Patent Grant
8516449
This application is based upon and claims priority from prior U.S. patent application Ser. No. 12502519, filed on Jul. 14, 2009, now [pending], the entire disclosure of which is herein incorporated by reference.
Not Applicable
Web applications are typically written in a combination of several programming languages (e.g., JavaScript on the client side, and PHP with embedded SQL commands on the server side), and generate structured output in the form of dynamically generated HTML pages that may refer to additional scripts to be executed. Since the application is built using a complex mixture of different languages, programmers may inadvertently make mistakes and introduce faults in the applications, resulting in web application crashes and malformed dynamically-generated HTML pages that can seriously impact usability.
Moreover, it is difficult to find errors and faults, especially errors and faults that lead to security vulnerabilities in a client-server environment. One reason for this difficulty is that a server application may dynamically generate HTML, JavaScript and other client code that violates information flow.
Disclosed is an automated detection of information-flow security problems in computer software applications that comprise multiple tiers and in which tiers dynamically generate code that could be executed in different tiers, causing both information and code to flow between tiers. Once a problem has been identified, this invention reports not only where the problem occurred, but also which tier was responsible for generating the faulty code, and which portion of the generating code should be fixed so that the generated code will no longer exhibit that vulnerability. This process is performed with a combination of static and/or dynamic analysis of the code of the various tiers that comprise the application under analysis.
The present invention overcomes many problems associated with automatic fault detection and localization in dynamic web applications. The present invention provides a system, computer program product, on-demand service, and a computer implemented method for analyzing a set of two or more communicating applications. The method includes executing a first application, such as a client application, and executing a second application, such as a server application, wherein the first application and the second application are communicating with each other. The communications can include message passing between the client application and the server application. Also the communications can in response to the server application causing a change in information flow with the client application. In still another embodiment, a portion of the code fragments as part of the first application and/or the second application are dynamically executed and/or statistically analyzed. The first application and the second application are communicating data and/or control with each other.
Static code analysis is the analysis of code fragments as part of an application that is performed without actually executing programs built from that software. The term is usually applied to the analysis performed by an automated tool, with human analysis being called program understanding, program comprehension or code review.
In contrast, analysis performed on executing code fragments is known as dynamic analysis or dynamically executed. Dynamic analysis is the analysis of an application that is performed by executing fragments built from that software system on a real or virtual processor. For dynamic program analysis to be effective, the target program must be executed with sufficient test inputs to produce interesting behavior. Use of software testing techniques such as code coverage helps ensure that an adequate slice of the program's set of possible behaviors has been observed. Also, care must be taken to minimize the effect that instrumentation has on the execution (including temporal properties) of the target program.
A correlation is recorded between the first application and the second application that have been executed and an execution characteristic that the code fragments exhibited on execution. An oracle is used to determine an analysis of the first application that has been executed. The oracle may be used to determine an analysis of the second application as well. In one embodiment, the execution of the first application causes a change of state in the second application and/or a change control flow in the second application. Code fragments in the first application and/or the second application are prioritized based on an evaluation produced by the oracle, and based on the correlation between the code fragments that have been executed and execution characteristics exhibited by these code fragments. In another embodiment, an analysis using a dynamic and/or static oracle of at least a portion of the code fragments that comprise the first application that have been executed.
Various security vulnerabilities and security related information can be analyzed by the oracle. For example, the security related information may include a violation of information flow. Or the security related information includes violation of information flow such as untrusted information flowing into a trusted application. Another type of security information includes private data flowing into a public application.
Also described is a technique that generates additional user inputs based on execution information associated with path constraints and based on information from the oracle. Advantageously, the embodiment is useful in a test generation tool that generated many similar inputs when a failure-inducing input is found in order to enhance fault localization. Further, described is a computer-implemented flow for extending the existing idea of combined concrete and symbolic testing to applications that interact with persistent state.
The present invention leverages two existing techniques—combined concrete and symbolic execution, and the Tarantula algorithm for fault localization—to create a uniquely powerful method for finding and localizing faults. The present invention extends combined concrete and symbolic execution to the domain of dynamic web applications by automatically simulating user interaction. The method automatically discovers inputs required to exercise paths through a program, thus overcoming the limitation of many existing fault localization techniques that a test suite be available upfront. Shown is how the effectiveness of Tarantula can be significantly improved by utilizing a correlation between executed statements and their execution characteristics, in combination with an oracle or that detects where errors occur in the execution characteristics. The present invention is implemented in Apollo, a tool for testing PHP applications, using an HTML validator as our oracle. When applied to a number of open-source PHP applications, Apollo found, and precisely localized, a significant number of faults.
The subject matter, which is regarded as the invention, is particularly pointed out and distinctly claimed in the claims at the conclusion of the specification. The foregoing and other features and also the advantages of the invention will be apparent from the following detailed description taken in conjunction with the accompanying drawings.
It should be understood that these embodiments are only examples of the many advantageous uses of the innovative teachings herein. In general, statements made in the specification of the present application do not necessarily limit any of the various claimed inventions. Moreover, some statements may apply to some inventive features but not to others. In general, unless otherwise indicated, singular elements may be in the plural and vice versa with no loss of generality. In the drawing like numerals, refer to like parts through several views.
Overview of Approach
The present invention is an extension of our previous work and is directed to detecting and localizing security vulnerabilities in client-server software environment such as a violation of information flow, such as untrusted information flowing into a trusted application or private data flowing into a public application.
In our previous work [3], the technique of combined concrete and symbolic execution [4, 7, 8, 19, 22] was adapted to web applications written in PHP. In this approach, the application is first executed on an empty input, and a path condition is recorded that reflects the control flow predicates in the application that have been executed. By changing one of the predicates in the path condition, and solving the resulting condition, additional inputs can be obtained. Execution of the program on these inputs will result in additional control flow paths being exercised. This process is repeated until either there is sufficient coverage of the statements in the application or until the time budget is exhausted. For each execution, a determination is made if an execution error occurs, or if the generated HTML page is malformed, using an HTML validator as an oracle. This technique has been implemented in a tool called Apollo (version 1.0), and in previous experiments on 4 open-source PHP applications, Apollo 1.0 found a total of 214 failures [3].
The coverage achieved by Apollo 1.0 was limited, since Apollo 1.0 ignored changes to the state of the environment by the executed scripts. That is, each script was executed from a single initial environment state (usually a populated database). However, the desired execution characteristics of a PHP application are often only achieved by a series of interactions between the user and the server (e.g., a minimum of five inputs are needed from opening Amazon to buying a book). In the present invention combined concrete and symbolic testing is enhanced by supporting automatic dynamic simulation of user interactions, and implemented in a new version of our tool, Apollo 2.0. Apollo 2.0 records the environment state (database, sessions, cookies) after executing each script, analyzes the output of the script to detect the possible user options that are available, and restores the environment state before executing a new script based on a detected user option.
More importantly, the present invention determines where in the source code changes need to be made in order to fix these failures. This task is commonly referred to as fault localization, and has been studied extensively in the literature (see, e.g., [5, 11-13, 18, 25]). In the present invention the Tarantula fault localization technique by Jones et al. [11,12] is combined with concolic execution in order to perform fully automated failure detection and localization for web applications written in PHP. The Tarantula technique predicts for each statement, how suspicious it is in causing the error. This is calculated from the percentage of passing tests that execute the statement and the percentage of failing tests that execute the statement. From this, a suspiciousness rating is computed for each executed statement. Programmers are encouraged to examine the statements in order of decreasing suspiciousness, and this has been demonstrated to be quite effective in experiments with the Siemens suite [10] of versions of small C programs into which artificial faults have been seeded [11].
