center!dontDoAnything+0x1030

Pull out individual code

artifacts from traces.

If code appears on a

crash dump stack trace,

it’s on the attack surface.

4

Called Risk-Based Attack Surface Approximation, or RASA.

5

Why “Risk Based?”

• Crashes are empirical evidence of…

– Data paths through software – with flaws

• Code that is covered by RASA are therefore…

– More likely to have vulnerabilities, as we have

evidence that there are flaws on RASA

– More likely to be exploited, as they’re on known

traversable paths

6

Methodology - RASA

7

Methodology - RASA

8

Methodology - RASA

9

Proposed Solution

10

Proposed Solution

11

Proposed Solution

12

How effective is risk-based attack surface approximation in

predicting the location of security vulnerabilities in a

software system?

We are covering the majority of vulnerabilities seen!

Windows 8 User Crashes

%binaries 48.4%

%vulnerabilities 94.6%

Feasibility Study

13

Practitioner Problems

• Feasibility study used tens of millions of crashes.

• Previous study was per binary.

• Practitioners had some issues with it…

– “Binary prioritization isn’t actionable.”

– “We don’t have that much data!”

– “We don’t store every crash we received, we don’t

see the value in that.”

– “This is a new product; we don’t have reported

vulnerabilities to use as a goodness measure.”

14

Methodology - RASA

15

Methodology - Sampling

16

10% of…

Methodology - Sampling

17

10% of…20% of…

Methodology - Sampling

18

10% of…20% of…

• Sample at each “level”

• Record stdev of files,

vulnerabilities covered

19

12%

13%

14%

15%

16%

17%

70%

71%

72%

73%

74%

75%

Random Sample Size

Files

Vulnerabilities

10%

12%

14%

16%

18%

20%

22%

24%

26%

30%

32%

34%

36%

38%

40%

42%

44%

46%

Random Sample Size

20

Files

Vulnerabilities

Better Precision?

Can crash dump stack traces be used to prioritize beyond

the attack surface?

“Change,” or code being added or falling off attack surface

“Complexity” of code in software systems

“Boundary,” or code appearing on the edge of software

21

Dropped Code (DC): Code that appears on earlier version, not later

New Code (NC): Code that appears on later version, not earlier

Persistent Code (PC): Code that appears on both versions

22

Total Files Percentage

Windows 8 to 8.1

DC 7951 23.1%

NC 5490 15.9%

PC 21025 61.0%

Windows 8.1 to 10

DC 13645 43.7%

NC 4677 15.0%

PC 12870 41.3%

Windows 10 – 2014 to 2015

DC 232 1.3%

NC 15497 89.0%

PC 1674 9.6%

Change in Attack Surface

Dropped Code (DC): Code that appears on earlier version, not later

New Code (NC): Code that appears on later version, not earlier

Persistent Code (PC): Code that appears on both versions

23

Total Files Percentage

Windows 8 to 8.1

DC 7951 23.1%

NC 5490 15.9%

PC 21025 61.0%

Windows 8.1 to 10

DC 13645 43.7%

NC 4677 15.0%

PC 12870 41.3%

Windows 10 – 2014 to 2015

DC 232 1.3%

NC 15497 89.0%

PC 1674 9.6%

Change in Attack Surface

The attack surface should be recomputed

between version changes and more

frequently during the development process.

Calculate spearman correlation coefficient for fan-in and fan-out values

Fan-in: number of unique incoming calls to a file as measured by crashes

Fan-out: number of unique outgoing calls to a file as measured by

crashes

24

Coefficient p-value

Windows 8In 0.63 p < 0.01

Out 0.75 p < 0.01

Windows 8.1In 0.82 p < 0.01

Out 0.90 p < 0.01

Windows 10In 0.65 p < 0.01

Out 0.83 p < 0.01

Complexity of Code

Calculate spearman correlation coefficient for fan-in and fan-out values

Fan-in: number of unique incoming calls to a file as measured by crashes

Fan-out: number of unique outgoing calls to a file as measured by

crashes

25

Coefficient p-value

Windows 8In 0.63 p < 0.01

Out 0.75 p < 0.01

Windows 8.1In 0.82 p < 0.01

Out 0.90 p < 0.01

Windows 10In 0.65 p < 0.01

Out 0.83 p < 0.01

Complexity of Code

More complex code is more likely to have

security vulnerabilities.

Boundary Code (BC): percentage of code that appears on the boundary,

or edge of a software system

Boundary Vulnerabilities (BV): percentage of vulnerabilities on

Boundary Code (BC).

26

BC BV Ratio

Windows 82014 4.5% 17.2% 3.8

2015 4.6% 18.6% 4.0

Windows 8.12014 4.6% 16.5% 3.6

2015 6.9% 23.7% 3.4

Windows 102014 3.4% 10.5% 3.1

2015 3.9% 25.1% 6.4

Boundary of the System

Boundary Code (BC): percentage of code that appears on the boundary,

or edge of a software system

Boundary Vulnerabilities (BV): percentage of vulnerabilities on

Boundary Code (BC).

27

BC BV Ratio

Windows 82014 4.5% 17.2% 3.8

2015 4.6% 18.6% 4.0

Windows 8.12014 4.6% 16.5% 3.6

2015 6.9% 23.7% 3.4

Windows 102014 3.4% 10.5% 3.1

2015 3.9% 25.1% 6.4

Boundary of the System

The boundary (or outer edge) of a software

system is more likely to have exploited

vulnerabilities.

Systematic Literature Review

28

Systematic Literature Review

29

Systematic Literature Review

30

Developer Tools: Visualization

• We have a list of vulnerable files; now what?

– Further prioritization to assist developers.

• We looked at

– Creating an approximation of the attack surface.

– How the attack surface changes over time.

– How the complexity of the attack surface predicts

vulnerabilities.

– How proximity to the boundary of a software

system predicts vulnerabilities.

31

Developer Tools: Visualization

32

Comparing Against VPMs

There are many Vulnerability Prediction Models (VPMs)

• String Mining (Scandariato)

• Churn/Complexity (Shin)

• Development Metrics (Zimmermann)

• Static Analysis Alerts (Gegick)

How does RASA perform as standalone predictor?

Can RASA improve precision, recall of these approaches?

33

34

foo!foobarDeviceQueueRequest+0x68

foo!fooDeviceSetup+0x72

foo!fooAllDone+0xA8

bar!barDeviceQueueRequest+0xB6

bar!barDeviceSetup+0x08

bar!barAllDone+0xFF

[email protected]

@theisencr

theisencr.github.io