Vulnerability detection and repair is a demanding and expensive part of the software development process. As such, there has been an effort to develop new and better ways to automatically detect and repair vulnerabilities. DifFuzz is a state-of-the-art tool for automatic detection of timing side-channel vulnerabilities, a type of vulnerability that is particularly difficult to detect and correct. Despite recent progress made with tools such as DifFuzz, work on tools capable of automatically repairing timing side-channel vulnerabilities is scarce. In this paper, we propose DifFuzzAR, a tool for automatic repair of timing side-channel vulnerabilities in Java code. The tool works in conjunction with DifFuzz and it is able to repair 56% of the vulnerabilities identified in DifFuzz’s dataset. The results show that the tool can automatically correct timing side-channel vulnerabilities, being more effective with those that are control-flow based. In addition, the results of a user study show that users generally trust the refactorings produced by DifFuzzAR and that they see value in such a tool, in particular for more critical code.
Hinweise
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
1 Introduction
Software vulnerabilities are a serious threat to the security of software systems and can have disastrous consequences. For that reason, the detection of software vulnerabilities is an important problem that has received a lot of attention from the software security community. However, the detection of vulnerabilities can be difficult since a vulnerable application can pass all tests or even fulfil its correctness specification. Different types of vulnerabilities have different difficulty levels of detection. Arguably, one of the hardest types of vulnerabilities to be detected are side-channel vulnerabilities.
A side-channel is any observable side effect of a computation which can manifest in several different ways: for example, in the difference in computation time, in power consumption, sound production, or electromagnetic radiation emitted (Zhou and Feng 2005; Koeune and Standaert 2005). Most of the side-effects require that the attackers have physical access to the system they are trying to attack, since they need to gather information directly from the system (e.g. measuring the power consumption). On the other hand, side-effects such as the difference in computation time or in response size do not require the attackers to be in direct contact with the systems under attack. This enables the possibility of remote attacks, thus exposing systems to a larger number of attackers.
Anzeige
Moreover, side-channel vulnerabilities based on measuring differences in computation time, also known as timing side-channel vulnerabilities, can occur at multiple program points: for example, they can occur on a simple method to compare strings, or on a large and complex parallel computation. There are multiple real-world applications that were found to be vulnerable to timing side-channel attacks. For instance, Nate Lawson et al. discovered a timing side-channel vulnerability in Google’s Keyczar Library (Lawson 2020); another example is the timing side-channel vulnerability discovered in Xbox 360 (IVC Wiki 2020).
As timing side-channel vulnerabilities are difficult to detect, there has been a substantial effort to develop tools capable of automatically detecting these vulnerabilities (Antonopoulos et al. 2017; Chen et al. 2017; Nilizadeh et al. 2019). Despite this, once vulnerabilities are found, developers must correct them manually, which in some cases can be difficult, time-consuming and prone to errors. As such, we propose to facilitate the correction of vulnerabilities by developing a tool capable of automatically repairing timing side-channel vulnerabilities. Even though the ideas presented in this paper are general and can be applied to different programming languages, we focus on Java, since according to GitHub (2019), Java is the second language with more contributors in public and private repositories and is still the most used language for enterprise applications (Cloud Foundry 2020; IBM 2020). The tool developed, called DifFuzzAR, is designed to work in conjunction with the state-of-the-art detection tool DifFuzz Nilizadeh et al. (2019). We evaluated DifFuzzAR using the same dataset that was used to evaluate DifFuzz: although DifFuzzAR has some limitations, it repaired 56% of the vulnerabilities identified in DifFuzz’s dataset. This shows that DifFuzzAR has the potential to simplify substantially the debugging process. We also performed a quantitative and subjective user study to gather insights on whether potential users would like to use DifFuzzAR and if they trust its results, and to collect suggestions for improvement. The results are positive and show that DifFuzzAR has the potential to be used and trusted by its target users and that these do see value in the tool, in particular for more critical code.
This paper is an extension of the work originally presented in Lima et al. (2021). In particular, we improve the presentation, we clarify descriptions, we extend parts of the related work, and we add new examples of refactorings that aid understanding. We also substantially extend the evaluation section by describing a new user study. In particular, we address the three following research questions:
RQ1.
Do users trust DifFuzzAR’s refactorings?
RQ2.
What would users like to see in a tool such as DifFuzzAR?
RQ3.
Do users value the use of DifFuzzAR differently in critical and non-critical applications?
1.1 Structure of the paper
We present background and related work in Sect. 2. After presenting an overview of the system in Sect. 3, we describe the main components of DifFuzzAR. In Sect. 4 we present how vulnerable methods are identified and in Sect. 5 we describe how vulnerabilities are fixed. The evaluation of the tool is presented in Sect. 6. We discuss threats to validity in Sect. 7 and we conclude the paper in Sect. 8, where we also present current limitations and discuss future work.
Anzeige
2 Background and related work
This section presents background and related work on timing side-channel vulnerabilities and automated repair methods.
2.1 Timing side-channel vulnerabilities
A timing side-channel vulnerability happens when a secret1 can be learned based on the time a computation takes to complete. To put it differently, an application is vulnerable to timing side-channel attacks when the time it takes to complete a computation depends on a given secret, e.g., a password. These attacks are practical and apply to general software systems. For example, in their influential work, Brumley and Boneh were able to extract private keys from an OpenSSL-based web server running on a machine in the local network (Brumley and Boneh 2005). As mentioned in the previous section, timing side-channel vulnerabilities were found in Google’s Keyczar Library (Lawson 2020) and Xbox 360 (IVC Wiki 2020). A timing side-channel vulnerability can appear in multiple ways, as detailed below.
2.1.1 Early-exit vulnerabilities
An early-exit timing side-channel vulnerability happens when the method contains exit points that are dependent on the value of a secret. An example is when checking if an array has a certain length and exiting immediately once it is established that it does not. If that array is a secret, then the execution time of the method is dependent on the value of the secret, in this case, on the size of the array.
Listing 1 shows a classic example of an early-exit vulnerability, consisting of code that checks if the given password is correct. In this code, an exit condition is dependent on the secret, the parameter sec. Resourceful attackers can find the password by trying different combinations of characters until they find the correct one, as they know they found a correct character simply by the time the function takes to respond. For example, if the password is “bcdef” and the attackers use the password “abcde”, the function returns false immediately. If they use the password “bbcde”, they will know that the first character of the password is correct simply because the function took extra clock cycles to return. With that, the attackers know that the first character of the password is “b”. They can now repeat the process for the other characters of the password.
