U.S. patent application number 15/079719 was filed with the patent office on 2016-09-29 for systems and methods for detecting copied computer code using fingerprints.
The applicant listed for this patent is Terbium Labs, Inc.. Invention is credited to Dionysus BLAZAKIS, Michael MOORE, Daniel J. ROGERS.
Application Number | 20160283229 15/079719 |
Document ID | / |
Family ID | 56975717 |
Filed Date | 2016-09-29 |
United States Patent
Application |
20160283229 |
Kind Code |
A1 |
ROGERS; Daniel J. ; et
al. |
September 29, 2016 |
SYSTEMS AND METHODS FOR DETECTING COPIED COMPUTER CODE USING
FINGERPRINTS
Abstract
Systems and methods of detecting copying of computer code or
portions of computer code involve generating unique fingerprints
from compiled computer binaries. The unique fingerprints are
simplified representations of functions in the compiled computer
binaries and are compared with each other to identify similarities
between functions in the respective compiled computer binaries.
Copying can be detected when there are sufficient similarities
between fingerprints of two functions.
Inventors: |
ROGERS; Daniel J.;
(Baltimore, MD) ; MOORE; Michael; (Columbia,
MD) ; BLAZAKIS; Dionysus; (Parkville, MD) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Terbium Labs, Inc. |
Baltimore |
MD |
US |
|
|
Family ID: |
56975717 |
Appl. No.: |
15/079719 |
Filed: |
March 24, 2016 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
62138543 |
Mar 26, 2015 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06F 8/751 20130101 |
International
Class: |
G06F 9/44 20060101
G06F009/44 |
Claims
1. A method comprising: receiving, by a computer, a first compiled
computer binary; generating, by the computer, a first fingerprint
of a first function of the first compiled computer binary by
generating a block rank score for each block in the first function;
generating a path of blocks of the first function based on the
block rank score of each block; generating the first fingerprint
using the generated path of blocks of the first function;
receiving, by the computer, a second compiled computer binary;
generating, by the computer, a second fingerprint of a second
fingerprint of the second compiled computer binary by generating a
block rank score for each block in the second function; generating
a path of blocks of the second function based on the block rank
score of each block; generating the second fingerprint using the
generated path of blocks of the second function; comparing, by the
computer, the first fingerprint of the first function with the
second fingerprint of the second function; and determining, by the
computer, whether the second function includes at least some of
code from the first function based on the comparison.
2. The method of claim 1, wherein the generation of a block rank
score for each block in the first and second functions involves
reversing all edges in a control flow graph for the first and
second functions and assigning a block rank score to each block
based on block rank scores of blocks having a forward path to the
each block.
3. The method of claim 1, wherein the generation of the path of
blocks for the first and second functions comprises: starting from
a first block in the first and second functions; selecting a next
block in the path based on the block rank score of possible next
blocks; and storing, for each block in an ordered list based on a
position within the path, a 3-tuple comprising a number of
instructions in the block, an in-degree of the block, and an
out-degree of the block.
4. The method of claim 3, wherein the path of blocks of the first
function includes less than all blocks of the first function and
the ordered list for the first function only includes 3-tuples for
blocks in the path of blocks for the first function.
5. The method of claim 1, wherein the first and second complied
computer binaries are complied using different compilers, one of
the different compilers adds at least one block to the first or
second function that is not present in the other of the first and
second functions, and a match is determined between the first and
second functions even though the added at least one block is not
present in both the first and second functions.
6. The method of claim 1, wherein the comparison of the first and
second fingerprints involves comparing 3-tuples in a first ordered
list for the first function with 3-tuples in a second ordered list
for the second function, wherein the 3-tuples in the first and
second ordered lists comprise a number of instructions in the
block, an in-degree of the block, and an out-degree of the
block.
7. The method of claim 6, wherein the comparison of the 3-tuples in
the first and second ordered lists comprises: generating a
two-dimensional array of all possible pair combinations between the
3-tuples in the first and second ordered lists; generating a
maximum match pathway through the two-dimensional array.
8. The method of claim 1, wherein a plurality of functions are
generated from a plurality of received complied computer binaries,
and each of the plurality of functions is fingerprinted and
compared to one of the first and second fingerprints to determine
whether each of the plurality of functions matches the one of the
first and second fingerprints.
9. A method comprising: receiving, by a computer, a plurality of
complied computer binaries; generating, by the computer, a
plurality of fingerprints for a corresponding plurality of
functions in each of the plurality of compiled computer binaries;
storing, by the computer, the plurality of fingerprints in a
database; receiving, by the computer, a query function; generating,
by the computer, a query fingerprint for the query function by
generating a block rank score for each block in the query function;
generating a path of blocks of the query function based on the
block rank score of each block; generating the query fingerprint
using the generated path of blocks of the query function;
comparing, by the computer, the query fingerprint to each of the
plurality of fingerprints; and determining, by the computer, that
the query fingerprint contains copied code from one of the
plurality of fingerprints based on the comparison.
