0-knowledge fuzzing white paper

0-knowledge fuzzing
Vincenzo Iozzo
vincenzo.iozzo@zynamics.com

February 9, 2010

Abstract and widely established methodology employed in
COTS software vulnerability discovery process.
Nowadays fuzzing is a pretty common technique The first appearance of fuzzing in software test-
used both by attackers and software developers. ing dates back to 1988 by Professor Barton
Currently known techniques usually involve Miller[1]; since then the technique has evolved
knowing the protocol/format that needs to be a lot and it is not only used by attackers to dis-
fuzzed and having a basic understanding of how cover vulnerabilities but also internally by many
the user input is processed inside the binary. companies to find bugs in their software.
In the past since fuzzing was little-used obtain- Over the course of time a lot of different imple-
ing good results with a small amount of effort mentations of fuzz testing have been researched,
was possible. nonetheless it is commonly believed that there
Today finding bugs requires digging a lot inside are two predominant approaches to fuzzing:
the code and the user-input as common vulner- Mutation-based and Generation-based.
abilies are already identified and fixed by devel- The former is based on random mutations of
opers. This paper will present an idea on how known well-formed data, whereas the latter cre-
to effectively fuzz with no knowledge of the user- ates testing samples using templates describing
input and the binary. the format of the software input.
Specifically the paper will demonstrate how tech- Both approaches have their advantages and pit-
niques like code coverage, data tainting and in- falls. The former requires little effort to be im-
memory fuzzing allow to build a smart fuzzer plemented and it is reusable across different soft-
with no need to instrument it. ware. Nonetheless given the raising interest com-
panies have shown in properly testing and devel-
oping products this approach will generally yield
1 Introduction worse results than generation-based fuzzers.
The second approach has the advantage of ob-
Fuzzing, or fuzz testing, is a software test-
taining better results in terms of bugs found, al-
ing methodology whose aim is to provide in-
though it requires knowledge of the input format
valid, unexpected or random inputs to a pro-
the binary expects and its reusability is bounded
gram. Although the idea behind this technique
to binaries that deal with the same input format.
is conceptually very simple it is a well known

2

The difficulty of creating input models can range of view.
from low for public data formats to almost infea- To the best of the author‘s knowledge there are
sible for proprietary formats. no public attempts at combining together these
In order to ease the process of creating input techniques for fuzz-testing purposes. A notable
templates various approaches have been stud- exception is Flayer which nonetheless only fo-
ied, most notably evolutionary fuzzers and in- cuses on dynamic analysis and program paths
memory fuzzers. manipulation in order to discover software de-
Both are derived from mutation-based fuzzers fects.
but for different purposes. The first type of The rest of this paper is organized as follows.
fuzzers, in fact, by employing genetic algorithms In section 2 we provide basic background infor-
attempts to generate sets of data which resem- mation on the metrics used. Section 3 discusses
ble as precisely as possible the input format. The related work. Section 4 presents our approach
latter, instead, first requires a human to manu- and implementation. Finally we conclude and
ally identify specific functions inside the binary discuss future work directions in Section 5.
then mutates the input in-memory in order to
prevent data validation which could lead to dif-
ferent code paths thus resulting in not fuzzing 2 Background
crucial pieces of an application.
Evolutionary based fuzzers suffer from the diffi- In this section, we present background informa-
culty of identifying proper scoring and mutation tion on static analysis metrics, data tainting, and
functions and for this approach to be effective it in-memory fuzzing.
usually requires more time than the generation- In our implementation we use primary two static
based one. In-memory fuzzing on the other hand analysis techniques: cyclomatic complexity and
has a high rate of false positives and negatives loop detection.
and it requires an expert reverse engineer in or- Cyclomatic complexity is a software metric used
der to identify proper test cases. to determine how complex in terms of code paths
In this paper the author presents an approach a function is. The computation is done on the
to fuzz testing based on in-memory fuzzing aim- number of edges and nodes a function contains.
ing at limiting human intervention and minimiz- Intuitively the more the structure of the function
ing the number of false positives and negatives is complicated the more complex the function
that currently affects this technique. The pro- is. In [2] the connection between function com-
posed methodology employs a range of known plexity and bugs presence has been discussed.
metrics from both static and dynamic program Although there is not always a correlation be-
analysis together with a new technique for in- tween the two, it is reasonable to assume that
memory fuzzing. Specifically we will use data more complex functions are prone to contain
tainting for tracking user input, thus being able bugs given the amount of code they contain.
to identify locations in-memory suitable for test- Another metric employed is loop detection. This
ing; we will also employ static analysis metrics algorithm takes advantage of some properties of
in order to identify functions in the binary that a function flowgraph and its dominator tree in
can be interesting from a security testing point order to detect loops present in compiled code.