×
2.1.2 Control-flow based vulnerabilities
A control-flow based timing side-channel vulnerability happens when there is a significantly slow operation that happens only when a certain condition is met. This means that an attacker can take notice of the time a method takes to return and learn that the slow operation is executed.
In the example shown in Listing 2, if the condition in line 7 is true, two slow methods will be executed, while if the condition is false, the execution time will be much faster. In this case, it is possible to learn information about the exponent through the timing behavior of the program, since the value of the exponent influences the execution time.
×
2.1.3 Mixed vulnerabilities
Some methods might suffer from both early-exit and control-flow based timing side-channel vulnerabilities. When this happens, we say that the method has a mixed timing side-channel vulnerability. To fix it, it is necessary to correct the early-exit and the control-flow parts of the vulnerability. This can be done in different ways. First, they can be corrected simultaneously. Another option is to first correct one of the types of vulnerability and then correct the other type on the corrected version of the first type. As an example, Listing 3 shows a method with a mixed timing side-channel vulnerability, where the parameter a is a secret.
×
2.2 Detection of timing side-channel vulnerabilities
Automated detection of timing side-channel vulnerabilities has received substantial attention in recent years. Antonopoulos et al. (2017) developed a new way to prove the absence of timing side-channels, by using decomposition instead of self-composition. Their approach divides the program’s execution traces into smaller and less complex partitions. Then each partition has their resilience to timing side-channels attacks checked through a time complexity analysis. The authors’ idea is that the resilience of each component proves the resilience of the whole program. To ensure that any pair of program traces with the same public input has a component containing both traces, the construction of the partition is done by splitting the program traces at secret-independent branches. The authors’ approach follows the demand-driven partitioning strategy that uses a regex-like notion that they call trails, which identifies sets of execution traces, particularly those influenced by tainted (or secret) data. The authors prove a non-relational property about a trace, instead of proving a relational property about every pair of execution traces. Their method is implemented in a tool called Blazer.
Chen et al. (2017) presented the notion of \(\upepsilon \)-bounded non-interference, a variation of Goguen and Meseguer’s non-interference principle (Goguen and Meseguer 1982). The execution time of an application can be affected by sources external to the application. As such, a minimum difference in execution time should be expected and must be accepted. This minimum change is what the authors denote as \(\upepsilon \). To simplify, \(\upepsilon \)-bounded non-interference means that regardless of the secret, the execution time of an application will not vary by more than \(\upepsilon \). To verify the \(\upepsilon \)-bounded non-interference property, the authors present a new program logic called Quantitative Cartesian Hoare Logic (QCHL), which is at the core of their technique. With QCHL the authors can “[...] prove triples of the form\(\langle \phi \rangle \)S \(\langle \psi \rangle \),where Sis a program fragment and \(\phi \), \(\psi \)are first-order formulas that relate the program’s resource usage (e.g., execution time) between an arbitrary pair of program runs”. The authors implemented their technique in a tool called Themis and showed that their tool can find previously unknown vulnerabilities in widely used Java programs.
These two works improve the field of detection of side-channel vulnerabilities. However, both use static analysis.
Nilizadeh et al. (2019) present a new approach based on dynamic analysis and introduce a new tool called DifFuzz that uses differential fuzzing.2DifFuzz instruments a program to record its coverage and resource consumption along the paths that are executed. As such, the inputs must maximize the code coverage. For that, they use the fuzz testing tool American Fuzzy Lop (AFL) (Zalewski 2017), which uses genetic algorithms and mutates the inputs using byte-level coverage. Given that AFL only supports programs written in C, C++, or Objective C, and DifFuzz is written in Java, the authors used Kelinci (Kersten et al. 2017) to connect the two tools, since Kelinci provides AFL-style instrumentation for Java programs. To use DifFuzz, one has to create a Fuzzing Driver (or Driver File), that parses the input provided by AFL and executes two copies of the code, measuring the cost difference between the two. That cost difference will be used to guide the AFL in the generation of more input values so that the difference can be increased. This process is repeated for a predetermined time or until the user cancels the execution of the tool. DifFuzz was evaluated with a dataset of widely-used Java applications and it found previously unknown vulnerabilities (later confirmed by the developers). It was also applied to complex examples from the DARPA STAC (2020) program. DifFuzz was able to find the same vulnerabilities as other tools and also found vulnerabilities on corrected versions of the benchmarks of Themis and Blazer.
The authors of DifFuzz proposed improvements such as adding statistical guarantees to the tool and adding automated repair methods to eliminate the vulnerabilities discovered by DifFuzz. Our work contributes to the latter.
2.3 Automated repair tools
Automated program repair consists in fixing software bugs with little to no human intervention. Usually, workflows for automated program repair start with fault localization, which aims at locating faults in the code; patch generation, which aims to generate valid patches; and patch validation, which checks the validity of the generated patches in the previous step. Validity of patches is often determined by checking whether the patched program passes a given test suite. As described in Goues et al. (2019), automated program repair techniques can be classified as heuristic-based (Forrest et al. 2009; Kim et al. 2013; Liu et al. 2019; Cornu et al. 2015), constraint-based (Nguyen et al. 2013; Xuan et al. 2016; Mechtaev and Roychoudhury 2016), and learning-based (Gupta et al. 2017; Chen et al. 2019; Lutellier et al. 2020; Ye et al. 2021; Chen et al. 2021; Yasunaga and Liang 2021; Allamanis et al. 2021; Yasunaga and Liang 2020).
Two influential works in the area are GenProg (Le Goues et al. 2011) and Nopol (Xuan et al. 2016). GenProg receives as input the faulty source code and a set of test cases. The set of test cases must contain a set of passing positive test cases and at least one failing negative test case. The negative test case encodes the fault to be repaired and the set of positive test cases encodes the functionalities that can not be lost while repairing the bug. GenProg uses genetic programming to search for a variant of the program that retains all required functionality but does not have the fault in question.
Xuan et al. (2016) presented Nopol, an approach to automatically repair buggy conditional statements. This approach takes as input a program and a set of test cases and outputs a patch for the input program with a conditional expression. The set of test cases passed as input is similar to the one expected by GenProg. However, unlike GenProg, which follows a generic approach for automatic software repair, Nopol was built to focus on buggy if conditions and missing precondition bugs. Buggy if conditions occur when a bug is the condition of an ‘if’ statement. Missing precondition bugs happen when there should be a condition before a statement, such as detecting a null pointer or an invalid index to access an array. Nopol uses Ochiai, a spectrum-based ranking metric that is used to rank statements in a descending order based on their suspiciousness score. The suspiciousness score indicates the likelihood that a statement contains a fault.