The use of combined concrete and symbolic execution to obtain passing and failing runs overcomes the limitation of Tarantula and many other existing fault localization techniques that a test suite with passing and failing runs be available up-front. Furthermore, the fact that PHP applications generate output in a format (HTML) that can be validated using an oracle (an HTML validator) enables us to enhance the effectiveness of fault localization. This is accomplished by maintaining, during program execution, an output mapping from statements in the program to the fragments of output they produce. This mapping, when combined with the report of the oracle that indicates what parts of the program's output are incorrect, provides an additional source of information about the possible location of the fault, and is used to fine-tune the suspiciousness ratings provided by Tarantula.
The contributions of the present invention are as follows:
Context: PHP and Web Applications
PHP is widely used for implementing Web applications, in part due to its rich library support for network interaction, HTTP processing and database access. A typical PHP web application is a client-server application in which data and control flows interactively between a server that runs PHP scripts and a client, which is usually a web browser. The PHP scripts that run on the server generate HTML that includes forms to invoke other PHP scripts, passing them a combination of user input and constant values taken from the generated HTML.
This section briefly reviews the PHP scripting language, and discusses the kinds of failures that may occur during the execution of a PHP application, focusing on those aspects of PHP that differ from mainstream languages.
The PHP Scripting Language
PHP is object-oriented, in the sense that it has classes, interfaces, and dynamically dispatched methods with syntax and semantics similar to that of Java. PHP also has features of scripting languages, such as dynamic typing, and an eval construct that interprets and executes a string value that was computed at run-time as a code fragment. For example, the following code fragment:
$code=“$X=3;”; $x=7; eval($code); echo $x;
prints the value 3 (names of PHP variables start with the $ character). Other examples of the dynamic nature of PHP are a predicate that checks whether a variable has been defined, and class and function definitions that are statements that may occur anywhere.
The code in
Failures in PHP Programs
In one embodiment, the present invention provides a technique that targets two types of failures that may occur during the execution of PHP applications and that can be automatically detected:
Fault Localization
Detecting failures only demonstrates that a fault exists; the next step is to find the location of the fault that causes each failure. There are at least two pieces of information that might help:
PHP Example
These fragments are part of the client-server work flow in a Web application: the user first sees the index.php page of
The example program contains an error in the HTML produced for the administrative details: the H2 tag that is opened on line 21 of
The next step is to determine the cause of the malformed HTML. Consider the two sources of information suggested in the section “Fault Localization”:
Neither of these estimates is precise, since the fault is clearly in the printing of the H2 line itself (line 21). Combining, however, the results of the validator and the sets of statements. Specifically, observing that the printing of /BODY on line 27 in view.php occurs in both passing and failing executions, and is therefore unlikely to be the location of the fault. Furthermore, observing that lines 12 and 14, each of which is only executed in one of the executions, is not associated with the failure according to the information received from the oracle or code-validator. Therefore, it can be concluded that the fault is most closely associated with line 21 in view.php.
Combined Concrete and Symbolic Execution in the Presence of Interactive User Input
The technique of the present invention for finding failures in PHP applications is a variation on combined concrete and symbolic execution [4, 7, 8, 19, 22], a well-established test generation technique. The basic idea behind this technique is to execute an application on some initial (e.g., empty or randomly chosen) input, and then on additional inputs obtained by solving constraints derived from exercised control flow paths. Failures that occur during these executions are reported to the user.
In our previous paper [3], described is how this technique can be adapted to the domain of dynamic web applications written in PHP. The resulting Apollo 2.0 tool takes into account language constructs that are specific to PHP, uses an oracle to validate the output, and supports database interaction. However, prior art solutions relied on a manual solution for the challenging problem of interactive user input already described in the section “PHP and Web Applications”. PHP applications typically generate HTML pages that contain user-interface features such as buttons that—when selected by the user—result in the execution of additional PHP scripts. Modeling such user input is important, because coverage of the application will typically remain very low otherwise. In our previous paper [3], a manually performed program transformation was used that translates interactive user input into additional script parameters. This manual step had several limitations:
The present invention, replaces this manual step with an automatic method that (i) tracks changes to the state of the environment (i.e., session state, cookies, and the database) and (ii) performs an “on the fly” analysis of the HTML output produced by PHP scripts to determine what user options it contains, with their associated PHP scripts. By determining the state of the environment as it exists when an HTML page is produced, the environment in which additional scripts are executed as a result of user interaction is determined.
This is important because a script is much more likely to perform complex execution characteristics when executed in the correct context (environment). For example, if the web application does not record in the environment that a user is logged in, most scripts will present only vanilla information and terminate quickly (e.g., when the condition in line 5 of
Algorithm
More specifically, in
The algorithm uses a queue of tests. Each test contains the program component to execute, a path constraint which is a conjunction of conditions on the program's input parameters, and the environment state before the execution. The queue is initialized with one test for each of the components executable from the initial state, and the empty path constraint (lines 3-5). The algorithm uses a constraint solver to find a concrete input that satisfies a path constraint from the selected test (lines 7-9). The algorithm restores the environment state (line 11), then executes the program component concretely on the input and checks if failures occurred (lines 12-14). Any detected failure is merged into the corresponding bug report (lines 15-16). Next, the program is executed symbolically on the same input (line 17). The result of symbolic execution is a path constraint, ^i=lnci, that is fulfilled if the given path is executed (here, the path constraint reflects the path that was just executed). The algorithm then creates new test inputs by solving modified versions of the path constraint (lines 18-21) as follows. For each prefix of the path constraint, the algorithm negates the last conjunct (line 19). A solution, if it exists, to such an alternative path constraint corresponds to an input that will execute the program along a prefix of the original execution path, and then take the opposite branch. Finally, the algorithm analyzes the output to find new transitions from the new environment state (line 22). Each transition is expressed as a pair of path constraints and an executable component. The algorithm then adds new tests for each transition that was not explored before (line 23-27).
Algorithm Example
Illustrated now is the algorithm of
Iteration 1. The First Iteration of the Outer Loop (Lines 6-27)
removes that item from the queue (line 7), uses an empty input to satisfy the empty path constraint (line 9), restores the empty initial state (line 11), and executes the script (line 12).
No failures are observed, so the next few lines (line 13-16) do nothing. The call to executeSymbolic on line 17 returns an empty path constraint, so the function analyzeOutput on line 22 is executed next, and returns one user option; (login.php, Ø,Ø) for executing login.php with no input, and the empty state. This test is added to the queue (line 27).
iteration 2-5. The next iteration of the top-level loop dequeues the new work item, and executes login.php with empty input, and empty state. No failures are found. The call to executeSymbolic in line 17 returns a path constraint user≠admin ^user≠reg, indicating that the call to check_password on line 8 in
Given this, the loop at lines 18-21 will generate several new work items for the same script with the following path constraints user≠admin ^user≠reg, and user=admin which are obtained by negating the previous path constraint. The loop on lines 23-27 is not entered, because no user input options are found. After several similar iterations, two inputs are discovered: user=admin ^pw=admin, and user≠reg^pw=reg. These correspond to alternate control flows in which the check_password test succeeds.
iteration 6-7. The next iteration of the top-level loop dequeues an item that allows the check_password call to succeed (assume it selected user=reg . . . ). Once again, no failures are observed, but now the session state with user and pw set is recorded at line 13. Also, this time analyzeOutput (line 22) finds the link to the script in
The next iteration of the top-level loop dequeues one work item. Assume that it takes the last one described above. Thus, it executes the script in
iteration 8-9. The next loop iteration takes that last work item, containing a user and password pair for which the call to check.password succeeds, with the user name as ‘admin’. Once again, no failures occur, but now the session state with user, pw and type set is recorded at line 13. This time, there are no new inputs to be derived from the path constraint, since all prefixes have been covered already. Once again, parsing the output finds the link to the script in
There are a few other things that happen, but at this point it should be noted that one successful and one failing execution for the script in
Fault Localization
In this section, reviewed first is the Tarantula fault localization technique. Next, an alternative technique is presented that is based on the output mapping and positional information obtained from an oracle. Finally, presented is a technique that combines the former with the latter.