10. The method of claim 9, wherein the generation of a block rank
score for each block in the query function involves reversing all
edges in a control flow graph for the query function and assigning
a block rank score to each block based on block rank scores of
blocks having a forward path to the each block.
11. The method of claim 9, wherein the generation of the path of
blocks for the query function comprises: starting from a first
block in the query function; selecting a next block in the path
based on the block rank score of possible next blocks; and storing,
for each block in an ordered list based on a position within the
path, a 3-tuple comprising a number of instructions in the block,
an in-degree of the block, and an out-degree of the block.
12. The method of claim 11, wherein the path of blocks of the query
function includes less than all blocks of the query function and
the ordered list for the query function only includes 3-tuples for
blocks in the path of blocks for the query function.
13. The method of claim 9, wherein the comparison of the query
fingerprint with the plurality of fingerprints involves comparing
3-tuples in a first ordered list for the query function with
3-tuples in ordered lists for each of the plurality of functions,
wherein the 3-tuples in the first ordered list and in the ordered
lists for each of the plurality of functions comprise a number of
instructions in the block, an in-degree of the block, and an
out-degree of the block.
14. The method of claim 3, wherein the comparison of the 3-tuples
in the first and second ordered lists comprises: generating a
two-dimensional array of all possible pair combinations between the
3-tuples in the first and second ordered lists; generating a
maximum match pathway through the two-dimensional array.
15. A method comprising: receiving, by a computer, a first compiled
computer binary; generating, by the computer, a first fingerprint
of a first function of the first compiled computer binary;
comparing, by the computer, the first fingerprint of the first
function with a second fingerprint of a second function by
comparing 3-tuples in a first ordered list for the first function
with 3-tuples in a second ordered list for the second function by
generating a two-dimensional array of all possible pair
combinations between the 3-tuples in the first and second ordered
lists; generating a maximum match pathway through the
two-dimensional array, wherein the 3-tuples in the first and second
ordered lists comprise a number of instructions in the block, an
in-degree of the block, and an out-degree of the block; and
determining, by the computer, whether the second function includes
at least some of code from the first function based on the
comparison.
16. The method of claim 15, wherein the generation of the first
fingerprint comprises: generating a block rank score for each block
in the first function; generating a path of blocks of the first
function based on the block rank score of each block; generating
the first fingerprint using the generated path of blocks of the
first function.
17. The method of claim 16, wherein the generation of a block rank
score for each block in the first function involves reversing all
edges in a control flow graph for the first function and assigning
a block rank score to each block based on block rank scores of
blocks having a forward path to the each block.
18. The method of claim 16, wherein the generation of the path of
blocks for the first function comprises: starting from a first
block in the first function; selecting a next block in the path
based on the block rank score of possible next blocks; and storing,
for each block, a 3-tuple for the block in an ordered list based on
a position of the block within the path.
19. The method of claim 18, wherein the path of blocks for the
first function includes less than all blocks of the first function
and the ordered list for the query function only includes 3-tuples
for blocks in the path of blocks for the query function.
Description
CROSS REFERENCE TO RELATED APPLICATIONS
[0001] This application claims the benefit of Provisional
Application 62/138,543, filed Mar. 26, 2015, the entire disclosure
of which is herein expressly incorporated by reference. This
application is related to Provisional Application 61/973,125, filed
Mar. 31, 2014, U.S. Non-Provisional application Ser. No.
14/314,307, filed Jun. 25, 2014, and U.S. Non-Provisional
application Ser. No. 14/621,554, filed Feb. 13, 2015, the entire
disclosures of which are herein expressly incorporated by
reference.
BACKGROUND OF THE INVENTION
[0002] Software theft has been, and continues to be, pervasive.
Individuals and companies typically try various techniques to
combat software theft, including requiring a unique software key to
install software, requiring online activation of software,
requiring an active online connection to use software, encryption
of software, and the like. Although these techniques typically
prevent casual users from installing unauthorized copies, the
techniques can typically be overcome by sophisticated users.
[0003] Another way to combat software theft is to try to identify
the source of the stolen software using watermarks. This involves
applying unique watermarks to each copy of the software so that
when a stolen piece of software is found, the watermark in the
stolen software will corresponding to one of the unique watermarks
in the authorized software. This requires modification of the
computer code, which is undesirable. Further, this technique can be
overcome by removing the watermark from the stolen software or
removing the watermark from the authorized software so that all
further copies do not contain the unique watermark.
SUMMARY OF THE INVENTION
[0004] In addition to the issues identified above with the known
techniques for combating software theft, these techniques focus on
the software as a whole, and thus cannot identify when only
portions of the underlying code are stolen. For example, if a
watermark were applied to the software, the watermark would not
appear in the stolen software if less than the entire code were
used. Similarly, if software theft were identified by comparing
hash values generated from the authorized and stolen software, the
hash values would not match when less than the entire underlying
code is present in the stolen software. Thus, a thief could simply
modify some portion of the code to defeat these techniques.