For a comprehensive survey on automated program repair, we recommend Monperrus’s survey (Monperrus 2015). Moreover, since this is a fast-moving research topic with new approaches being regularly proposed, we also recommend Monperrus’s living review on automated program repair (Monperrus 2018).
2.4 Automated repair of timing side-channel vulnerabilities
Wu et al. (2018) proposed a method based on program analysis and transformation to eliminate timing side-channel vulnerabilities. Their solution produces a transformed program functionally equivalent to the original program but without instruction and cache timing side-channels. They ensure that the number of CPU cycles taken to execute any path is independent of the secret data, and the cache behaviour of memory accesses is independent of the secret data in terms of hits and misses. Their method is implemented in LLVM and uses static analysis to identify the set of variables whose values depend on the secret inputs. To decide if those variables lead to timing side-channel vulnerabilities, they check if the variables affect unbalanced conditional jumps, for instruction timing side-channel, or accesses to memory blocks across multiple cache lines, for cache-related timing side-channel vulnerabilities. After this analysis, to mitigate the leaks, code transformation is performed to equalize the execution time. The method is implemented in a tool called SC-Eliminator.
3 System overview
DifFuzzAR is designed to work in conjunction with DifFuzz. The tool needs to first identify the vulnerable method to be repaired. For this, the tool assumes the existence of a Driver file that can be used with DifFuzz. Once the vulnerable method is identified using the Driver, the tool will attempt to repair the method. In its current version, DifFuzzAR will attempt to repair Early-Exit Timing Side-Channel vulnerabilities, Control-Flow Based Timing Side-Channel vulnerabilities, and, by combining both, Mixed Timing Side-Channel vulnerabilities.
DifFuzzAR was designed to be as modular as possible. This way, if someone wants to add functionality to repair another type of vulnerability, they simply have to create a new independent module with all the code capable of repairing it. The one thing that is intrinsic to the tool is the analysis of the Driver to identify the vulnerable method and the class it belongs to. Once this identification is done, the tool searches for that method and sends it to the module responsible for correcting an early-exit timing side-channel vulnerability. That module then creates a repaired version of the method, which is then sent to the module responsible for correcting a control-flow based timing side-channel vulnerability. That module then creates another repaired version of the method and, given that it is the final module, the tool outputs a new class with the corrected method. An overview of the architecture of the tool is shown in Fig. 1.
Fig. 1
Overview of DifFuzzAR
×
4 Identification of vulnerable methods
The first task of DifFuzzAR is to uncover the vulnerable method that is to be repaired. As mentioned above, the driver used for DifFuzz is used to identify the method. This means that the driver must be properly created so that the correct method is retrieved. We assume that drivers are similar to the drivers provided by DifFuzz, where each driver calls the vulnerable method twice, each time immediately after a call to the method Mem.clear()3. However, there are three groups of variations that we consider:
Group 1.
The simplest variation occurs when an object is created between the invocation of Mem.clear() and the invocation of the vulnerable method. This normally happens when the vulnerable method is an instance method and the object needed to invoke it is created before the invocation.
Group 2.
A second variation is when after the instruction Mem.clear() a ‘try’ block appears. When this happens, the vulnerable method is considered to be the first instruction of the ‘try’ block.
Group 3.
The third variation occurs when after the invocation of the instruction Mem.clear() an ‘if’ statement appears. When this happens, the invocation of the vulnerable method is considered to be the first statement of either the ‘then’ or ‘else’ block. This normally happens when the driver used for the safe and unsafe versions of an example are similar and the difference is only in the value assigned to a boolean variable. That variable will then be used as a condition of an ‘if’ to decide which method to invoke (either the safe or unsafe version). To resolve this case, it is necessary to record the variable and its value. When the tool finds the ‘if’ statement where its condition is the variable found, the value of the variable is used to decide whether to look in the first instruction of the ‘then’ block or the ‘else’ block.
The search for the method after the Mem.clear() instruction is done twice since in the driver the vulnerable method will be invoked twice. Following DifFuzz’s convention, the parameter that changes in both invocations of the method is considered to be the secret. We thus assume that the driver uses the same arguments for the public parameters and different ones for the secret. For example, consider the method invocations vulnMethod(a, b, c) and vulnMethod(a, d, c); here, the second parameter is considered by the tool as the secret, since it is the only parameter that changes. In the identification of the vulnerable method, the tool also finds the path to the class file where the method definition is, even if that class is an inner class in some package. The tool also validates its findings of the vulnerable method by comparing the two instances and checking if name, class, return type and the number of parameters are the same, while at least one parameter is different. A basic overview of this process can be seen in Algorithm 1.
×
After implementing this strategy, the tool was tested with the 58 drivers of all the examples provided with the DifFuzz dataset (Nilizadeh et al. 2019). The tool was capable of finding the correct vulnerable method in all examples.
5 Correction of vulnerabilities
In the current version of DifFuzzAR, the correction of a vulnerability is done in two separate phases: the correction of an early-exit timing side-channel vulnerability followed by the correction of a control-flow based timing side-channel vulnerability. This way, there are two separate modules, each responsible for the correction of one type of vulnerability. As mentioned above, the addition of the correction of a new type of side-channel vulnerability is as simple as writing the code responsible for that correction and adding the module to the tool, as well as its invocation.
From the previous identification step, the tool knows which method was identified as vulnerable by DifFuzz. However, it does not know the specific instruction or set of instructions that cause the vulnerability. As such, the tool has to analyze the code and produce a correction that consists in a modification of the code to make its execution time as independent of the secret as possible. Algorithm 2 shows a basic overview of the correction process. If the vulnerable method has more than one return statement, then the tool considers it to potentially have an early-exit and so the tool starts by correcting that vulnerability. Afterwards, the tool executes the module responsible for the correction of control-flow based timing side-channel vulnerabilities.
The tool is implemented in Java and uses the open-source library Spoon (Pawlak et al. 2015) for the refactoring process. Examples of corrected vulnerabilities, which can be useful to understand in more detail the descriptions presented in the next subsections, are available in our GitHub repository.4
The correction of early-exit timing side-channel vulnerabilities consists in the elimination of all ‘return’ statements except the last one. However, the result of the execution of the method should be the same after the modification. For that reason, every ‘return’ statement of the method will be replaced with an assignment of the value being returned to a variable. That variable will either be the variable returned in the final return (if it returns a variable) or a new one created with the return type of the method.