Tarantula
Jones et al. [11, 12] presented Tarantula, a fault localization technique that associates with each statement a suspiciousness rating that indicates the likelihood that it contributes to a failure. Note line numbers are used to identify statements, because that enables us to present the different fault localization techniques in a uniform manner. The suspiciousness rating Star(l) for a statement that occurs at line3 l is a number between 0 and 1 that is defined as follows:
where Passed(l) is the number of passing executions that execute statement l, Failed(l) is the number of failing executions that execute statement l, TotalPassed is the total number of passing test cases, and TotalFailed is the total number of failing test cases. After suspiciousness ratings have been computed, each of the executed statements is assigned a rank, in order of decreasing suspiciousness. Ranks do not need to be unique: The rank of a statement l reflects the maximum number of statements that would have to be examined if statements are examined in order of decreasing suspiciousness, and if l were the last statement of that particular suspiciousness level chosen for examination.
Jones and Harrold [11] conducted a detailed empirical evaluation in which they apply Tarantula to faulty versions of the Siemens suite [10], and compare its effectiveness to that of several other fault localization techniques (see the section “Related Work”). The Siemens suite consists of several versions of small C programs into which faults have been seeded artificially. Since the location of these faults is given, one can evaluate the effectiveness of a fault localization technique by measuring its ability to identify these faults. In the fault localization literature, this is customarily done by reporting the percentage of the program that needs to be examined by the programmer, assuming statements are inspected in decreasing order of suspiciousness [1, 5, 11, 18].
Specifically, Jones and Harrold compute for each failing test run a score (in the range of 0%-100%) that indicates the percentage of the application's executable statements that the programmer need not examine in order to find the fault. This score is computed by determining a set of examined statements that initially contains only the statement(s) at rank 1. Then, iteratively, statements at the next higher rank are added to this set until at least one of the faulty statements is included. The score is now computed by dividing the number of statements in the set by the total number of executed statements. Using this approach, Jones and Harrold found that 13.9% of the failing test runs were scored in the 99-100% range, meaning that for this percentage of the failing tests, the programmer needs to examine less than 1% of the program's executed statements to find the fault. They also report that for an additional 41.8% of the failing tests, the programmer needs to inspect less than 10% of the executed statements.
Fault Localization Using the Output Mapping
An oracle that determines whether or not a failure occurs can often provide precise information about which parts of the output are associated with that failure. For instance, an HTML validator will typically report the location of malformed HTML. Such information can be used as a heuristic to localize faults in the program, provided that it can be determine which portions of the program produced which portions of the output. The basic idea is that the code that produced the erroneous output is a good place to start looking for the causative fault. This is formalized as follows. Assume the following two functions:
Given these two functions, a suspiciousness rating Smap(l) of the statement at line l for failure f is defined as follows:
Note that this is a “binary” rating: program parts are either highly suspicious, or not suspicious at all.
Combined Technique
The algorithm presented in the section “Tarantula” localizes failures based on how often statements are executed in failing and passing executions. However, in the web applications domain, a significant number of lines are executed in both cases, or only in failing executions. Thus, the fault localization technique presented in the section “Fault Localization using the Output Mapping” can be used to enhance the Tarantula results by giving a higher rank to statements that are blamed by both Tarantula and the mapping technique. More formally, a new suspiciousness rating Scomb(l) for the statement at line l is defined as follows:
Informally, given the suspiciousness rating 1.1 to any statement that is identified as highly suspicious by the oracle, and for which Tarantula indicates that the given line is positively correlated with the fault (indicated by the fact that Tarantula's suspiciousness rating is greater than 0.5).
Generating Inputs for Tarantula
As discussed previously, Tarantula computes suspiciousness ratings using a formula that considers how many times a statement is executed by passing and failing executions. But which passing executions and failing executions should be supplied as inputs to Tarantula?
To answer this question, assume that the algorithm of the section “Combined concrete and symbolic Execution in the Present of Interactive User Input” has exposed a number of failing executions. This set can be partitioned into subsets that pertain to the same failure. Here, two failures are assumed to be “equivalent” (i.e., due to the same fault) if the oracle produces the same message for them, and if the same program constructs are correlated with these messages according to the output mapping. In the section “Evaluation”, separate fault localization are conducted for experiments for each subset of equivalent failing executions.
This leaves the question of what set of passing executions should be supplied to Tarantula as inputs along with these failing executions. Currently two options are considered:
1. Supply all passing executions that were identified by the algorithm of the section “Combined concrete and symbolic Execution in the Present of Interactive User Input.”
2. Supply a randomly selected subset of 10% of the passing tests that were identified by the algorithm of the section “Combined concrete and symbolic Execution in the Present of Interactive User Input”.
Note that the above strategies can be applied to both the Tarantula and the combined algorithms.
Example
As described in the section “Example Algorithm”, the test input generation algorithm produced two runs of the script in
To understand how the Tarantula ratings are computed, consider statements that are only executed in the passing run. Such statements obtain a suspiciousness rating of 0/(1+0)=0.0. By similar reasoning, statements that are only executed in the failing run obtain a suspiciousness rating of 1/(0+1)=1.0, and statements that are executed in both cases obtain a suspiciousness rating of 1/(1+1)=0.5.
The suspiciousness ratings computed by the mapping based technique can be understood by examining the output of the validator in
The suspiciousness ratings for the combined technique follow directly from its definition in the section entitled “Combined Technique”.
As can be seen from the table, the Tarantula technique identifies lines 12 and 21 as the most suspicious ones, and the output mapping based technique identifies lines 21 and 27 as such. In other words, each of these fault localization techniques—when used in isolation—reports one nonfaulty statement as being highly suspicious. However, the combined technique correctly identifies only line 21 as the faulty statement.
Example Flow Diagrams
Turning now to
In
In step 410, the code fragments in the first application and/or the second application are prioritized in the application based on the evaluation produced by the oracle, and based on the correlation between the executed code fragments and the execution characteristics exhibited by those code fragments and the process ends at step 412.
In another embodiment, the evaluation produced by the oracle corresponds to security vulnerabilities in the generated application. Various security vulnerabilities and security related information can be analyzed by the oracle. For example, the security information may include a violation of information flow. For example, a violation of information flow is untrusted information flowing into a trusted application. Another example of a violation of information flow is private data flowing into a public application.
In one embodiment, the evaluation produced by the oracle is a fault localization evaluation. In another embodiment, the execution characteristics exhibited by the executed code fragments is an output of the application, and the evaluation produced by the oracle corresponds to errors in at least one part of this output. The execution characteristics in another embodiment is a creation of at least one new generated application. Still, in another embodiment, the execution characteristics include a characteristic of performance of the code fragments. Further, in one embodiment, the evaluation produced by the oracle corresponds to errors in the generated application.