Further, it is often the case that only a portion of the underlying
code is truly unique and provides the overall value to the
software, and accordingly a thief may only want to use this unique
portion in different software. For example, a thief could copy one
or several functions from someone else's software and incorporate
the function(s) in his/her own software. Because these functions
may comprise only a small part of the overall code in the thief's
software, simple watermarking and hash value techniques across an
entire set of code would not identify this theft of several
functions from someone else's code.
[0005] Exemplary embodiments of the present invention are directed
to techniques for combating software theft by identifying whether
at least a portion of one piece of software appears in another
piece of software. Thus, the present invention allows the
identification of whether portions of one piece of software appear
in a different piece of software, even when the overall operation
of the two pieces of software is different. The inventive technique
is particularly useful because it operates using compiled computer
binaries, and thus does not require access to the underlying source
code.
[0006] In accordance with exemplary embodiments of the present
invention, fingerprints are generated for functions in compiled
computer binaries. The fingerprint generation involves generating a
block rank score for each block of a function and then generating
order list of blocks in a path through the function using the block
rank scores. The comparison of fingerprints involves the use of the
ordered list of blocks for each corresponding function to identify
similarities between the fingerprints.
BRIEF DESCRIPTION OF THE DRAWINGS
[0007] FIG. 1 is a flow diagram of the overall process of exemplary
embodiments of the present invention;
[0008] FIG. 2 is a block diagram of an exemplary system for
generating and matching fingerprints in accordance with the present
invention;
[0009] FIG. 3 is a flow diagram of an exemplary process for
generating a fingerprint using binary code in accordance with the
present invention;
[0010] FIG. 4 is a block diagram of an exemplary call graph in
accordance with the present invention;
[0011] FIG. 5 is a block diagram of an exemplary control flow graph
in accordance with the present invention;
[0012] FIG. 6 is a block diagram of an exemplary fingerprint in
accordance with the present invention;
[0013] FIGS. 7A and 7B are flow diagrams of an exemplary process
for matching fingerprints in accordance with the present
invention;
[0014] FIG. 8 is a flow diagram of an exemplary fingerprinting
technique in accordance with another embodiment of the present
invention;
[0015] FIG. 9 is a block diagram illustrating application of block
ranking to an exemplary function in accordance with the present
invention;
[0016] FIGS. 10A and 10B are flow diagrams of an exemplary process
for matching fingerprints in accordance with another embodiment of
the present invention; and
[0017] FIGS. 11A-11C are block diagrams illustrating exemplary
arrays used for calculating fingerprint similarity scores in
accordance with the present invention.
DETAILED DESCRIPTION OF THE DRAWINGS
[0018] FIG. 1 is a flow diagram of the overall process of exemplary
embodiments of the present invention. As illustrated in FIG. 1, the
overall process involves generating a fingerprint from compiled
computer binaries (step 300 or 800) and matching the generated
fingerprints to determine whether there is sufficient similarity
(step 700 or 1000). It should be recognized that fingerprint
generation from compiled computer binaries that are later compared
for matching can be performed at approximately the same time or can
be performed at different times. For example, a fingerprint of a
first compiled computer binary can be generated for purposes of
identifying theft of the underlying code. Other compiled computer
binaries can then be collected over a period of time, and then
fingerprints can be generated using the other compiled computer
binaries for comparison with the first compiled computer binaries.
These other compiled computer binaries can be obtained in any
manner, such as, for example, using a web spider that crawls across
the Internet and collects compiled computer binaries. The other
compiled computer binaries can also be manually input. For example,
the owner of a first compiled computer binary may suspect that a
second compiled computer binary contains code stolen from the first
compiled computer binary. In this case, the fingerprints can be
generated from the first and second compiled computer binaries at
approximately the same time and then compared using the inventive
fingerprint matching technique.
[0019] FIG. 2 is a block diagram of an exemplary system for
generating and matching fingerprints in accordance with the present
invention. The system 200 can comprise one or more computers that
include a processor 205 coupled to memory 210, input 215, and
output 220. The disclosed processes can be performed by processor
205 executing computer code stored in memory 210. Processor 205 can
be any type of processor, including a microprocessor, field
programmable gate array (FPGA), and/or an application specific
integrated circuit (ASIC). Memory 210 can be any type of
non-transitory memory. In addition to storing computer code for
executing the processes described herein, memory 210 can also store
the generated fingerprints. Alternatively or additionally, a
separate storage medium can store the generated fingerprints. For
example, the computer binaries, fingerprints, and comparison scores
can be stored in a distributed file system and non-relational,
distributed database. Input 215 provides mechanisms for controlling
the disclosed processes, including, for example, a keyboard, mouse,
trackball, trackpad, touchscreen, etc. Further, input 215 can
include a connection to an external storage device for providing
compiled computer binaries, such as an external hard drive or flash
storage memory, as well as a network connection. Output 220 can
include a display, printer, and/or the like. Additionally, output
220 can include a network connection for notifying an owner of a
compiled computer binary of any identified potential infringement,
such as by electronic mail, posting on a website or webpage, a text
message, and/or the like.