Algorithm 3 shows an overview of the correction process for early-exit timing side-channel vulnerabilities. The tool starts by obtaining the element returned in the final return of the method. If this element is not a variable, the tool creates a new variable of the same type as the return type of the method, and initializes it with the element obtained, referred from now on as the return variable. Then, the tool analyses every instruction of the method in search for a return statement, which will be replaced by an assignment to the return variable with the value being returned. If that return statement happens after a ‘while’ block, then the instruction is added before the ‘while’ block. If it is the last return statement, then the value being returned is altered to be the return variable. If the return statement is inside an ‘if’ statement, then the condition of the ‘if’ statement is saved to be used to protect the variables used in the condition. If the instruction under analysis uses any variable saved to be protected, then that statement will be inside the ‘then’ block of a new ‘if’ statement, where the condition is the combination of the negation of every condition that variable was part of.
In the end, a new version of the class that contains the vulnerable method is created. This version is a copy of the original version, except that it contains an extra method called VulnerableMethodName$Modification. If the users want to use the corrected version, they must replace the original method with the corrected method.
×
This correction can create or exacerbate a control-flow based timing side-channel vulnerability. For instance, if an early-exit happens inside a loop, where the stopping condition depends on a secret, then the effect of the existing control-flow based timing side-channel vulnerability becomes more prominent, i.e., the difference in execution time depending on the secret is greater since it will have more instructions to execute.
Listing 4 shows another example of code which contains an early-exit timing side-channel vulnerability. The solution for early-exit timing side-channel vulnerabilities is to remove the early exit from the code. Listing 5 shows the corrected code (as proposed by DifFuzzAR). The solution in this case consists in removing the early-exit, which was in line 17 in the original code, and changing the logic of the program to have a single exit point.
The correction of control-flow based timing side-channel vulnerabilities involves (1) the modification of the stopping condition of loops that depend on a secret to depend on a public argument or (2) the replication of the block of instructions of the ‘then’ block to the ‘else’ block, and vice-versa, of an ‘if’ statement where the condition depends on the secret.
Algorithm 4 shows an overview of the correction process for this type of vulnerabilities. The pseudo-code is divided into two parts to improve presentation. In this process, the tool starts by creating a list of the secrets and a list of the public arguments. The list of public arguments is final, while the list of secrets is updated during the analysis of the method. Every time a variable is assigned with a value dependent on a secret, that variable is added to the list of secrets. The tool also creates a map to connect the newly created variables with the old variables being replaced. The tool then starts to analyse each instruction, taking actions according to the type of instruction and where the instruction happens.
×
×
×
If the instruction found is an assignment that needs to be modified, then a new variable is created and it is added to the map of replacements with the existing variable. The instruction is also changed so that the variable being assigned to is the newly created one. If the instruction found is a ‘for’ statement and the stopping condition uses a secret, the tool will attempt to change the condition to use a public argument instead of the secret. This public argument must be of the same type as the secret in the stopping condition. When the tool finds a ‘for’ statement it will retrieve the body of the ‘for’ and will analyse each instruction of that block. If the instruction found is an ‘if’ statement then the tool will retrieve the ‘then’ and ‘else’ blocks. If the condition uses a secret, then the tool will try to modify the instructions of the ‘then’ block and then of the ‘else’ block, producing two new blocks with the modified versions of the instructions. Then, the modified version of the ‘then’ block is added to the ‘else’ block and the modified version of the ‘else’ block is added to the ‘then’ block. Otherwise, the tool will analyse each instruction of both blocks without adding new instructions to either block. If the instruction is a method invocation, the tool will retrieve the target of that invocation. If the target is a secret, then the tool will create a new variable to replace the target. If the instruction is a local variable, the tool will retrieve the assigned value. If that value uses a secret, then the variable assigned to will be considered a secret. If the value being assigned does not use any variable that is used in the condition of the ‘if’ statement this instruction belongs to, then a new variable to replace the variable assigned to is created. If the instruction is a loop statement, then the tool will retrieve its body and will analyse each instruction of the body. If the instruction is an operator assignment, then the tool will create a new variable to replace the one being assigned to. If the instruction is a ‘try’ block, then the tool will retrieve its body and will analyse its instructions. If the instruction is a unary operator, the tool will retrieve the variable used. If that variable was already replaced, then the tool will obtain the variable created as a replacement and will replace the variable in the unary operator with the variable created for replacement. If the instruction is a ‘while’ statement, the tool will replace the variables used in the stopping condition, either by variables already created as replacements or with newly created variables. Then the tool retrieves the body of the loop and will analyse its instructions.
In the end, a new method is created with the control-flow based timing side-channel vulnerability corrected. An example is the code in Listing 6. Here, the method has a control-flow based timing side-channel vulnerability, since its execution time depends on the value of a secret, in this case the parameter taint. The ‘then’ and ‘else’ block of the ‘if’ statement has different instructions with different execution times. As such, to correct this vulnerability, the tool modifies the instructions of both blocks and adds the modified version of the ‘then’ block to the ‘else’ block and the modified version of the ‘else’ block to the ‘then’ block. The difference in the instructions is that the assignment is not made to the same variables, so as not to alter the value of the original variables. As a result, the tool produces the code in Listing 7.
Sometimes a method has both an early-exit and a control-flow based timing side-channel vulnerability. If the method has more than one return statement, the tool tries to repair an early-exit timing side-channel vulnerability producing a modified version of the method. Then, the tool tries to correct the control-flow based timing side-channel vulnerability in the modified version of the method, producing its final version. This means that each module responsible for correcting a type of timing side-channel vulnerability must return its modified version of the method. Since both repair processes create new variables in the method, and a method can not have two variables with the same name, the naming of a variable is global to the tool and it keeps a record of the names used.
6 Evaluation
In this section, we describe how the developed tool was evaluated. The evaluation is divided in two main parts:
testing the tool to ensure that the refactored code is semantically correct and to ensure that it has no timing side-channel vulnerabilities detected by DifFuzz, and
a user study to assess if potential users would like to use DifFuzzAR, if they trust the code refactorings produced by the tool, and to gather suggestions for improvement.
6.1 Testing the tool
The evaluation consists in ensuring that the refactored code is semantically correct and that it has no timing side-channel vulnerabilities detected by DifFuzz. This section presents both types of evaluation, explaining how they are done as well as why they are necessary.
This evaluation was performed in a remote server with a 32-processor Intel Xeon Silver 4110 at 2.10GHz with 64GB of RAM running Debian Linux 10. DifFuzz was configured to run for 2.5 h. The results of the evaluation can be seen in Table 1.