Further, in one embodiment, the execution characteristics of the executed code fragments is an output of the application, and the evaluation produced by the oracle corresponds to errors in at least one part of this output. The execution characteristics in another embodiment is a creation of at least one new generated application. Still, in another embodiment, the execution characteristics of the executed code fragments is a characteristic of performance of the code fragments. Further, in one embodiment, the evaluation produced by the oracle corresponds to errors in the generated application. And in another embodiment, the evaluation produced by the oracle corresponds to security vulnerabilities in the generated application.
Turning now to
In
In one embodiment, the execution characteristics of the executed code fragments is an output of the application, and the evaluation produced by the oracle corresponds to errors in at least one part of this output. In another embodiment, execution characteristics of the executed code fragments is a characteristic of performance of the executed code fragments. Moreover, in another embodiment, the execution information associated with code fragments that have been executed consists of path constraints. In still another embodiment, the execution characteristics of the executed code fragments is a creation of at least one new generated application. Further, the evaluation produced by the oracle corresponds to errors in the generated application. Still, further, in another embodiment, the evaluation produced by the oracle corresponds to security vulnerabilities in the generated application.
Turning now to
In
In another embodiment, the execution information associated with code fragments that have been executed consists of path constraints. In another embodiment, the examining of the execution characteristics of the executed code fragments is a creation of at least one new generated application, and where the examining of the execution characteristics involves performing a program analysis to detect security vulnerabilities in this generated application.
Implementation
The Apollo 1.0 tool [3] has been extended with the algorithm for combined concrete and symbolic execution in the presence of interactive user input and persistent session state that was presented in the section “Combined concrete and symbolic Execution in the Presence of Interactive User Input”, and with the fault localization algorithm that was presented in the section “Fault Localization”. This section discusses some key features of the implementation.
interactive user input and session state. As was mentioned in the section “Combined concrete and symbolic Execution in the Presence of Interactive User Input”, it is important to determine what PHP scripts the user may invoke by selecting buttons, checkboxes, etc. in the HTML output of previously executed scripts. To this end, Apollo 2.0 automatically extracts the available user options from the HTML output. Each option contains the script to execute, along with any parameters (with default value if supplied) for that script. Apollo 2.0 also analyzes recursive static HTML documents that can be called from the dynamic HTML output, i.e. Apollo 2.0 traverses hyperlinks in the generated dynamic HTML that link to other HTML documents on the same site. To avoid redundant exploration of similar executions, Apollo 2.0 perform state matching (performed implicitly in Line 26 of
The use of session state allows a PHP application to store user supplied information on the server for retrieval by other scripts. The PHP interpreter has been enhanced to record when input parameters are stored in session state, to enable Apollo 2.0 to track constraints on input parameters in all scripts that use them.
web server integration. Apollo 1.0[3] only supported the execution of PHP scripts using the PHP command line interpreter. However, dynamic web applications often depend on information supplied by a web-server, and some PHP constructs are simply ignored by the command line interpreter (e.g., header). Apollo 2.0 supports execution through the Apache web-server in addition to the stand-alone command line executor. A developer can use Apollo 2.0 to silently analyze the execution and record any failure found while manually using the subject program on an Apache server.
Evaluation
This evaluation aims to answer two questions:
Subject Programs
For the evaluation, six open-source PHP programs (from http://sourceforge.net) have been selected, for which the characteristics are shown in
Coverage/Failures Detected
Apollo was run with and without the simulation of user interaction for 10 minutes on each subject program. This time limit was chosen arbitrarily, but it allows each strategy to generate hundreds of inputs and there is no reason to program believe that the results would be much affected by a different time limit. This time budget includes all experimental tasks. Line coverage was measured, i.e., the ratio of the number of executed lines to the total number of lines with executable PHP code that was shown in
Although the effect varies, it is clear that the user input simulation allows Apollo to achieve better results on all subject programs. For example, on phpsysinfo the effect on coverage is marginal (55.7% vs. 55.5%) because this program is not interactive. On the other hand, for phpBB2 the effect is significant for both the coverage obtained (28.0% vs. 11.4%), and for the number of failures detected (24 instead of 4), and similarly for timeclock because these applications only performs most operations when starting in the correct state (e.g., when the user is logged in).
Localizing Faults
To answer the second research question, localized faults database was created by manually localizing up to 20 faults in 3 of the subject programs (webchess contained only 9 faults that caused the 34 failures reported previously). The three fault localization methods were applied that were discussed in the section “Fault Localization” to each failure f: (i) one implementation of Tarantula (see the section “Tarantula”), (ii) a fault localization technique that uses only the output mapping (see the section “Fault Localization using the Output Mapping”), and (iii) a fault localization technique that combines Tarantula with the output mapping (see the section “Combined Technique”). As mentioned in the section “Generating Inputs”, two sets of inputs were tried for each technique: (a) the set of executions exposing fin combination with all passing executions, and (b) the set of executions exposing fin combination with 10% of randomly selected passing executions.
The effectiveness was measured of these fault localization algorithms as the minimal number of statements that need to be inspected until all the faulty lines are detected, assuming that statements are examined in order of decreasing suspiciousness (See the section “Tarantula”).
The results show that the combined technique is clearly superior to each of the Tarantula and mapping-based techniques that it builds upon. For webchess, the programmer would need to inspect 19.00% of the statements on average when Tarantula is supplied with all passing executions, 25.12% when the mapping-based technique is used, but only 6.94% using the combined technique. Using the same set of executions, the programmer needs to inspect 29.94% of schoolmate's statements using Tarantula, 15.06% using the output mapping, and only 5.09% using the combined technique. Similar results are obtained for timeclock. The use of a randomly selected subset of the passing tests yields slightly worse results for each of the techniques.
Related Work
This section discusses three categories of related work: (i) combined concrete and symbolic execution, (ii) testing of web applications, and (iii) fault localization.
DART [7] is a tool for finding combinations of input values and environment settings for C programs that trigger errors such as assertion failures, crashes and non-termination. DART combines random test generation with symbolic reasoning to keep track of constraints for executed control-flow paths. A constraint solver directs subsequent executions towards uncovered branches. CUTE [19] is a variation (called combined concrete and symbolic (concolic) testing) on the DART approach. The authors of CUTE introduce a notion of approximate pointer constraints to enable reasoning over memory graphs and handle programs that use pointer arithmetic.
Subsequent work extends the original approach of combining concrete and symbolic executions to accomplish two primary goals: 1) improving scalability [26], [28], [31], [32], [8], [35], and 2) improving execution coverage and fault detection capability through better support for pointers and arrays [4], [19], better search heuristics [8], [33], [34], or by encompassing wider domains such as database applications [30].
Godefroid [31] proposed a compositional approach to improve the scalability of DART. In this approach, summaries of lower level functions are computed dynamically when these functions are first encountered. The summaries are expressed as pre- and post-conditions of the function in terms of its inputs. Subsequent invocations of these lower level functions reuse the summary. Anand et al. [26] extend this compositional approach to be demand-driven to reduce the summary computation effort.
Exploiting the structure of the program input may improve scalability [32], [35]. Majumdar and Xu [35] abstract context free grammars that represent the program inputs to produce a symbolic grammar. This grammar reduces the number of input strings to enumerate during test generation.