[0020] FIG. 3 is a flow diagram of an exemplary process for
generating a fingerprint using binary code in accordance with the
present invention. This process is performed using the system of
FIG. 2. Initially, processor 205 receives a compiled computer
binary (step 305) via input 215 and/or memory 210 and measures bulk
file characteristics and meta-data (step 310). Next, processor 205
disassembles the compiled computer binary, generates a call graph
from the disassembled binary, and generates control flow graphs for
each function in the call graph (step 315). Once the compiled
computer binary is disassembled, the remainder of the processing
can be performed independent of the particular language, operating
system, or architecture that the code was written for.
[0021] FIG. 4 is a block diagram of an exemplary call graph in
accordance with the present invention. As will be appreciated by
those skilled in the art, a call graph describes the relationship
between various functions in a compiled binary. Thus, in FIG. 4, a
main function 405 (also commonly referred to as a routine) has
calls to sub-functions 410 and 415 (also commonly referred to as
sub-routines). In turn, function 415 has calls to functions 420,
425, and 430, and function 425 has calls to functions 435 and 440.
Those skilled in the art will recognize that in a call graph each
function is commonly referred to as a node and the connections
between functions are commonly referred to as edges. It will be
recognized that the call graph of FIG. 4 is a highly simplified
graph and that compiled computer binaries typically will be
disassembled into much more extensive call graphs.
[0022] FIG. 5 is a block diagram of an exemplary control flow graph
in accordance with the present invention. Those skilled in the art
will recognize that a control flow graph for a function describes
all possible paths that may be traversed during execution of the
particular function. Examples of paths that may be executed in the
function of FIG. 5 include blocks 505, 515, 520, 525, 530, 535, and
545; 505, 515, 520, 525, 530, 535, 540, 550, 560, and 565; 505,
510, 515, 550, 560, and 565; and 505, 510, 515, 550, 555, and 565.
In addition to these paths, any path with connections illustrated
in FIG. 5 can be traversed during the execution of the function.
Further, it should be recognized that this is merely an exemplary
function and that other functions are within the scope of the
invention.
[0023] Returning to FIG. 3, after processor 205 generates the
control flow graphs for each function (step 315), processor 205
selects one of the functions for processing (step 320). As
illustrated, this processing is performed along two parallel paths.
It should be recognized, however, that these two paths can be
performed serially, if desired. It should also be recognized that
this parallel processing does not require all of the control flow
graphs to be generated, and accordingly this parallel processing
can be performed as control flow graphs are generated. Turning to
the path on the left-hand side of FIG. 3, first processor 205
calculates the function out-degree using the call graph for the
particular function (step 325). The function out-degree represents
the number of paths or calls from a particular function to other
functions in the call graph. Next processor 205 calculates the
total number of blocks within the selected function's control flow
graph (step 330), and finally processor 205 calculates the total
number of edges within the selected function's control flow graph
(step 335). The edges in the control flow graph are the connections
between the different blocks within this graph.
[0024] Turning now to the path on the right-hand side of FIG. 3,
first processor 205 calculates the leading Eigenvector of the
adjacency matrix (step 340). The adjacency matrix is comprised of
the function coordinates 620 (described below in connection with
FIG. 6). Next, processor 205 calculates the Markov chain (step
345). The Markov chain is calculated starting with the leading
Eigenvector and then appending connected nodes corresponding with
successively smaller elements of the leading Eigenvector. The
Markov chain provides a good low-rank approximation, or
serialization, of the control flow graph that is relatively unique
and particularly well-suited for further statistical analysis of
the control flow graph.
[0025] Processor 205 then calculates count block size, in-degree,
and out-degree along the Markov chain (step 350). These three
spectra are relatively unique among and within compiled computer
binaries. An example of the count block size, in-degree, and
out-degree will now be described in connection with FIG. 5, and
assuming a chain between the blocks as follows:
505=>515=>550=>560=>565. In this example the block
count spectrum would be [12, 2, 3, 3, 9] because block 505 has 12
instructions, block 515 has two instructions, blocks 550 and 560
each have three instructions, and block 565 has nine instructions.