6.1.1 Dataset used
We started with the 32 examples distributed with DifFuzz. One of the examples suffers from a size side-channel. For other examples, the correction seems to involve some domain-specific changes, which are outside the scope of our proposed method. For instance, the example blazer modpow2, seems to contain a control-flow based timing side-channel vulnerability that is corrected by changing the multiplication operation, which is outside the scope of our work. Since we are interested in assessing the effectiveness of our tool within the scope in which it can operate, we decided to exclude these examples. Also, it was not possible to understand why some examples are vulnerable. As such, only 25 of those examples were used. Those examples were categorized according to the type of vulnerability.
One of the examples suffers from an early-exit timing side-channel vulnerability; eight of the examples contain a control-flow based timing side-channel vulnerability; the remaining 16 examples have a mixed timing side-channel vulnerability.
6.1.2 Semantics preservation
The modifications proposed can ‘break’ the code, in the sense that for the same inputs, the output can be different from that of the original version. As such, it is important that after any modifications to a method, the method is tested again to ensure that its functionality remains.
During the development of the tool, the application examples used by the authors of DifFuzz were used to ensure that the tool was capable of correcting a vulnerability. However, these examples do not include tests, so it was not possible to ensure that the correction kept the functionality of the method. Since creating manual tests is a time-consuming and error-prone activity, we decided to use (EvoSuite 2020) to generate tests automatically. The tests were created and first run on the original, vulnerable, code. We only retained tests that pass. Then, the vulnerable method was replaced with the method created by the tool and the tests were executed again. If all retained tests passed, then the solution created by the tool to correct the timing side-channel vulnerability was considered to be semantically correct.
Table 1 shows that 22 of the 25 attempted corrections (88%) are semantically correct. Regarding the 3 corrections that are not semantically correct, the first of them fails at compile time because the correction introduced a new variable that was used before being declared; the remaining two fail because, when removing a return to deal with the early-exit vulnerability, an exception ArrayIndexOutOfBoundsException is introduced. Regarding the missing variable declaration, one solution would be to add the creation of a new variable as the first instruction after the declaration of the variable that was replaced. However, this would add several extra instructions to the code and would create several instructions with the same name. All that would make it so that it would not be possible to minimize the execution time of the branches. It should be noted, however, that this can easily be fixed manually by the user.
6.1.3 Vulnerability correction
Once the tool repairs a vulnerable method and that repair is shown to be semantically correct, it is necessary to verify if the repair produced by the tool repairs the vulnerability. We use DifFuzz to determine if the repaired version contains any timing side-channel vulnerabilities.
Table 1 shows that out of 25 examples, the tool successfully corrected 14 of them, a success rate of 56%. Not all corrected versions produced by the tool are semantically correct, meaning that the code lost some of the functionality after the repair. When considering only semantically correct examples, the total of examples is 22, which makes a success rate of 63,6%.
Table 1
Results of the application of DifFuzzAR to the DifFuzz dataset
Example name
Has secure version?
Type
Correction attempted
Semantically correct
Vulnerability corrected
Apache FtpServer Clear
Yes
Mixed
Yes
No
–
Apache FtpServer Md5
Yes
Early-Exit (If dependent)
Yes
No
–
Apache FtpServer Salted
Yes
Mixed
Yes
No
–
Apache FtpServer StringUtils
Yes
Mixed
Yes
Yes
Yes
Blazer Array
Yes
Control-Flow
Yes
Yes
Yes
Blazer Gpt14
Yes
Control-Flow
Yes
Yes
No
Blazer K96
Yes
Control-Flow
Yes
Yes
Yes
Blazer Modpow1
Yes
Control-Flow
Yes
Yes
Yes
Blazer PasswordEq
Yes
Mixed
Yes
Yes
Yes
Blazer Sanity
Yes
Mixed
Yes
Yes
Yes
Blazer StraightLine
Yes
Control-Flow
Yes
Yes
Yes
Blazer UnixLogin
Yes
Control-Flow
Yes
Yes
Yes
Example PWCheck
Yes
Mixed
Yes
Yes
Yes
GitHub AuthmReloaded
Yes
Mixed
Yes
Yes
Yes
STAC Ibasys
No
Control-Flow
Yes
Yes
No
Themis Boot-Stateless-Auth
Yes
Mixed
Yes
Yes
No
Themis Dynatable
No
Mixed
Yes
Yes
No
Themis Jdk
Yes
Mixed
Yes
Yes
Yes
Themis Jetty
Yes
Mixed
Yes
Yes
Yes
Themis OACC
No
Mixed
Yes
Yes
Yes
Themis OrientDb
Yes
Mixed
Yes
Yes
No
Themis Pac4j
Yes
Control-Flow
Yes
Yes
Yes
Themis PicketBox
Yes
Mixed
Yes
Yes
No
Themis Spring-Security
Yes
Mixed
Yes
Yes
No
Themis Tomcat
Yes
Mixed
Yes
Yes
No
Some of the attempted vulnerability corrections fail due to the unavailability of public arguments to fix control-flow vulnerabilities. For instance, consider the example Themis Spring-Security shown in Listing 8, which contains a mixed vulnerability: an early-exit vulnerability in line 8 and a control-flow vulnerability in line 12 (since the loop depends on the secret). For this example, our tool fixes the early-exit vulnerability, as shown in Listing 9. However, to repair the control-flow vulnerability, the tool attempts to replace the secret used in the stopping condition of the loop (expectedLength) with a public argument of the same type (as shown in Algorithm 4). Since there are no public arguments of the same type, the tool does not change the stopping condition. This problem also applies to the examples Themis Dynatable, Themis OrientDb, Themis PickedBox. Without additional information, these examples cannot be fully repaired automatically. A possible future improvement is to bring the developer into the loop to obtain the additional information needed to repair the vulnerability. The other examples fail due to issues that go from the difficulty in identifying secrets due to complex data dependencies (e.g., in Blazer Gpt14, the secret is a parameter in the constructor of the class where the vulnerable method is defined) to the short circuiting semantics of conditional Java statements (which cause time variations when evaluating conditions, as in Themis Boot-Stateless-Auth). We have no immediate suggestions on how to fix these. More information about these failed corrections can be found at Lima’s Master’s thesis (Lima 2021).
×
6.2 User study
In this section, we provide the motivation for the user study, its goals, and the methods used. We end the section with the results obtained.