Majumdar and Sen [34] describe hybrid combined concrete and symbolic testing, which interleaves random testing with bounded exhaustive symbolic exploration to achieve better coverage. Inkumsah and Xie [33] combine evolutionary testing using genetic mutations with combined concrete and symbolic testing to produce longer sequences of test inputs. SAGE [8] also uses improved heuristics, called white-box fuzzing, to achieve higher branch coverage
The PHP language that we used as example in this invention is quite different from the focus of previous testing research. PHP poses several new challenges such as dynamic inclusion of files, and function definitions that are statements. Existing techniques for fault detection in PHP applications use static analysis and target security vulnerabilities such as SQL injection or cross-site scripting (XSS) attacks [40], [42], [45], [50], [51]. In particular, Minamide [45] uses static string analysis and language transducers to model PHP string operations to generate potential HTML output—represented by a context free grammar—from the web application. This method can be used to generate HTML document instances of the resulting grammar and to validate them using an existing HTML validator. As a more complete alternative, Minamide proposes a matching validation which checks for containment of the generated context free grammar against a regular subset of the HTML specification. However, this approach can only check for matching start and end tags in the HTML output, while our technique covers the entire HTML specification. Also, flow-insensitive and context-insensitive approximations in the static analysis techniques used in this method result in false positives, while our method reports only real faults.
Kiezun et al. present a dynamic tool, Ardilla [43], to create SQL and XSS attacks. Their tool uses dynamic tainting, combined concrete and symbolic execution, and attack-candidate generation and validation. Like ours, their tool reports only real faults. However, Kiezun et al. focus on finding security faults, while we concentrate on functional correctness. Their tool builds on and extends the input-generation component of Apollo but does not address the problem of user interaction. It is an interesting area of future research to combine Apollo's user-interaction and state-matching with Ardilla's exploit-detection capabilities.
McAllister et al. [44] also tackle the problem of testing interactive web application. Their approach attempts to follow user interactions. Their method relies on pre-recorded traces of user interactions, while our approach automatically discovers allowable interactions. Moreover, their approach to handling persistent state relies on instrumenting one particular web application framework, Django. In contrast, our approach is to instrument the PHP runtime system and observe database interactions. This allows handling state of PHP applications regardless of any framework they may use.
Benedikt et al. [52] present a tool, VeriWeb, for automatically testing dynamic webpages. They use a model checker to systematically explore all paths (up to a certain bound) of user navigate in a web site. When the exploration encounters HTML forms, VeriWeb uses SmartProfiles. SmartProfiles are user-specified attribute-value pairs that are used to automatically populate forms and supply values that should be provided as inputs. Although VeriWeb can automatically fill in the forms, the human tester needs to pre-populate the user profiles with values that a user would provide. In contrast, Apollo automatically discovers input values by looking at the branch conditions along an execution path.
Dynamic analysis of string values generated by PHP web applications has been considered in a reactive mode to prevent the execution of insidious commands (intrusion prevention) and to raise an alert (intrusion detection) [41], [46], [49]. As far as we know, our work is the first attempt at proactive fault detection in PHP web applications using dynamic analysis. Finally, our work is related to implementation based (as opposed to specification based e.g., [47]) testing of web applications. These works abstract the application behavior using a) client-side information such as user requests and corresponding application responses [36], [38], or b) server-side monitoring information such as user session data [37], [48], or c) static analysis of server-side implementation logic [39]. The approaches that use client-side information or server-side monitoring information are inherently incomplete, and the quality of generated abstractions depends on the quality of the tests run.
Halfond and Orso [39] use static analysis of the server-side implementation logic to extract a web application's interface, i.e., the set of input parameters and their potential values. They implemented their technique for JavaScript. They obtained better code coverage with test cases based on the interface extracted using their technique as compared to the test cases based on the interface extracted using a conventional web crawler. However, the coverage may depend on the choices made by the test generator to combine parameter values—an exhaustive combination of values may be needed to maximize code coverage. In contrast, our work uses dynamic analysis of server side implementation logic for fault detection and minimizes the number of inputs needed to maximize the coverage. Furthermore, we include results on fault detection capabilities of our technique.
Early work on fault localization relied on the use of program slicing [21]. Lyle and Weiser [16] introduce program dicing, a method for combining the information of different program slices. The basic idea is that, when a program computes a correct value for variable x and an incorrect value for variable y, the fault is likely to be found in statements that are in the slice w.r.t. y, but not in the slice w.r.t. x. Variations on this idea technique were later explored by Pan and Spafford [17], and by Agrawal et al. [2].
In the spirit of this early work, Renieris and Reiss [18] use set-union and set-intersection methods for fault localization, so that they compare with their nearest neighbor fault localization technique (discussed below). The set-union technique computes the union of all statements executed by passing test cases and subtracts these from the set of statements executed by a failing test case. The resulting set contains the suspicious statements that the programmer should explore first. In the event that this report does not contain the faulty statement, Renieris and Reiss propose an SDG-based ranking technique in which additional statements are considered based on their distance to previously reported statements along edges in a System Dependence Graph [9]. The set-intersection technique identifies statements that are executed by all passing test cases, but not by the failing test case, and attempts to address errors of omission, where the failing test case neglects to execute a statement.
The nearest neighbors fault localization technique by Renieris and Reiss [18] assumes the existence of a failing test case and many passing test cases. The technique selects the passing test case whose execution spectrum most closely resembles that of the failing test case according to one of two distance criteria, and reports the set of statements that are executed by the failing test case but not by the selected passing test case. One similarity they use measures the distance between two test cases as the cardinality of the symmetric set difference between the statements that they cover. The other measure considers the differences in the relative execution frequencies. In the event that the report does not contain the faulty statement, Renieris and Reiss use the SDG-based ranking technique mentioned above to identify additional statements that should be explored next. Nearest Neighbor was evaluated on the Siemens suite [10], a collection of small C programs for which faulty versions and a large number of test cases are available, and was found to be superior to the set-union and set-intersection techniques.
Cleve and Zeller [5, 25] present a fault-localization technique based on Delta Debugging [24], a binary search and minimization technique. Delta debugging is first employed to identify the variables responsible for a failure, by selectively introducing values that occur in the program state of a failing run into the state obtained during a passing run, and observing whether or not the failure reoccurs. Then, delta debugging is applied again in order to identify cause transitions, i.e., points in the program where one variable ceases to be the cause for a failure, and where another variable starts being the origin of that failure. Cleve and Zeller report finding a real failure in GCC using the technique, and also evaluate their work on the Siemens suite.
Dallmeier et al. [6] present a fault localization technique in which differences between method call sequences that occur in passing and failing executions are used to identify suspicious statements. They evaluate the technique on buggy versions of the NanoXML Java application.
Two recent papers by Jones and Harrold [11] and by Abreu et al. [1] present empirical evaluations of several fault localization techniques, including several of the techniques discussed above, using the Siemens suite. Yu et al. [23] evaluated the sensitivity of several of the fault localization techniques discussed above to test suite reduction. Here, the goal was to determine to what extent the effectiveness of fault localization techniques was reduced as a result of applying several test-suite minimization techniques.
Other fault localization techniques analyze statistical correlations between control flow predicates and failures (see, e.g., [14, 15]), and correlations between changes made by programmers and test failures [20].
In the present invention, the Tarantula technique is applied in a different domain (open-source web applications written in PHP instead of C programs), and adapted to take into account positional information that was obtained from the PHP interpreter. Instead of using artificially seeded faults such as the ones in the Siemens suite, real faults are studied that were exposed by our Apollo 2.0 tool. Moreover, the present invention does not use an existing test suite but relies on Apollo 2.0 to generate a large number of (passing and failing) test cases instead.