The in-degree spectrum would be [0, 2, 3, 1, 2] because block 505
does not have any incoming edges, block 515 has two incoming edges,
block 550 has three incoming edges, block 560 has one incoming
edge, and block 565 has two incoming edges. The out-degree spectrum
would be [2, 2, 2, 1, 1] because blocks 505, 515, and 550 each have
two outgoing edges and blocks 560 and 565 each have two outgoing
edges. The values for each spectra and ordering of values provides
a unique signature for a particular control flow graph that can be
used to identify other functions that have the same or similar
unique signatures.
[0026] After the processing of the two parallel paths is complete,
processor 205 determines whether there are any further functions to
process (step 355). If there are ("Yes" path out of decision step
355), then the next function is selected (step 360) and the
parallel processing is repeated. If not ("No" path out of decision
step 355), then processor 205 generates the fingerprint of the
binary using the calculated information (step 365).
[0027] FIG. 6 is a block diagram of an exemplary fingerprint in
accordance with the present invention. As illustrated in FIG. 6,
the fingerprint includes bulk file characteristics and meta-data
610, function coordinates 620, and unique spectra 630. The function
coordinates 620 includes, for each function, the calculated
out-degree, number of blocks, and number of edges. The unique
spectra includes, for each function, the calculated block size,
in-degree, and out-degree. These fingerprints are then used for
comparison against fingerprints of other compiled computer binaries
as described below in connection with FIGS. 7A and 7B.
[0028] FIGS. 7A and 7B are flow diagrams of an exemplary process
for matching fingerprints in accordance with the present invention.
This process is performed using the system of FIG. 2. Initially,
processor 205 receives two fingerprints of compiled computer
binaries via input 215 and/or memory 210 (step 705) and compares
their bulk file characteristic and meta-data (step 710). The bulk
file characteristic can include measuring the Jaccard Similarity,
which is a numerical score representing the amount of overlap
between strings contained in both fingerprints, and is typically
between 0 and 1.
[0029] Next, processor 205 computes distances between each pair of
possible functions of the two fingerprints using the function
coordinates 620 of the respective fingerprints (step 720).
Processor 205 then sorts the distances (step 720) and discards
function pairs having distances greater than a threshold distance
(step 725). This step reduces the processing load because only the
most-likely related functions will have distances below the
threshold. Thus, the particular threshold value can be selected
depending upon the available processing power of the computer and
the desired run-time of the fingerprint comparison process.
Furthermore, those skilled in the art will recognize that above a
certain distance it is highly unlikely that functions will be
related, and thus at least some thresholding should be performed to
reduce unnecessary processing.
[0030] Processor 205 then generates a list using the remaining
function pairs (step 730), and one of the function pairs from the
reduced list is selected for further processing of the unique
spectra 630 (step 735). Specifically, processor 205 calculates a
cross-correlation for the block-size spectra (step 740), the
in-degree spectra (step 745), and the out-degree spectra (step
750). It will be recognized that the cross-correlation is a measure
of how closely the spectra of the two fingerprints are related.
Next processor 205 determines whether any function pairs remain to
be processed (step 755). If so ("Yes path out of decision step
755), then the next function pair is selected from the reduced list
(step 735) and the cross-correlation of the unique spectra are
calculated (steps 740-750). The cross-correlation can produce a
correlation coefficient indicating the degree of similarity or
correlation. For example, a coefficient of -1 indicates complete
anti-correlation and 1 indicates complete correlation (i.e., the
two fingerprints have the same control flow graph).
[0031] If there are no remaining function pairs to process ("No"
path out of decision step 755), then processor 205 calculates a
block-size, in-degree and out-degree spectra ratios (step 760-770).
These ratios are calculated by dividing a total number of
respective correlation coefficients above a threshold by a total
number of correlation coefficients. The threshold used for the
calculation of the three ratios can be the same or different.
Processor 205 then selects the maximum ratio of the unique spectra
ratios (step 775) and generates a comparison score based on the
selected maximum ratio (step 780). The generated comparison score
is then used by processor 205 to identify infringement of one of
the compiled computer binaries (step 785). The comparison score is
generated by comparing the selected maximum ratio of the unique
spectra ratios to a threshold, and accordingly infringement is
identified when the selected maximum ratio is above the threshold.
The threshold can be set, for example, by training the system using
known data and a particular compiled computer binary for which it
is to be determined whether there are other compiled computer
binaries infringing the particular compiled computer binary. This
training identifies commonalities between the known data and the
particular compiled computer binary so that the threshold can be
set to avoid false positives indicating infringement due to code
commonly used across different pieces of software that would not be
an indicator of infringement.
[0032] When, based on the generated comparison score, there is
sufficient similarity between the compiled computer binaries or
portions of the compiled computer binaries, processor 205 can
notify the owner of one of the compiled computer binaries of the
potential infringement via output 220 (step 790). The notification
can include details of the regions of the allegedly infringing
computer binary that is most likely involved in the
infringement.