3

This technique is widely used in compilers for op- the technique so that basic blocks execution is
timization purposes, and it has some interesting being traced. This implementation does not
aspects from a security prospective as well. It take into account code paths and therefore might
is commonly known, in fact, that memory write be imprecise in some circumstances, nonetheless
often happens inside loops and that most compil- we consider this trade-off to be acceptable as it
ers usually inline functions like memcpy so that avoids to overly complicating the implementa-
the function will effectively result in a loop. tion and improves the fuzzer performance.
Another crucial piece of infrastructure for the
proposed fuzzer is the data tainting engine. The
goal of data tainting is to gather information on
3 Related work
how user input is propagated through a binary. In this section we will briefly describe exist-
The concept of data tainting is intuitively very ing approaches to data tainting and in-memory
simple, one or more markings are associated with fuzzing together with a brief description of
some data supposedly representing the user in- Flayer[3] being it the closest work to the one de-
put and those markings are propagated follow- scribed in this paper.
ing the program flow. Although it is possible to
perform data tainting using static analysis the
3.1 Existing in-memory fuzzing im-
complexity of the task and the possible incom-
plementations
plete set of information led the author to choose
a dynamic analysis approach to the problem by
taking advantage of an existing dynamic data
tainting framework called Dytan[6]. Using dy-
namic data tainting has the benefit of obtain-
ing more precise and richer information on data
propagation although it will not be able to ex-
plore program paths that are not executed at (a) Mutation loop insertion
run-time. Given the nature of the fuzzer, ob-
taining information on non-executed code paths
is of no interest as in-memory fuzzing relies on
the ability to reach code paths by mutating a set
of known good data.
Finally in order to monitor the effectiveness
of our fuzzer, we employ a software testing mea-
sure known as code coverage. This technique
(b) Snapshot restoration
verifies the degree to which the code of a pro-
mutation
gram has been tested by tracing the execution
of the binary. Although there are many differ- Figure 1: Known implementations of in-memory
ent implementations of code coverage all using fuzzing
different criteria in terms of the kind of informa-
tion to record, the author decided to implement To the best of the author‘s knowledge in-