Conclusions
The present invention provides a method and system for failure detection and fault localization that leverages combined concrete and symbolic execution [4, 7, 8, 19, 22] and the Tarantula algorithm [11, 12] to automatically find and localize failures in PHP web applications.
The present invention adapts combined concrete and symbolic execution to the domain of web applications by performing dynamic simulation of user interaction in different environment states. Unlike previous fault localization methods, ours does not require a test-suite with passing and failing test cases to be available up front. An output mapping is used between PHP statements and the output they produce in combination with positional information about HTML errors obtained from the oracle to improve on Tarantula's fault localization.
The technique is implemented in Apollo 2.0. In experiments on 6 open-source PHP applications, discovered is our new automatic method for simulating user input significantly improved line coverage and the number of failures found. Discovered also was that a fault localization technique that combines Tarantula with information retrieved from the output mapping is significantly more precise than either Tarantula or the output mappings alone.
The main topic for future work is to explore the use of combined concrete and symbolic execution to generate passing test cases that are highly similar to failing test cases, to further improve the effectiveness of Tarantula.
On Demand Deployment
Detecting and localizing security vulnerabilities in client-server application, in one embodiment, is implemented in an on-demand environment. This on demand embodiment provides a shared architecture to simultaneous serve multiple customers in a flexible, automated fashion. It is standardized, requiring little customization and it is scalable, providing capacity on demand in a pay-as-you-go model.
The process software can be stored on a shared file system accessible from one or more servers. The process software is executed via transactions that contain data and server processing requests that use CPU units on the accessed server. CPU units are units of time such as minutes, seconds, hours on the central processor of the server. Additionally the accessed server may make requests of other servers that require CPU units. CPU units are an example that represents but one measurement of use. Other measurements of use include but are not limited to network bandwidth, memory usage, storage usage, packet transfers, complete transactions etc.
When multiple customers use the same process software application, their transactions are differentiated by the parameters included in the transactions that identify the unique customer and the type of service for that customer. All of the CPU units and other measurements of use that are used for the services for each customer are recorded. When the number of transactions to any one server reaches a number that begins to affect the performance of that server, other servers are accessed to increase the capacity and to share the workload. Likewise when other measurements of use such as network bandwidth, memory usage, storage usage, etc. approach a capacity so as to affect performance, additional network bandwidth, memory usage, storage etc. are added to share the workload.
The measurements of use used for each service and customer are sent to a collecting server that sums the measurements of use for each customer for each service that was processed anywhere in the network of servers that provide the shared execution of the process software. The summed measurements of use units are periodically multiplied by unit costs and the resulting total process software application service costs are alternatively sent to the customer and or indicated on a web site accessed by the customer which then remits payment to the service provider.
In another embodiment, the service provider requests payment directly from a customer account at a banking or financial institution.
In another embodiment, if the service provider is also a customer of the customer that uses the process software application, the payment owed to the service provider is reconciled to the payment owed by the service provider to minimize the transfer of payments.
Non-Limiting Hardware Embodiments
Computer System
Overall, the present invention can be realized in hardware or a combination of hardware and software. The processing system, according to a preferred embodiment of the present invention can be realized in a centralized fashion in one computer system, or in a distributed fashion where different elements are spread across several interconnected computer systems and image acquisition sub-systems. Any kind of computer system—or other apparatus adapted for carrying out the methods described herein—is suited. A typical combination of hardware and software is a general-purpose computer system with a computer program that, when loaded and executed, controls the computer system such that it carries out the methods described herein.
An embodiment of the processing portion of the present invention can also be embedded in a computer program product, which comprises all the features enabling the implementation of the methods described herein, and which—when loaded in a computer system—is able to carry out these methods. Computer program means or computer programs in the present context mean any expression, in any language, code or notation, of a set of instructions intended to cause a system having an information processing capability to perform a particular function either directly or after either or both of the following a) conversion to another language, code or, notation; and b) reproduction in a different material form.
A computer system may include, inter alia, one or more computers and at least a computer readable medium, allowing a computer system to read data, instructions, messages or message packets, and other computer readable information from the computer readable medium. The computer readable medium may include non-volatile memory, such as ROM, flash memory, disk drive memory, CD-ROM, and other permanent storage. Additionally, a computer readable medium may include, for example, volatile storage such as RAM, buffers, cache memory, and network circuits 1112 connected to network 1138. Furthermore, the computer readable medium may comprise computer readable information in a transitory state medium such as a network link and/or a network interface, including a wired network or a wireless network, that allow a computer system to read such computer readable information.
An example of a computer system 1100 is shown in
Computer system 1100 includes a display interface 1110 that forwards graphics, text, and other data from the communication infrastructure 1102 (or from a frame buffer not shown) for display on the display unit 1120. Computer system 1100 also includes a main memory 1106, preferably random access memory (RAM), and optionally includes a secondary memory 1112. The secondary memory 1108 includes, for example, a hard disk drive 1116 and/or a removable storage drive 1118, representing a floppy disk drive, a magnetic tape drive, an optical disk drive, etc. The removable storage drive 1116 reads from and/or writes to a removable storage unit 1118 in a manner well known to those having ordinary skill in the art. Removable storage unit 1118, represents a floppy disk, magnetic tape, optical disk, etc. which is read by and written to by removable storage drive 1116. As will be appreciated, the removable storage unit 1118 includes a computer usable storage medium having stored therein computer software and/or data.
In alternative embodiments, the secondary memory 1112 includes other similar means for allowing computer programs or other instructions to be loaded into computer system 1100. Such means include, for example, a removable storage unit 1118 and an interface 1108. Examples of such include a program cartridge and cartridge interface (such as that found in video game devices), a removable memory chip (such as an EPROM, or PROM) and associated socket, and other removable storage units 1116 and interfaces 1108 which allow software and data to be transferred from the removable storage unit 1118 to computer system 1100.
N-Tiered Architecture
Referring to
This multi-tiered system has evolved from a more conventional system architecture in which clients retrieve information from a database, process the data according to instructions from a user, and store the data in the database. The clients in the conventional system architecture have three types of computer instructions installed and running on them to process information: code for the user interface (displaying buttons and lists of data), code for interacting with the database to fetch or store data, and code that processes the fetched data according to commands from the user interface or business logic. In contrast, in the multi-tiered system architecture, the client may contain only user interface code. The code for interacting with the database and processing the data is installed and operating on a middle-tier of servers such as application servers of
There are a variety of ways of implementing this middle tier, such as transaction processing monitors, message servers, or application servers. The middle tier can perform queuing, application execution, and database staging. For example, if the middle tier provides queuing, the client can deliver its request to the middle layer and disengage because the middle tier will access the data and return the answer to the client. In addition, the middle tier adds scheduling and prioritization for work in progress.
The exemplary web server 1204 of
The system of
The system of
The system of
Some caution is advised in use of the terms ‘client’ and ‘server’ because whether a particular computer acts as a client or a server depends upon role. In the system of
The system of
The arrangement of servers and other devices making up the exemplary system illustrated in
Although specific embodiments of the invention have been disclosed, those having ordinary skill in the art will understand that changes can be made to the specific embodiments without departing from the spirit and scope of the invention. The scope of the invention is not to be restricted, therefore, to the specific embodiments. Furthermore, it is intended that the appended claims cover any and all such applications, modifications, and embodiments within the scope of the present invention.