[0033] The collection of compiled computer binaries, fingerprint
generation, and fingerprint matching can be automated and scheduled
to execute using any type of task scheduling technique. Thus, the
present invention provides a particularly cost- and time-effective
way to discover, remediate, and enforce intellectual property
rights, and accordingly acts as a deterrence against the theft of
software code. Further, by identifying infringement based on the
functions contained within compiled computer binaries, the present
invention can identify an entirely copied compiled computer binary,
as well as copied portions of a compiled computer binary.
[0034] FIG. 8 is a flow diagram of an exemplary fingerprinting
technique 800 in accordance with another embodiment of the present
invention, which is performed using the system of FIG. 2.
Initially, processor 205 receives a binary via input 215 and/or
memory 210 (step 805) and disassembles the binary in order to
generate a control flow graph for each function in the binary (step
810). Processor 205 selects a control flow graph for a first
function (step 815), reverses all of the edges of the selected
control flow graph and performs a block rank algorithm on the
selected control flow graph in order to generate a block rank score
for each block in the control flow graph of the selected function
(step 820). The block rank algorithm is similar to the PageRank
algorithm disclosed in the article "The PageRank Citation Ranking:
Bringing Order to the Web" by L. Page et al. 1999, the entire
disclosure of which is herein expressly incorporated by reference.
Those skilled in the art will recognize that the PageRank algorithm
is a type of Eigenvector centrality algorithm that assigns a value
corresponding to the relative importance of each block in a graph.
In PageRank the blocks are web pages and the relative importance is
based on the number of forward links from the page, as well as the
relative importance of the web pages associated with these forward
links. In the present invention the blocks within a function and
the relative importance of the blocks is based on the number of
incoming edges (because the edges are reversed), as well as the
relative importance of the blocks from which the incoming edges
originate.
[0035] The application of block rank scoring to an exemplary,
simplified function is illustrated in FIG. 9, which illustrates a
function with the edges reversed going from the end of the function
to the start. In this exemplary function block 906 has a block rank
score of 3 because block 916 has only a single forward path and
therefore contributes its entire block rank score of 2 and block
914 has a block rank score of 2, which is divided between blocks
904 and 906. Block 904 has forward links from blocks 908, 910, and
912, each of which have only a single forward path and each having
a block rank score of 1, and a forward link from block 914, which
splits its block rank score between blocks 904 and 906. Thus, block
904 is assigned a value of 4. Finally, block 902 is assigned a
block rank score of 7 because blocks 904 and 906 only have a single
forward path, and therefore contribute their entire respective
block rank scores of 4 and 3.
[0036] Once a relative importance value is assigned to each block
within the control flow graph for the first function, a path is
constructed starting from the function start based on the block
rank score. Specifically, processor 205 begins path construction at
the function start (step 825) and identifies the blocks connected
to the function start in the control flow graph and the block rank
score for each connected block (step 830). Processor 205 then
constructs the path to a next block by selecting the one of the
identified blocks having a highest block rank score (step 835).
When there is more than one of the identified blocks having the
same highest block rank score then the block with the larger number
of instructions is selected as the next block. If there is still a
tie, an arbitrary block is chosen among the blocks having the same
highest block rank score.
[0037] For each selected block in the path processor 205 saves a
3-tuple in memory 210 as an ordered list based on the order of the
blocks in the path (step 840). The 3-tuple comprises the number of
instructions in the block, the in-degree of the block, and the
out-degree of the block. Next processor 205 determines whether
there are any out paths from the selected block (step 845). When
processor determines there are out paths ("Yes" path out of
decision step 845) then processor 205 uses the block rank score to
select a next connected block (step 835).
[0038] The path construction continues until a block is reached
that does not have any out paths ("No" path out of decision step
845). Once the end block is reached processor 205 generates the
fingerprint of the selected function by storing the ordered list of
3-tuples in memory 210 along with the control flow graph block
count, the control flow graph edge count, and the function call
count (step 850). Each element of the ordered list corresponds to
each block along the chosen characteristic path (or the Markov
chain referred to above).
[0039] Returning to the exemplary function of FIG. 9, the path
would be constructed from block 902 to block 904 because block 904
has a higher block rank value than block 906. The path would then
continue from block 904 to block 914 because block 914 has the
highest block rank value of any of the outgoing paths from block
904. Thus, it should be appreciated that the ordered list does not
include all blocks of the function, only those along the path.
Accordingly, this reduces the amount of information required to
generate a fingerprint, and in turn improves the overall
functioning of a computer executing this method because the overall
processing power is reduced.
[0040] Returning to FIG. 8, next processor 205 determines whether
there are any remaining functions in the binary that have not yet
been processed (step 855). If processor 205 determines there are
any remaining functions ("Yes" path out of decision step 855),
processor 205 selects a next function (step 860) and the function
is processed in the same manner as the first selected function
(steps 820-850). When processor 205 has processed all functions of
the binary ("No" path out of decision step 855) the fingerprints of
each of the functions are saved in memory 210 along with an
association with the received binary from which the functions
originated (step 865).