4

memory fuzzing was first introduced to the pub- of software. A lot of implementations of data
lic by Greg Hoglund of HBGary in [4] and later tainting frameworks exist, for this reason the
further developed by Amini at al[5]. Currently author decided to use a framework previously
there are two public methods: Mutation loop in- created by James Clause and Alessando Orso of
sertion and snapshot restoration mutation. Gatech called Dytan[6]. The decision was made
The first method works by inserting an uncon- based on a number of requirements.
ditional jump from the function being tested to First and foremost the ability to instrument bi-
a function responsible for mutating the data re- naries without any recompilation or access to the
siding in the process address space of the fuzzed source code.
binary. At the end of the mutation function an- Another very important requirement was porta-
other unconditional jump to the beginning of the bility, most of the existing implementations are
currently tested function is inserted. The control based on Valgrind[7] which does not support the
flow graph of this approach is shown in 1(a). Windows platform. The two most appealing
This approach suffers of a number of drawbacks candidates were Temu[9] and Dytan.
with a high rate of false negatives and stack con- The first one is built on the top of a modified
sumption being the two major ones. Another version of Qemu[8]. Although this would have
disadvantage of this method is the general insta- respected both the initial requirements we think
bility of the memory after a few fuzzing itera- that a data tainting framework based on a vir-
tions. tual machine emulator is overkill for our goals.
The second approach works by inserting an un- Besides the implementation in the author‘s opin-
conditional jump from the beginning of the func- ion is not yet robust enough.
tion being tested to a function responsible of tak- Dytan is implemented as a pintool[16]. It is a
ing a memory snapshot. This function will later flexible framework and can run on both Linux
call again the tested function. At the end of the and Windows.
analyzed function another unconditional jump is
inserted. The jump points to a function respon- 3.3 Additional related work
sible of restoring the memory, fuzzing data and
executing again the fuzzed function. A control As already mentioned in the previous section
flow diagram employing this approach is shown Flayer[3] is the most similar work to the ap-
in Figure 1(b). Although this method has some proach discussed in this paper. The software
advantages in respect to the first one described, combines data tainting and the ability to force
code paths. Differently from many other data
it still suffers from a high false positives rate and
it is also slower given the need of continuously tainting tools Flayer has bit-precision markings.
having to restore process memory. Although this grants a higher degree of precision
in obtaining information on data propagation for
3.2 Existing data tainting implemen- the purpose of our work byte-precision markings
are detailed enough.
tations
Another limitation is the software the tool is
Dynamic data tainting has gained momentum in based on; as already mentioned Valgrind does
the last few years given the increase complexity not support Windows which severly impairs the

5

usefulness of the tool. should be replaced by a more sophisticated ap-
Finally even if the main aim of the tool is not proach which takes into account scores coming
fuzzing it has the ability of forcing code paths from various metrics and weights them in respect
and therefore it can be used to test various code to their relevance from a security prospective.
paths. This method has three main drawbacks;
the first one is a high number of false positives,
the second one is the absence of a sample which
can be later used by the attacker to reproduce
the bug and finally a problem known as code-
path explosion. This problem arises because the
number of code paths to force increases exponen-
tially with the complexity of the software.

Figure 3: The edge in red is missed by the ap-
4 Proposed approach and im- proximative cyclomatic complexity formula.
plementation
Cyclomatic complexity Cyclomatic com-
plexity was first described by Robert McCabe
in [10]. The purpose of this metric is to calcu-
late the number of independent paths in a code
section. Many formulation of this metric have
been given, we briefly explain the ones that are
Figure 2: Fuzzer components relevant to our fuzzer.

In this section we will present the idea and Definition Let G be a flowgraph, E the num-
implementation of our work. As shown in Figure ber of edges in G, N the number of nodes in G
2 our fuzzer can be divided into 4 parts. and P the number of connected components in
G.
Cyclomatic complexity is defined as:
4.1 Static analysis metrics
M = E − N + 2P (1)
Static analysis algorithms are used to determine
which functions could be potentially of inter- A connected component is a subgraph in which
est for our fuzzer. We assign a higher score to any two vertices are connected to each other by
functions that have a high cyclomatic complex- paths. This formula originates from the cyclo-
ity score and at least one loop in them; we then matic number:
consider all the functions that have loops but a
low cyclomatic complexity score and finally we Definition Let G be a strongly connected
take into account the remaining functions. Ide- graph, E the number of edges in G, N the
ally we will add more metrics to the implementa- number of nodes in G and P the number of
tion, therefore this rather trivial scoring system connected components in G.