6.2.1 Motivation
DifFuzzAR tries to repair timing side-channel vulnerabilities by automatically refactoring the Java code. To understand the impact that changes proposed by DifFuzzAR may have on users, we reviewed previous work on tools that automatically change code (i.e. refactoring tools).
Murphy-Hill et al. (2011) studied refactoring usage from Eclipse’s user data. Their findings seem to suggest that users refrain from using refactoring tools because of three main factors:
1.
lack of awareness of their existence;
2.
lack of opportunity to use refactoring;
3.
lack of trust in refactoring.
Users refrain from using refactoring tools in part because of a lack of trust about the full impact of the tools in the code. Murphy-Hill et al. (2011) stated that several developers mentioned they would avoid using a refactoring tool because of worries about introducing errors or unintended side-effects.
Another study by Eilertsen (2012) investigated the usability of refactoring tools. They interviewed and performed a usability study with 17 developers. They concluded that users of refactoring tools often complain about a lack of control and usability (Eilertsen 2012). Refactoring tools may change a program in unpredictable ways and users often like to review the code.
6.2.2 Goals
Given the human aspects affecting refactoring tools, we decided to perform a quantitative and subjective user study about DifFuzzAR. The goal is to understand if potential users would use our tool, if they trust its results, and what improvements they suggest. In particular, we are interested in answering the following research questions (RQs):
RQ1.
Do users trust DifFuzzAR’s refactorings?
RQ2.
What would users like to see in a tool such as DifFuzzAR?
RQ3.
Do users value the use of DifFuzzAR differently in critical and non-critical applications?
6.2.3 Design and methods used
In order to answer our RQs we designed a survey study. We followed best practices from Redmiles et al. (2017) and followed methods similar to Eilertsen (2012). We recruited participants through our network (e.g., past students and colleagues). The participants did not receive payment upon survey completion. Participants were shown a consent form before filling in the survey. They could remove their consent at any point without giving justification. We did not collect any personal data.
As we wanted to study potential users of our tool, we recruited participants with Java programming experience. We pre-screened participants and only accepted those that have been working with Java for at least two years in the previous 10 years. To characterize our sample, we also asked participants to rate their expertise in Java and if they knew what timing side-channel vulnerabilities were before our study.
We then asked all participants to go through a brief explanation of timing side-channel vulnerabilities with examples, and another explanation about DifFuzzAR and about what the tool does.
To understand if users would trust the results of our tool we provided four vignette scenarios. A vignette, as described in Lavrakas encyclopedia (Lavrakas 2008), describes a protagonist (or group of protagonists) faced with a realistic situation pertaining to the construct under consideration. The respondent is asked to make a judgment about the protagonist, the situation, or the correct course of action, using some form of closed-ended response. In our vignette scenarios, we presented the following scenario to participants:
Imagine you apply DifFuzzAR to automatically repair timing side-channel vulnerabilities in Java code.
We then provided an example of code before and after a vulnerability is fixed by the tool. We used one scenario for control-based vulnerabilities, one for early-exit, and two for mixed vulnerabilities (one simple and one more complex). For each scenario, we asked participants to indicate, using a 5-point Likert agreement scale, if they would trust that refactoring. We also asked them for feedback. The code snippets used in these scenarios are taken from the examples used in the first part of the evaluation (Sect. 6.1) and can be seen in Appendix A. It is important to mention that in the early-exit example we modified the “before” code to remove the control-flow vulnerability. We did this to provide an example with only an early-exit vulnerability. While the “before” was adapted, the correction of the early-exit vulnerability showed to participants is a direct output of DifFuzzAR.
DifFuzzAR differs from usual refactoring tools as it aims at producing code that is more secure (instead of “cleaner” code). So, we also wanted to gauge if different situations impact users’ willingness to use a tool like DifFuzzAR. With this goal in mind, we provided users with two scenarios of DifFuzzAR’s usage. One scenario described a programmer that is coding a sensitive part of a program (e.g., authentication) and the other describes a less critical situation (e.g., programming the GUI of an application). We call them critical and non-critical scenarios, respectively, and they can be reviewed in more detail in Appendix B. For each, we asked users to indicate their willingness to use DifFuzzAR using a 5-point Likert agreement scale. We also asked them for feedback. It should be noted that the non-critical scenario (focused on the dark mode change) can also leak information through observing timing differences, since the processing time for switching to dark mode might be slightly longer or shorter than switching to light mode. Despite the absence of secrets, in such a scenario, DifFuzzAR could potentially suggest a refactoring that ensures a constant-time execution. Our goal in including this scenario is to understand whether developers see value in using DifFuzzAR in situations where no potentially sensitive information nor secrets can be retrieved. This can provide insights on how to best integrate tools such as DifFuzzAR into the developer’s workflow.
We finished our survey by asking participants for suggestions to improve DifFuzzAR and by asking demographic questions.
User studies can suffer from bias due to the sample, how the survey questions are made, and even from the response options (Redmiles et al. 2017). To mitigate this problem, we did two cognitive interviews,5 followed best practices by offering “don’t know” or “prefer not to answer” responses (Redmiles et al. 2017), and used methods previously used in studies on related subjects (Eilertsen 2012).
6.2.4 Results
Our survey was answered by 20 users but only 11 meet our requirements and passed the pre-screening. To pass this pre-screening they needed to answer the question “How many of the last ten years (2012-2022) have you spent developing or maintaining software in Java?” with at least two years. We chose to do this because we wanted participants that could understand Java code well. Our sample of 11 participants has, on average, four years of experience with Java.
We followed the same methods as Eilertsen (2012) and asked participants to report their self-described proficiency with changing Java code. The respondents rated their proficiency on a scale from 1 to 5 (where 1 indicates no proficiency and five indicates expert proficiency). Most respondents (10 participants) reported at least three (i.e., above average) proficiency in Java. Our sample of users also reported a good familiarity with timing side-channel vulnerabilities as almost half of the respondents (45.45%) self-reported they knew what they were. Moreover, 27.27% of participants were not familiar with the concept and 27.27% were unsure. Most of the participants had a Master’s degree (55%, i.e., 6 participants) with the remaining ones having a Bachelor’s degree (45%). All participants were male with the exception of one female participant. The most common age group was 18-24 (63.6%), followed by 25-34 (18.2%), 35-44 (9.1%), and 45-54 (9.1%).