[0041] Although FIG. 8 involves fingerprinting functions of a
single set of code it should be recognized that this can be
performed on additional sets of code so that the fingerprint
database is populated with functions from more than one set of
code. For example, a software developer could fingerprint all of
its code so that the database can later be used to identify copied
functions from third-party code. Similarly, an entity can offer a
service in which code from one or more software developers is
collected, for example, by a web spider as described above, is
stored in the database and a software developer can bring its code
to the entity to determine whether the database includes any code
containing copied functions.
[0042] FIGS. 10A and 10B are a flow diagrams of an exemplary
process 1000 for matching fingerprints in accordance with another
embodiment of the present invention, which is performed using the
system of FIG. 2. Because the database of fingerprinted functions
is generated using disassembled complied binaries the fingerprinted
functions may contain one or more blocks that are added by the
complier during compilation of the source code. Accordingly, a
function may have been copied from another function even though two
functions are not identical, i.e., there may not be an exact
correspondence between the blocks and/or paths of one function and
those of another function. These types of discrepancies are
addressed in the method of FIGS. 10A and 10B, and thus the
fingerprint matching of the present invention can identify
functions containing copied code even when the functions are not
identical.
[0043] Accordingly, when one desires to determine whether a
function has been copied, processor 205 receives the function as a
query function (step 1002) and fingerprints it (step 1004) using
the method detailed above in connection with FIG. 8. Next processor
205 determines a neighborhood size for identifying potentially
copied functions (step 1006) using the following formula:
Neighborhood_Size=Eclidean_distance([0,0,0],
coordinate_triple_of_query_function)*0.20. [0044] where
coordinate_triple_of_query_function comprises the (1) control flow
graph block count, (2) control flow graph edge count, and (3)
function call count of the query function.
[0045] The value 0.20 is a weighting factor controlling the size of
the neighborhood, which in this instance is a 20% weighting factor.
It should be recognized that other weighting factors could be
employed, as well as no weighting factor, depending upon the query
function. The weighting factor adjustment depends on the density of
the distribution of the measurements of the function features,
specifically, how many other functions within the call graph have
similar feature metrics necessitating an expansion or contraction
of the definition of a function's neighborhood.
[0046] Now that the search neighborhood has been defined, processor
205 can search a database of fingerprints, which can be stored in
memory 210, to identify those within the search neighborhood and
then select the fingerprint of a first function in the neighborhood
(step 1008). Processor then initiates a comparison of the selected
neighborhood fingerprint corresponding 3-tuple to the 3-tuple of
the fingerprint of the query function (step 1010). Processor 205
begins the comparison by independently normalizing each dimension
of the query fingerprint and the selected neighborhood fingerprint
by dividing the distance from the mean for each dimension of all
fingerprints in the neighborhood by the standard deviation for each
dimension of all fingerprints in the neighborhood (step 1012).
[0047] Processor 205 calculates a similarity score between the
query fingerprint and the selected neighborhood fingerprint using
the Needleman-Wunsch algorithm (step 1014), which is described in
the article "A general method applicable to the search for
similarities in the amino acid sequence of two proteins" S.
Needleman et al., Journal of Molecular Biology 48 (3): 443-53.
doi:10.1016/0022-2836 (70)90057-4. PMID 5420325, the entire
disclosure of which is herein expressly incorporated by
reference.
[0048] The details of the application of the Needleman-Wunsch
algorithm in the present invention are illustrated in FIG. 10B,
which will be described in connection with FIGS. 11A-11C. First,
processor 205 generates a two-dimensional array
A.sub.j.times.B.sub.i of all possible pair combinations of 3-tuples
between the query fingerprint and the selected neighborhood
fingerprint (step 1016), where, starting from the beginning of the
ordered list of 3-tuples, A.sub.j represents jth block in the
ordered list for the selected neighborhood function the and B.sub.i
represents the ith block in the ordered list for the query
function. The assignment of the neighborhood function to A.sub.j
and the query function to B.sub.i is not critical and the
assignments can be reversed. Each cell of the array is assigned a
value of "1" if there is a match between the 3-tuples for the query
fingerprint and the selected neighborhood fingerprint and a "0" if
there is no match. The array allows a representation of every
possible comparison between the blocks of the two fingerprints by
generating pathways through the array and the pathway having the
sum of assigned cell values that is the largest is selected as the
similarity score.
[0049] Next, processor 205 initializes the indices i and j to
correspond to the last cell in the array (step 1018), which in this
example is setting i=y and j=z. Processor 205 then performs a
successive summation procedure according to steps 1020-1028, in a
row-by-row manner starting from the right side of each row.
Specifically, for a cell under evaluation at i=y and j=z, processor
205 adds its value to the maximum value of the cells having indices
i=y-1 and/or j=z-1 (step 1020). For example, referring now to FIGS.