6

The Cyclomatic number is defined as:

V (G) = E − N + P (2)

It should be notice that the cyclomatic num-
(a) A (b) Domi- (c) The
ber can be calculated only on strongly connected
function nator tree nodes
graphs, that is a graph in which from every pair flowgraph, of the in green
of vertices there is a direct path connecting them nodes in previous dominate
in both directions. McCabe proved that the flow- blue belong function, the node in
to a loop nodes red in the
graph of a function with a single entry point and
in green dominator
a single exit point can be considered a strongly correspond tree
connected graph and therefore the cyclomatic to the
number theorem applies and that P = 1, thus blue ones
the resulting simplified formula is: highlighted
in picture
(a)
M =E−N +2 (3)

Intuitively when a flowgraph has multiple exit Figure 4: Graphs used in loop detection algo-
points the aforementioned formula doesnt hold rithm
true anymore. Another one should be therefore
used: Loop detection algorithm As previously
mentioned another metric, loop detection, is
Definition Let G be a flowgraph, π the number used to select functions. The first required step
of decision points in G and s the number of exit is to extract the dominator tree out of a function.
points in G. Cyclomatic complexity is defined Formally:
as:
Definition A dominator tree is a tree where
M =π−s+2 (4) each node‘s children are the nodes it immedi-
ately dominates.
Applying (3) to functions with multiple exit
A node d is said to dominate node k if every
points we will have, in fact, lower cyclomatic
path from the start node s to node k must go
complexity values by a minimum factor of 2. Fig-
throught node d.
ure 3 shows typical edges and connected com-
ponents missed by using (3). To give a visual example of a dominator tree
Nonetheless the author believes that the less pre- of a function please refer to Figure 4. Nodes in
cise measurement can be used without impairing blue in Figure 4(a) are highlighted in the dom-
the results. inator tree in green in Figure 4(b).
We implemented cyclomatic complexity calcula- There are two known algorithms used to cal-
tion for each function in a module by using Bin- culate the dominator tree of a flowgraph. It
Navi API. A detailed explanation of the imple- is out of the scope of this paper to discuss
mentation can be found in[11]. them. It should be noticed, though, that the

7

tool upon which we built our loop detection Each data tainting implementation can choose
algorithm, BinNavi[12], implements Lengauer- the type of markings to use, more precisely it
Tarjan[13] dominator tree algorithm which is al- is possible to determine the granularity of those
most linear thus granting us a higher computa- markings.
tional speed. Dytan is able to either assign a single marking to
The second step is to calculate for each node its each piece of input or have byte-level markings.
dominators. In Figure 4(c) the dominators of We chose to use the second type of markings as
the node in red are the ones in green. it is more precise but at the same time does not
The last step is to search for edges from a node cause an excessive overhead during the execu-
to one of its dominators. Recalling the definition tion.
of domination it is trivial to show that if there In order to make data tainting work it is impor-
is an edge from a node to one of its dominators tant to define what data needs to be tracked. In
a loop is present. Dytan it is possible to track user input coming
Most complex assembly instruction sets have from network operations, files access and com-
what are called implicit loops instruction, for in- mand line arguments passed to the main() func-
stance rep movs in x86 ISA. Applying this al- tion. That is system calls and functions respon-
gorithm to a flowgraph will therefore miss this sible for the aforementioned input sources are
type of loops. monitored and their output is tracked through
In order to overcome this problem we will trans- the binary.
late the function to an intermediate language Another important factor to take into account
called REIL[14] implemented in BinNavi. This while implementing a data tainting tool is a
intermediate language provides a very small set propagation policy.
of instructions which helps in the process of un- A propagation policy is a set of rules followed
folding implicit loops. while taint markings are assigned during pro-
In [15] a detailed implementation of this algo- gram execution.
rithm can be found. Dytan currently is able to perform control and
data-flow or data-flow only analysis. The former
4.2 Data tainting tracks direct or transitive data assignments as
well as indirect propagation due to control flow
As stated before the author did not implement dependencies upon user input. The latter in-
the data tainting framework employed by the stead can only track direct and transitive data
fuzzer, nonetheless given the critical importance assignments. In our fuzzer we use the second
of data tainting for this project the author thinks approach as control flow analysis does not add
it is important to briefly describe how dytan any useful information on data locations to be
works and how we use this framework for our tested.
purposes. Another problem to tackle while creating a prop-
We previously mentioned that data tainting is a agation policy is how to deal with multiple mark-
technique to track user input inside a binary. ings assigned to the same input. Dytan currently
Tracking is usually performed by assigning mark- assigns to the resulting taint marking the union
ings to data while executing the binary. of all the taint markings related to it. Although