DifFuzzAR Scenarios
Respondents went through four scenarios. For each scenario, they were presented with a concrete example of DifFuzzAR’s usage, where we provided Java code before and after applying the tool. After each scenario, we asked participants to use a 5-point Likert scale to indicate if they trust the refactorings produced by DifFuzzAR. When analyzing Likert scale data, it is recommended to use a non-parametric statistical test as the answers are not normally distributed (Lazar et al. 2017). Therefore, we analyzed the data using the Wilcoxon signed-rank non-parametric test with continuity correction. This test’s null hypothesis is that there is not a significant difference between the samples. To disprove this hypothesis, the resulting p-value should be less than \(0.05\) (\(p < 0.05\)). When this happens, we conclude that there is a significant difference between the samples (Lazar et al. 2017).
The first two scenarios (control-flow and early-exit) were found to be very trustworthy by participants, i.e., most respondents stated they “agreed” or “strongly agreed” that they trusted the refactorings. Only one participant did not trust the early-exit scenario and, from their feedback, this was due to the wrong understanding that the refactoring did not preserve the behavior of the original code. Moreover, the Wilcoxon signed-rank non-parametric test did not find a statistical difference between the first two scenarios (\(p > 0.05\)).
The remaining two scenarios (mixed vulnerabilities, one simple and one more complex) while generally trusted (55% “agreed” or “strongly agreed” that they trusted the refactorings) had a more mixed response from participants. Two (2) out of 11 participants stated they did not trust the refactoring done in scenarios 3 and 4. Their answers seem to indicate that this may be due to two reasons: (a) the participant was unable to understand the changes that had been made to the code (e.g., the participant did not understand the final version of the code); (b) the scenario was complex (e.g., the participant associated complexity with distrust). This lack of trust in more complex refactorings is something that previous literature on refactoring also found to be the case (Eilertsen 2012).
While these scenarios (3 and 4) were not found as trustworthy as the first ones (1 and 2) there was no statistical difference between any of them (\(p > 0.05\)).
Critical and non-critical scenarios
Now that we have established that most participants trust DifFuzzAR’s refactorings, we turn our attention to the critical and non-critical scenarios. The critical scenario describes programming authentication code and the non-critical describes programming an application’s GUI. All users found DifFuzzAR to be useful in the critical scenario (100% answered “agree” or “strongly agree”) but to be less useful in the non-critical scenario ( 27% “agree” or “strongly agree”). These results are statistical different with \(p = {0.00072}\).
After analyzing participants’ answers, our results suggest that users value more DifFuzzAR’s usage in critical use cases, and less in day-to-day coding procedures as one participant stated “In the second (scenario), there is no risk associated with the functionality, no secrets are involved.”. The results also seem to indicate that the prevention of timing side-channel vulnerabilities is more important when their exploitation can lead to loss of valuable secrets (like in the critical scenario).
Further improvements
After the previous questions, we asked participants to suggest improvements and future features that DifFuzzAR could implement. We coded their answers with an emergent coding scheme (Lazar et al. 2017) as we had no previous insights about what their answers could be. The frequency of their answers and codes can be seen in Table 2. It is important to note that respondents’ answers may be coded with more than one code and some users may have not answered at all, so, the total frequency of the codes does not necessarily match the number of participants in the study.
Table 2
Coded answers to “What functionalities/improvements would you like to see in this tool?” and frequency of answers
Code description
Frequency
Correct other common vulnerabilities
3
Provide more information about the changes
3
I don’t know
2
Add analysis of nested functions
1
Less complex changes
1
Apply the tool to interpreted languages (e.g., Python)
1
Improve variable names
1
Some participants stated that they would like to see additional features in DifFuzzAR such as the correction of other common vulnerabilities in other languages. One of the participants even goes as far as suggesting they would like “to do this (use DifFuzzAR ) in an interpreted language (e.g., Python) instead of a compiled language”. Another common theme in users’ suggestions for improvements was that DifFuzzAR could provide more information about the changes it makes (e.g., by adding comments explaining the changes, with better variable names or, as one user mentioned, with a “detailed profiling of the modified code, before and after it was modified (...)”.) Overall users’ comments were positive and their feedback suggested useful future features for DifFuzzAR.
6.2.5 Discussion
We were able to successfully answer the proposed research questions (see Sect. 6.2.2) and the insights gathered seem to confirm the usefulness of DifFuzzAR. We now address briefly each of our research questions.
RQ1. Do users trustDifFuzzAR’s refactoring?
Our results suggest that users trust DifFuzzAR’s code transformations. They also seem to indicate that the complexity of the transformations affect trust, as simpler code transformations were seen as more trustworthy by participants.
RQ2. What would users like to see in a tool such asDifFuzzAR? This study’s results seem to indicate that our participants value
a thorough tool that also corrects other common vulnerabilities in other programming languages. Users’ also seem to value a tool that is transparent about its functioning, this is, a tool that informs the user about the changes it does to the code.
RQ3. Do users value the use ofDifFuzzAR differently in critical and non-critical applications?
Our results seem to suggest that users value the use of DifFuzzAR differently in distinct situations. From the data gathered, participants appear to be more willing to use DifFuzzAR in more critical Java code. If this is the case, then the potential impact of DifFuzzAR is greater as this type of code is impactful to the security of a product.
While we gathered significant data in this user study, this data is subjective as our sample size is relatively small. However, this study has still provided valuable insights that can inform future large-scale user studies. It also seems to confirm that users generally trust refactorings produced by DifFuzzAR and that they see value in a tool like this, in particular in more critical Java code. It has also provided suggestions for improvement of DifFuzzAR that can also be useful for other similar tools.
7 Threats to validity
Regarding our tool, a threat to validity is that it can have bugs and be incapable of fixing vulnerable code not considered in this study. We mitigate this risk by explicitly stating the scope of our tool (i.e., the goal is to work in conjunction with DifFuzz) and by considering in our study all the examples available in DifFuzz’s public benchmark. Moreover, all our code and data are publicly available for other researchers and potential users to check the validity of the results.
Regarding the user study, a threat to validity is the fact that we recruited participants through our professional network (e.g., past students and colleagues), as this can potentially bias the results. However, the sampling procedure that we use is commonly employed in studies of this nature. For example, as one of the reviewers of this work brought to our attention, works that can be used as guidelines for empirical studies, such as Ciolkowski et al. (2003), use similar approaches as they use their own industrial contacts.