11A and 11B, assume that the highlighted cell at indices 4, 4 is
under evaluation and that it was already assigned a value of 1 due
to a match between 3-tuples of the query fingerprint and the
selected neighborhood fingerprint. In FIG. 11A the cells in row 4
to the right of the cell at indices 4, 4 have already been subject
to the successive summation, as well as all of the cells in rows
1-3. The cells having an index of 3 (i.e., 4-1) are evaluated to
identify the maximum value, which in this example would be the
value of 2 in the cell at 3, 3. Accordingly, as illustrated in FIG.
11B, the value of 2 is added to the value of 1 already at cell 4,
4, and thus the successive summation results in a value of 3 at
index 4, 4.
[0050] Returning to FIG. 10B, after adding the value to a cell,
processor 205 determines whether the evaluated cell is the last
cell in a row (step 1022). If not ("No" path out of decision step
1022), processor 205 increments the column index by one (step 1024)
and the next cell in the row is subjected to the successive
summation (step 1020). If processor 205 determines that the
evaluated cell is the last cell of the row ("Yes" path out of
decision step 1022), then processor 205 determines whether the
evaluated cell is the last cell of all of the rows (step 1026). If
processor 205 determines there are remaining rows ("No" path of
decision step 1026), then processor 205 increments the row index by
one and resets the column index to the most right-hand column index
(step 1028) and processor 205 continues the successive summation in
the first column of the next row (step 1020). The result of the
successive summation across all cells is illustrated in FIG.
11C.
[0051] If processor 205 determines that all cells have been
evaluated ("Yes" path out of decision step 1026), then processor
205 selects, based on the Manhattan distance between each of the
3-tuples, a pathway having the largest sum of the assigned cell
values, less any gap penalty, as the maximum match pathway (step
1030). In an exemplary embodiment the gap penalty is 3. However, it
will be recognized that other gap penalties can be used. The value
in each cell in the outer row or outer column contains a value of
the maximum number of matches that can be obtained by starting a
pathway at that cell, and the cell in the outer row and column are
evaluated identify the cell having the largest value, which is
where the maximum match pathway begins. The maximum match pathway
then continues from this outer cell to the cell in each successive
row or column having a largest value. Referring again to FIG. 11C,
the largest number in the outer row or column is at index 6, 6,
which has a value of 4. From index 6, 6 the pathway continues to
the cell having the maximum value in each row of each successive
column, which is illustrated by the arrows in the figure.
[0052] Returning to FIG. 10A, once processor 205 identifies the
maximum match pathway (step 1030), then processor 205 uses the sum
of all of the cells along the pathway as the similarity score,
which is added to a similarity score ranking list, which is ordered
by score from closest to further (step 1034). If processor 205
determines there are functions in the neighborhood that have not
yet been compared ("Yes" path out of decision step 1034), then
processor 205 selects the next neighborhood function (step 1036)
and subjects it to the comparison (step 1010). If processor 205
determines that all neighborhood functions have been subject to the
comparison ("No" path out of decision step 1034), then processor
205 identifies any function having a similarity score greater than
or equal to a threshold similarity score as being a function having
code copied by the query function (step 1038). The threshold can be
set, for example, by training the system using known data and a
particular compiled computer binary for which it is to be
determined whether there are other compiled computer binaries
infringing the particular compiled computer binary. This training
identifies commonalities between the known data and the particular
compiled computer binary so that the threshold can be set to avoid
false positives indicating infringement due to code commonly used
across different pieces of software that would not be an indicator
of infringement.
[0053] Although FIGS. 10A and 10B involve a comparison between the
query function and a plurality of functions stored in a database of
fingerprints, the comparison technique can also be employed to
compare two functions to identify whether one contains copied code
of the other. In this case the selection of the neighborhood size
(step 1006) could be eliminated along with the iteration through
the functions in the neighborhood (the return path from step 1036
to 1010). Alternatively, the neighborhood size (step 1006) could be
maintained as an initial threshold to quickly eliminate the
possibility of copying when one of the fingerprints is outside of
the neighborhood. In this case the iteration through functions in
the neighborhood would still be eliminated because the comparison
involves only two functions.
[0054] The fingerprint generation and matching techniques of the
present invention advantageously improve the operation of the
computer performing these techniques by reducing the processing
load required to fingerprint code and compare fingerprints.
[0055] Although exemplary embodiments have been described above as
generating fingerprints using compiled computer binaries, the
present invention is equally applicable to computer source code,
byte code, and the like.
[0056] The foregoing disclosure has been set forth merely to
illustrate the invention and is not intended to be limiting. Since
modifications of the disclosed embodiments incorporating the spirit
and substance of the invention may occur to persons skilled in the
art, the invention should be construed to include everything within
the scope of the appended claims and equivalents thereof.
* * * * *