8

for our fuzzer a different approach might grant At point 1 we search for the address of the func-
better results we currently use the default dytan tion we are interested in fuzzing and install the
policy. analysis function for that function. At point 2
Finally we make dytan provide information on we iterate through function instructions locating
every instruction that assigns taint markings. the ones that are of interest in order to install an
That is for each of those instructions we obtain analysis function as described in 3.
the state of taint markings on machine registers The first approach consists of mutating memory
and on memory locations that are tainted at that locations and registers in place. That is instead
specific program point. of allocating new memory and pointing instruc-
tions operands to it we modify the content of
both memory locations and registers within their
4.3 In-memory fuzzing
length boundaries.
We presented in section 3 the two known ap- We then continue the program execution until
proaches to in-memory fuzzing. In this section the program quits or new data is obtained from
we are going to present two slightly different ap- a tainted source.
proaches which we believe to gain better results This approach is more conservative than all the
given the amount of information we can gather others as it does not change the memory layout
from data tainting analysis. thus the number of false positives is reduced but
We implemented our in-memory fuzzer on top at the expenses of an increased number of false
of PIN[16]. PIN has the ability to add instru- negatives.
mentation functions before and after a binary is The second approach works very similar to SRM
loaded in memory, functions and instructions. 1(b). In addition to the first three steps we also
Recalling that for each instruction that assigns add an instrumentation functions at the end of
taint markings we retrieve from data tainting the tested function. This function will be re-
analysis, we get the markings associated to ma- sponsible of restoring memory after fuzzing was
chine registers and memory locations, we are performed. With the second approach the mem-
able to precisely identify program points during ory layout is changed as the fuzzer will allocate
binary execution that are suitable for fuzzing. chunks of memory to be used during the fuzzing
For both approaches we perform a number of phase.
steps: As for the first approach the program execution
is continued until the application quits or new
1. Install an analysis function on image load- data is obtained from a tainted source.
ing. Although our second approach is similar to SRM
there are a few notable differences that have to
2. Install an analysis function before the func- be considered. First we do not take a full snap-
tion we are interesting in fuzzing is exe- shot of the process memory but we only track
cuted. modifications that occurred due to fuzzing dur-
ing the execution of the tested function. The
3. Install an analysis function before each in- second difference is that memory is not totally
struction that assigns taint markings. restored after the function was fuzzed, this can

9

allow us to reduce the number of false negatives The following must hold true:
since possible bugs caused by a faulty execution
of the function are not missed by restoring the C1 ≤ C + t (5)
full process memory.
The halting point is defined as:
It has to be noted that both approaches de-
scribed here although more effective cannot be C1 = C + t (6)
used without a proper amount of information
gathered by the means of data tainting analy- The code coverage score is calculated as fol-
sis or some similar techniques. lows:

Definition Let BBt be the totality of basic
4.4 Code coverage blocks in a binary, BBf the number of basic
blocks touched in a single execution.
The combination of code coverage with fuzz test-
The code coverage score is defined as:
ing has long been used in order to measure the
effectiveness of fuzzing. We implemented code BBf
coverage on the top of BinNavi debugging API. C= (7)
BBt
The choice of using BinNavi debugger serves a
double purpose, not only we are able to imple- A detailed implementation of code coverage
ment code coverage using lightwave breakpoints using BinNavi API can be found in [17].
which highly reduce execution overhead but we
are also able to monitor the execution for possi- 5 Results and future work
ble faults. We decided to implement code cov-
erage at basic blocks level, that is a breakpoint In this paper we have described a new approach
is set at the beginning of each basic block in the to fuzz testing which highly reduce instrumenta-
tested binary. We perform code coverage first tion costs thus resulting very useful when dealing
when the binary is executed with a known good with large proprietary applications.
sample, later it is calculated again every time We have also shown how it is possible to com-
the program is fuzzed. We require the fuzzing bine static and dynamic analysis techniques to
sample to perform at least as good as the known triage interesting functions from a security test-
good sample, we also set a threshold defining the ing point of view.
upper-bound after which the sample reaches the Finally we have proposed a new approach to in-
”halting point”. The ”halting point” is the point memory fuzzing which is more precise and less
where the fuzzing process is re-initialized with a prone to false negatives than previous known
new known good sample as shown in Figure 2. techniques.
Formally: We do not have enough data to determine
whether this approach has better results com-
Definition Let C be the code coverage score pared to other fuzzing techniques.
of a known good sample, C1 the code coverage The author believes that compared to other
score of a fuzzing sample, t a user supplied mutation-based and evolutionary-based method-
delta. ologies the one proposed in this paper will have

10

better results. In comparison to generation- ﬁrst USENIX workshop on Oﬀensive Tech-
based fuzzers our technique will have better re- nologies.
sults when dealing with complex software but
worse results when the software input is simple. [4] G. Hoglund: Runtime Decompilation: The
The main direction of future work will be focused GreyBox process for Exploiting Software,
on reducing false positives by employing con- Black Hat DC 2003
straint reasoners to determine whether a given [5] M. Sutton, A. Greene, P. Amini:
bug is reproducible with valid but unexpected Fuzzing:brute force vulnerability discovery.
input. Addison-Wesley.
Another important challenge is to implement
more static analysis metrics to triage functions [6] J.Clause, W. Li, A. Orso: Dytan:a generic
with a higher degree of precision. dynamic taint analysis framework, Proceed-
ings of the 2007 international symposium on
Software testing and analysis
Acknowledgments
[7] Valgrind: http://www.valgrind.org
The author would like to thank Thomas Dullien,
[8] Qemu: http://www.qemu.org
Dino Dai Zovi and Shauvik Roy Choudhary for
their suggestions and help while researching the [9] Temu:http://bitblaze.cs.berkeley.edu/temu.html
topic.
The author would also like to thank James [10] T. J. McCabe: A Complexity measure,
Clause and Alessandro Orso for having provided IEEE transactions on software engineering,
access to dytan source code and their help while vol. se-2, no.4, december 1976
testing and improving the original code base. [11] V. Iozzo: Scripting with BinNavi - Cyclo-
Finally we want to thank all the people who have matic Complexity
reviewed the paper.
[12] BinNavi: http://www.zynamics.com/binnavi.html

References [13] T. Lengauer and R. E. Tarjan: A fast algo-
rithm for nding dominators in a owgraph,
[1] B.P. Miller, L. Fredriksen, and B. So: ”An ACM Transactions on Programming Lan-
Empirical Study of the Reliability of UNIX guages and Systems
Utilities”, Communications of the ACM 33,
[14] T. Dullien, S. Porst: REIL: A platform-
12 (December 1990)
independent intermediate representation of
[2] Kan: Metrics and Models in Software disassembled code for static code analysis,
Quality Engineering. Addison-Wesley. pp. CanSecWest 2009
316317. [15] V. Iozzo: Finding interesting loops us-
ing(Mono)REIL
[3] W. Drewry, T. Ormandy: Flayer:exposing
application internals, Proceedings of the [16] PIN: http://www.pintool.org

11

[17] V. Iozzo: Code coverage and BinNavi

12

0-knowledge fuzzing white paper

More Related Content

What's hot (16)

Viewers also liked (9)

Similar to 0-knowledge fuzzing white paper (20)

More from zynamics GmbH (10)

0-knowledge fuzzing white paper