While the user study might have limitations and the results might not be completely generalizable, it nonetheless offers valuable insights to the community and has the potential to shape and guide future research endeavors. We do not claim that our study is perfect, but we argue that it provides valuable information. Moreover, we followed best practices from Redmiles et al. (2017) and followed methods similar to Eilertsen (2012). Similar methods are also used in other research papers. For example, we use vignette-based surveys, as such surveys have been found to well-approximate real-world behavior (Cummings et al. 2021; Hainmueller et al. 2015). It is worth noting that, as stated by Berry and Tichy (2003), a reference kindly suggested by one of the reviewers of this paper, “Perfection in experiments, especially in those involving human subjects, is unattainable”.
8 Conclusions
This paper presents a tool for automatic repair of timing side-channel vulnerabilities in Java code that works in conjunction with DifFuzz (Nilizadeh et al. 2019). Patterns that lead to timing side-channel vulnerabilities were identified and algorithms capable of correcting those potential vulnerabilities were proposed and implemented. The tool developed was evaluated using the same dataset that was used to evaluate DifFuzz (Nilizadeh et al. 2019), a dataset that contains examples of applications with timing side-channel vulnerabilities. The results obtained show that 88% of the attempted corrections are semantically correct (i.e. the original behavior is preserved) and 56% of the corrections eliminate the existing timing side-channel vulnerabilities. Moreover, the results obtained in a quantitative and subjective user study seem to confirm that users generally trust refactorings produced by DifFuzzAR and that they see value in a tool like this, in particular in more critical Java code. The fact that users do not seem to value the use of DifFuzzAR in non-critical scenarios suggests that one should be careful when integrating a tool of this nature into the developer’s workflow. It might not be advisable to allow such a tool to provide refactoring suggestions for every component that might leak information through observing timing differences, as fixing non-critical issues might frustrate developers and reduce the adoption of such a tool. A more judicious use, where only components that deal with secrets are considered, is recommended. This might be achieved by attempting to automatically infer the secrets or by allowing the developers to manually annotate the code to identify such secrets.
Even though there is space for improvement, we believe that DifFuzzAR can be used as a starting point for the development of new and improved tools capable of correcting timing side-channel vulnerabilities and other related vulnerabilities. The tool is open-source and is available at: https://github.com/sr-lab/DifFuzzAR.
8.1 System limitations
Although DifFuzzAR was built in an attempt to correct timing side-channel vulnerabilities regardless of how they present themselves, it is still possible that sometimes the repair created by the tool, not only does not repair the vulnerabilities, but also breaks some of the functionality of the method. As such, it is important to do a manual analysis of the repaired method after the execution of the tool, not only to check if no functionality is broken but also to beautify the changes (e.g. improve the names of the variables). Besides that, it is important that after the execution of the tool, the produced code is analysed again with DifFuzz to see if the tool eliminated the vulnerability.
The tool assumes that the method referenced in the Driver is vulnerable and corrects it. As such, if the Driver is not properly written or the method referenced is not the vulnerable one, but one that calls the truly vulnerable method, then the tool will not be able to repair it. DifFuzzAR can automatically repair the patterns identified and described in this paper. For vulnerabilities that follow other types of patterns, the tool needs to be extended. It is thus necessary to continuously improve the tool to be able to correct different code patterns that contain a vulnerability, or different instructions that cause the vulnerability. If the tool is executed on the correction of a control-flow based timing side-channel vulnerability, it will always try to repair the vulnerability again, which means it might break the original correction. In the results presented in this paper, the corrected versions of some examples are presented as having no timing side-channel vulnerability. However, there is always the possibility that they might have a vulnerability that remained unnoticed. Despite this, all the work developed is open to others on GitHub.
8.2 Future work
There is still plenty of work that can be done to improve DifFuzzAR. An important direction is to add the ability to repair more examples of timing side-channel vulnerabilities (including patterns not considered). DifFuzzAR is designed to be used in conjunction with DifFuzz. This means that the user must create a Driver following the rules described in Sect. 4. A future direction that would greatly simplify the use of the tool is to automatically generate a Driver file. Another future improvement for the tool is to transform it from a tool into a plugin to be used in the build process of the application. This would reduce the amount of manual intervention needed by the user. Another advantage of this is that being part of the build process can make it easier for other users to use the tool. Another direction would be to adapt DifFuzzAR so that it could easily be used and distributed as an IDE plugin. For this, it is likely that adjustments to the code transformations (e.g. better variable names) and usability studies should be performed, to ensure that the code transformations are accepted by the programmers. It would also be interesting to apply the tool to public projects and submit any corrections found as pull requests, thus improving existing software and, simultaneously, obtaining code reviews from developers (as done in related refactoring projects by the authors Ribeiro et al. 2021; Pereira et al. 2022).
Further user studies can also be useful for future improvements of DifFuzzAR. The user study presented here allowed us to gather significant data, but this data is subjective due to the size of the sample. As such, we suggest that future work should repeat our study with a larger sample of users. Future work on this topic should also include a usability analysis of DifFuzzAR’s usage with a larger number of users. We also believe that more work should be done to understand users’ motivations when using software to automatically repair vulnerabilities. While there have been previous works done about refactoring there is still a gap in the literature when it comes to tools os the same nature as DifFuzzAR.
Acknowledgements
We thank the anonymous reviewers of the previous version of this work (Lima et al. 2021) for their valuable and constructive comments. We also thank the anonymous reviewers of this extended version for their comments, which substantially improved the paper. This work was partially funded by the PassCert project, a CMU Portugal Exploratory Project funded by Fundação para a Ciência e Tecnologia (FCT), with reference CMU/TIC/0006/2019 and supported by national funds through FCT under project UIDB/50021/2020.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix 1: Java Code snippets used in user study scenarios
×
×
×
×
×
×
×
×
Appendix 2: Scenarios in user study
In the user study, participants were presented with two scenarios, one critical and one non-critical. These scenarios are presented below.
Critical Scenario
Imagine you are developing an application in Java. You write the code that is used for the login module of your project. Your code receives a user-provided string and compares it with a secret password. After writing the code, you think about whether you should use DifFuzzAR.
Non-critical Scenario
Imagine you are developing an application in Java. You write the code that is used for the GUI module of your project. Your code receives user-input when a user clicks on the “dark mode” button and changes the interface accordingly. After writing the code, you think about whether you should use DifFuzzAR.
DifFuzz instruments executions and keeps track of execution costs (so that it can detect discrepancies between executions). The method Mem.clear() resets the value of the current cost, which is necessary to measure the cost for each execution separately. For concrete examples, see https://github.com/sr-lab/DifFuzzAR/tree/master/src/test/resources.
Cognitive interviews involve asking respondents to think aloud as they complete a survey as well as asking them questions about each survey item (Redmiles et al. 2017).