Each of the following fifty-two references are hereby incorporated by reference in their entirety.
| Number | Name | Date | Kind |
|---|---|---|---|
| 6173440 | Darty | Jan 2001 | B1 |
| 6442694 | Bergman et al. | Aug 2002 | B1 |
| 6701460 | Suwandi et al. | Mar 2004 | B1 |
| 6707795 | Noorhosseini et al. | Mar 2004 | B1 |
| 6823507 | Srinivasan et al. | Nov 2004 | B1 |
| 6895551 | Huang et al. | May 2005 | B1 |
| 6928449 | Ten-Hove et al. | Aug 2005 | B2 |
| 7096074 | Yulevitch et al. | Aug 2006 | B2 |
| 7131037 | LeFaive et al. | Oct 2006 | B1 |
| 7379994 | Collazo | May 2008 | B2 |
| 7647528 | Yilmaz et al. | Jan 2010 | B2 |
| 7665066 | Sauer et al. | Feb 2010 | B2 |
| 7681182 | Mistry et al. | Mar 2010 | B1 |
| 7720662 | Aldrich | May 2010 | B1 |
| 7926114 | Neystadt et al. | Apr 2011 | B2 |
| 8015139 | Bahl et al. | Sep 2011 | B2 |
| 8051330 | Cinato et al. | Nov 2011 | B2 |
| 8065554 | Herscovitz et al. | Nov 2011 | B2 |
| 8065565 | Chilimbi et al. | Nov 2011 | B2 |
| 8108513 | Kiciman et al. | Jan 2012 | B2 |
| 8132180 | Agarwal et al. | Mar 2012 | B2 |
| 8151146 | Ostrand et al. | Apr 2012 | B2 |
| 8356287 | Tzoref et al. | Jan 2013 | B2 |
| 8392515 | Kakivaya et al. | Mar 2013 | B2 |
| 8407673 | Terpolilli | Mar 2013 | B2 |
| 20040181710 | Sauer et al. | Sep 2004 | A1 |
| 20050193281 | Ide et al. | Sep 2005 | A1 |
| 20070016894 | Sreedhar | Jan 2007 | A1 |
| 20070038896 | Champlin et al. | Feb 2007 | A1 |
| 20070083933 | Venkatapathy et al. | Apr 2007 | A1 |
| 20070179746 | Jiang et al. | Aug 2007 | A1 |
| 20080016115 | Bahl et al. | Jan 2008 | A1 |
| 20080172576 | Kusko et al. | Jul 2008 | A1 |
| 20080276228 | Sreedhar | Nov 2008 | A1 |
| 20090113399 | Tzoref et al. | Apr 2009 | A1 |
| 20090248901 | Medin, Jr. | Oct 2009 | A1 |
| 20090249298 | Blount et al. | Oct 2009 | A1 |
| 20110016356 | Artzi et al. | Jan 2011 | A1 |
| 20110016456 | Artzi et al. | Jan 2011 | A1 |
| 20110016457 | Artzi et al. | Jan 2011 | A1 |
| 20110083123 | Lou et al. | Apr 2011 | A1 |
| Entry |
|---|
| Abreu, Rui, Peter Zoeteweij, and Arjan JC Van Gemund. “On the accuracy of spectrum-based fault localization.” Testing: Academic and Industrial Conference Practice and Research Techniques-Mutation, 2007. TAICPART-Mutation 2007. IEEE, 2007. |
| Agarwal, Manoj K., et al. “Problem determination using dependency graphs and run-time behavior models.” Utility Computing. Springer Berlin Heidelberg, 2004. 171-182. |
| Appleby, Karen, G. Goldszmidt, and Malgorzata Steinder. “Yemanja—A layered fault localization system for multi-domain computing utilities.” Journal of Network and Systems Management 10.2 (2002): 171-194. |
| Chen, Mike Y., et al. “Pinpoint: Problem determination in large, dynamic internet services.” Dependable Systems and Networks, 2002. DSN 2002. Proceedings. International Conference on. IEEE, 2002. |
| Steinder, Maigorzata, and Adarshpal S. Sethi. “The present and future of event correlation: A need for end-to-end service fault localization.” Proc. IIIS SCI: World Multi-Conf. Systemics Cybernetics Informatics. 2001. |
| Abreu, Rui, et al., “An Evaluation of Similarity Coefficients for Software Fault Localization,” Software Technology Department, Faculty of Electrical Engineering, Mathematics, and Computer Science, Delft University of Technology, The Netherlands. In PRDC 2006, pp. 39-46, 2006. |
| Agrawal, Hiralal, et al., “Fault Localization Using Execution Slices and Dataflow Tests,” Proceedings of the Sixth International Symposium on Software Reliability Engineering (ISSRE95), pp. 143-151, Toulouse, France, Oct. 1995. |
| Artzi, Shay, et al., “Finding Bugs in Dynamic Web Applications,” ISSTA '08, Jul. 20-24, 2008, Seattle, Washington, Copyright 2008 ACM 978-1-60558-050-0/08/07. |
| Cadar, Cristian, et al., “EXE: Automatically Generatng Inputs of Death,” CCS'06, Oct. 30-Nov. 3, 2006, Alexandria, Virginia, Copyright 2006, ACM 1-59593-518-5/06/0010. |
| Cleve, Holger, et al., “Locating Causes of Program Failures,” ICSE '05, May 15-21, 2005, St. Louis, Missouri, Copyright 2005 ACM 1-58113-963-2/05/0002. |
| Dallmeier, Valentin, et al., “Lightweight Defect Localization for Java,” Saar;amd moversotu. Saarbrucken, Germany, ECOOP, pp. 528-550, 2005. |
| Godefroid, Patrice, et al., “DART: Directed Automated Random Testing,” PLDI'05, Jun. 12-15, 2005, Chicago, Illinois, Copyright 2005, ACM 1-59503-056-6/05/0006. |
| Godefroid, Patrice, et al., “Automated Whitebox Fuzz Testing,” NDSS, 2008. |
| Horwitz, Susan, “Interprocedural Slicing Using Dependence Graphs,” SCM/SIGPLAN Conference on Programing Language Design and Implementation (1979-199): A Selection, 2003. Copyright 2003 ACM 1-58113-623-4. |
| Hutchins, M., et al., “Experiments on the Effectivenes of Dataflow—and Controlflow-Based Test Adequacy Critera,” 0270-5257/94, Copyright 1994 IEEE. |
| Jones, James A., et al., “Epirical Evaluation of the Tarantula Automatic Fault-Localization Technique,” ASE'05, Nov. 7-11, 2005, Long Beach, California, copyright 2005, ACM 1-58113-993-4/05/0011. |
| Jones, James A., et al., “Vizualization of Test Information to Assist Fault Localization,” Proceedings of the International Conference on Software Engineering, Orlando, Florida, May 2002, ACM Copyright. |
| Libit, Ben, et al, “Bug Isolation via Remote Program Sampling,” PLDI'03, Jun. 9-11, 2003, San Diego, California, copyright 2003 ACM 1-58113-662-5/03/006. |
| Liblit, Ben, et al., “Scalable Statistical Bug Isolation,” PLDI'05, Jun. 12-15, 2005, Chicago, Illinois, copyright 2005, ACM 1-59593-080-0/05/0006. |
| Liu, Chao, et al., “SOBER: Statistical Model-Based Bug Localization,” ESEC-FSE'05, Septeber 5-9, 2005, Lisbon, Portugal, copyright 2005 AC 1-59593,014,0/05/0009. |
| Pan, Hsin, et al., “Heiristics for Automatic Localization of Software Faults,” Technical Report SERC-TR-116-P, Purdue University, Jul. 1992. |
