Can explainability and deep-learning be used for localizing vulnerabilities in source code?

Cyber-attacks are fairly mundane. The misconfigurations of the source code can result in security vulnerabilities that potentially encourage the attackers to exploit them and compromise the system. This paper aims to discover various mechanisms of automating the detection and correction of vulnerabilities in source code. Usage of static and dynamic analysis, various machine learning, deep learning, and neural network techniques will enhance the automation of detecting and correcting processes. This paper systematically presents the various methods and research efforts of detecting vulnerabilities in the source code, starting with what is a software vulnerability and what kind of exploitation, existing vulnerability detection methods, correction methods and efforts of best researches in the world relevant to the research area. A plugin will be developed which is capable of intelligently and efficiently detecting the vulnerable source code segment and correcting the source code accurately in the development stage.

Static analysis of source code security: Assessment of tools against SAMATE tests

To increase and maintain web application security, developers could use some different methods, one of them is static code analysis. This method could find security vulnerabilities inside a source code without the need of running the program. It could also be automated by using tools, which considered more efficient than manual reviews. One specific method which is commonly used in static code analysis is taint analysis. Taint analysis usually utilizes source code modeling to prepare the code for analysis process to detect any untrusted data flows into security sensitives computations. While this kind of analysis could be very helpful, static code analysis tool for Laravel-based web application is still quite rare, despite its popularity. Therefore, in this research, we want to know how static code (taint) analysis could be utilized to detect security vulnerabilities and how the projects (Laravel-based) should be modeled in order to facilitate this analysis. We then developed a static analysis tool, which models the application’s source code using AST and dictionary to be used as the base of the taint analysis. The tool first parsed the route file of Laravel project to get a list of controller files. Each file in that list would be parsed in order to build the source code representation, before actually being analyzed using taint analysis method. The experiments was done using this tool shows that the tools (with taint analysis) could detect 13 security vulnerabilities from 6 Laravel-based projects with one False Negative. An ineffective sanitizer was the suspected cause of this False Negative. This also shows that proposed modeling technique could be helpful in facilitating taint analysis in Laravel-based projects. For future development and studies, this tool should be tested with more Laravel and even other framework based web application with a wider range of security vulnerabilities.

Insecurity Refactoring is a change to the internal structure of software to inject a vulnerability without changing the observable behavior in a normal use case scenario. An implementation of Insecurity Refactoring is formally explained to inject vulnerabilities in source code projects by using static code analysis. It creates learning examples with source code patterns from known vulnerabilities.Insecurity Refactoring is achieved by creating an Adversary Controlled Input Dataflow tree based on a Code Property Graph. The tree is used to find possible injection paths. Transformation of the possible injection paths allows to inject vulnerabilities. Insertion of data flow patterns introduces different code patterns from related Common Vulnerabilities and Exposures (CVE) reports. The approach is evaluated on 307 open source projects. Additionally, insecurity-refactored projects are deployed in virtual machines to be used as learning examples. Different static code analysis tools, dynamic tools and manual inspections are used with modified projects to confirm the presence of vulnerabilities.The results show that in 8.1% of the open source projects it is possible to inject vulnerabilities. Different inspected code patterns from CVE reports can be inserted using corresponding data flow patterns. Furthermore the results reveal that the injected vulnerabilities are useful for a small sample size of attendees (n=16). Insecurity Refactoring is useful to automatically generate learning examples to improve software security training. It uses real projects as base whereas the injected vulnerabilities stem from real CVE reports. This makes the injected vulnerabilities unique and realistic.

The study explores the capabilities of these large language models in the domain of security analysis. Specifically, it investigates the models' ability to classify vulnerability types and descriptions from the CWE Top 25 category, which includes the most dangerous vulnerabilities for 2024 according to the CWE catalog ranking. To evaluate the models, several sets of specialized queries are constructed. These queries test the models' knowledge of vulnerability types as well as their ability to detect these types of vulnerabilities in source code. The experiments conducted with the models yielded a body of statistical data, allowing for an assessment of the models' effectiveness in the tasks of security analysis and vulnerability detection.