| Renieris, Manos, et al., “Fault Localization with Nearest Neighbor Queries,” In ASE, pp. 30-39, 2003. |
| Sen, Koushik, et al., “CUTE: A Concolic Unit Testing Engine for C,” ESEC-FSE'05, Sep. 5-9, 2005, Lisbon, Portugal, copyright 2005 ACM 1-59593-014/0/05/0009. |
| Stoerzer, Maximilian, et al., “Finding Failure-Inducing Changes in Java Programs Using Change Classification,” SIGSOFT'06/FSE-14, Nov. 5-11, 2006, Portland Oregon, copyright 2006, ACM 1-59593-468-5/06/0011. |
| Tip, Frank, “A Survey of Program Slicing Techniques,” Journal of Programming Languages, 3(3):121-189, 1995. |
| Wassermann, Gary, et al., “Dynamic Test Input Generation for Web Applications,” ISSTA'08, Jul. 20-24, 2008, Seattle, Washington, copyright 2008, ACM 978-1-59593-904-3/08/07. |
| Yu, Yanbing, et al., “An Empirical Study of the Effects of Test-Suite Reduction on Fault Localization,” ISCSE'08, May 10-18, 2008, Leipzig, Germany, copyright 2008, ACM 978-1-60558-079-1/08/05. |
| Zeller, Andreas, “Yesterday, my program worked. Today, it does not. Why?”, FSE< vol. 1687 of Lecture Notes in Computer Science, pp. 253-267, Springer, Sep. 1999. |
| Zeller, Andreas,“ Isolating Cause-Effect Chains from Computer Programs,” SIGSOFT 2002/FSE-10, Nov. 18-22, 2002, Charleston, SC, copyright 2002 ACM 1-58113-514-9/02/0011. |
| Anand, Saswat, et al., “Demand-Driven Compositional Symbolic Execution,” in TACAS, 2008. |
| Brumley, David, et al., “Towards Automatic Discovery of Deviations in Binary Implementations with Applications to Error Detection and Fingerprint Generation,” Proceedings of 16th USENIX Security Symposium on USENIX Security Syposium, 2007. |
| Cadar, Cristian, et al, “KLEE: Unassisted and Automatic Generation of High-Coverage Tests for Complex Systems Programs,” in OSDI, 2008. |
| Cadar, Cristian, et al., “Execution Generated Test Cases: How to Make Systems Code Crash Itself,” in SPIN 2005. |
| Wasserman, Gary, et al., “Dynamic Test Input Generation for Web Applications,” ISSTA'08, Jul. 20-24, 2008, Seattle, Washington, copyright 2008 AC 978-1-59593-904-3/08/07. |
| Godefroid, Patrice, “Compositional Dynamic Test Generation,” POPL'07, Jan. 17-19, 2007, Nice, France, copyright 2007, ACM 1-59593-575-4/07/0001. |
| Godefroid, Patrice, et al., “Grammar-based Whitebox Fuzzing,” PLDI'08, Jun. 7-13, 2008, Tucson, Arizona, copyright 2008 ACM 978-1-59593-860-2/08/06. |
| Inkumsah, Kobi, et al., “Evacon: A Framework for Integrating Evolutionary and Concolic Testing for Object-Oriented Programs,” in ASE 2007. |
| Majumdar, Rupak, et al., “Hybrid Concolic Testing,” 29th International Conference on Software Engineering (ICSE'07), 0-7695-2828-7/07, copyright 2007 IEEE. |
| Majumdar, Rupak, et al., “Directed Test Generation Using Symbolic Grammars,” ASE'07, Nov. 4-9, 2007, Atlanta, Georgia, copyright 2007 ACM 978-1-59593-882-4/07/0011. |
| Elbaum, Sebastian, et al., “Web Application Characterization Through Directed Requests,” WODA'06, May 23, 2006, Shanghai, China, copyright 2006 ACM 1-59593-085-X/06/0005. |
| Elbaum, Sebastian, et al., “Leveraging User-Session Data to Support Web Application Testing,” IEEE Trans. Softw. Eng., 31(3), 0098-5589/05, 2005. |
| Fisher, Marc, et al., “Dynamic Characterization of Web Application Interfaces,” FASE 2007, LNCS 4422, pp. 260-275, 2007, copyright Sprger-Verlag Berlin Heidelberg 2007/. |
| Halfond, William, G.J., et al., “Improving Test Case Generation for Web Applications Using Automated Interface Discovery,” ESEC/FES'07, Sep. 3-7, 2007, copyright 2007 ACM 978-1-59593-811-4/07/009. |
| Huang, Yao-Wen, et al., “Verifyiing Web Applications Using Bounded Model Checking,”Proceedings of International Conference on Dependable Systems and Networks, 2004. |
| Johns, Martin, et al., “SMask: Preventing Injection Attacks in Web Applications by Approximating Automatic Data/Code Separation,” SAC'07 Mar. 11-15, 2007, Seoul, Korea, copyright 2007, ACM 1-59593-480-4/07/0003. |
| Jovanovic, Nenad, et al., “Pixy: A Static Analysis Tool for Detecting Web Application Vulnerabilities,” in Security and Privacy, 2006. |
| Kiezun, Adam, et al., “Automatic Creation of SQL Injection and Cross-Site Scripting Attacks,” in Proceedings of International Conference of Software Engineering, ICSE, 2009. |
| McAllister, Sean, et al., “Leveragng User Interations for In-Depth Testing of Web Applications,” in Raid '08: Proceedings of the 11th International Symposium on Recent Advances in Intrusion Detection, pp. 191-210, Berlin, Heidelberg, 2008, Springer-Verlag. |
| Minamide, Yasuhiko, “Static Approximation of Dynamically Generated Web Pages,” WWW 2005, May 10-14, 2005, Chiba, Japan, ACM 1-59593-046-9/05/005. |
| Pietraszek, Tadeusz, et al., “Defending Against Injection Attacks through Context-Sensitive String Evaluation,” in RAID 2005. |
| Caprile, Bruno, et al., “Analysis and Testing of Web Applications,” in ICSE 2001. |
| Sprenkle, Sara, et al., “Automated Replay and Failure Detection for Web Applications,” ASE'05, Nov. 7-11, 2005, Long Beach, California, copyright 2005 ACM 1-58113-993-4/05/0011. |
| Su, Zhengong, et al., “The Essence of Command Injection Attacks in Web Applications,”POPL'06, Jan. 11-13, 2006, Charleston, South Carolina, copyright 2006 ACM 1-59593-02702/06/0001. |
| Wasserman, Gary, et al, “Sound and Precise Analysis of Web Applications for Injection Vulnerabilities,” PLDI'07, Jun. 11-13, 2007, San Diego, California, copyright 2007, ACM 978-1-59593-633-2/07/0006. |
| Xie, Yichen, et al., “Static Detection of Security Vulnerabilities in Scripting Languages,” in USENIX-SS, 2006. |
| Freire, Juliana, et al., “VeriWeb: Automatically Testing Dynamic Web Sites,” in WWW 2002. |
| Number | Date | Country | |
|---|---|---|---|
| 20110030061 A1 | Feb 2011 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 12502519 | Jul 2009 | US |
| Child | 12902423 | US |