Static code analysis is widely used to support the development of high-quality software. It helps developers detect potential bugs and security vulnerabilities in a program's source code without executing it. While the potential benefits of static analysis tools are beyond question, their usability is often criticised and prevents software developers from using static analysis to its full potential. In the past decade, researchers have studied developer needs and contrasted them to available static analysis tool functionalities. In this paper, we summarize the main design challenges for building usable static analysis tools, and show that they revolve around the notion of explainability, which is a subarea of usability. We present existing analysis tools and current research in static analysis usability, and detail how they approach those challenges. This leads us to proposing potential lines of future work in explainability for static analysis, namely turning static analysis tools into assistants and teachers.

Code vulnerability detection is a crucial approach to ensuring software security, aiming to automatically identify potential security vulnerabilities in source code. However, existing static vulnerability detection methods often face challenges such as feature loss and an insufficient expressive power when extracting program features. To address these issues, in this paper, we propose a source code vulnerability detection method based on joint graph and multimodal feature fusion. Innovatively, we construct a joint graph that integrates multiple program dependencies and semantic edges to achieve more comprehensive feature extraction. Additionally, by combining Graph Attention Networks (GATs) with the Transformer architecture, both structural and sequential features of code snippets are captured, further enhancing the model’s expressive capabilities. Finally, we introduce pre-fusion and post-fusion strategies to fully integrate multimodal features, thereby improving the accuracy and performance of vulnerability detection. Experimental results on the SARD dataset demonstrate the method’s excellent performance in detecting various types of vulnerabilities, achieving an F1 score of 85.20% and an accuracy of 86.50%. On the Real-Vul real-world dataset, the method achieves an F1 score of 73.9% and an accuracy of 86.50%. The detection results exhibit remarkable stability, ensuring a reliable and consistent performance. Overall, the proposed method surpasses the performance of existing mainstream detection approaches.

Automated detection of vulnerabilities in source code is an essential cybersecurity challenge, underpinning trust in digital systems and services. Graph Neural Networks (GNNs) have emerged as a promising approach as they can learn the structural and logical code relationships in a data-driven manner. However, the performance of GNNs is severely limited by training data imbalances and label noise. GNNs can often learn “spurious” correlations due to superficial code similarities in the training data, leading to detectors that do not generalize well to unseen real-world data. In this work, we propose a new unified framework for robust and interpretable vulnerability detection—that we call VISION—to mitigate spurious correlations by systematically augmenting a counterfactual training dataset. Counterfactuals are samples with minimal semantic modifications that have opposite prediction labels. Our complete framework includes: (i) generating effective counterfactuals by prompting a Large Language Model (LLM); (ii) targeted GNN model training on synthetically paired code examples with opposite labels; and (iii) graph-based interpretability to identify the truly crucial code statements relevant for vulnerability predictions while ignoring the spurious ones. We find that our framework reduces spurious learning and enables more robust and generalizable vulnerability detection, as demonstrated by improvements in overall accuracy (from 51.8% to 97.8%), pairwise contrast accuracy (from 4.5% to 95.8%), and worst-group accuracy increasing (from 0.7% to 85.5%) on the widely popular Common Weakness Enumeration (CWE)-20 vulnerability. We also demonstrate improvements using our proposed metrics, namely, intra-class attribution variance, inter-class attribution distance, and node score dependency. We provide a new benchmark for vulnerability detection, CWE-20-CFA, comprising 27,556 samples from functions affected by the high-impact and frequently occurring CWE-20 vulnerability, including both real and counterfactual examples. Furthermore, our approach enhances societal objectives of transparent and trustworthy AI-based cybersecurity systems through interactive visualization for human-in-the-loop analysis.

The large number of industrial control systems based on PLCs (Programmable Logic Controllers) available at CERN implies a huge number of programs and lines of code. The software quality assurance becomes a key point to ensure the reliability of the control systems. Static code analysis is a relatively easy-to-use, simple way to find potential faults or error-prone parts in the source code. While static code analysis is widely used for general purpose programming languages (e.g. Java, C), this is not the case for PLC program languages. We have analyzed the possibilities and the gains to be expected from applying static analysis to the PLC code used at CERN, based on the UNICOS framework. This paper reports on our experience with the method and the available tools and sketches an outline for future work to make this analysis method practically applicable.

Because of the language’s simplicity, flexibility and massive libraries supporting artificial intelligence, web development and data science python has become one of the widely used programming language. But this speedy growth has led to software security and vulnerability concerns. This paper examines the security provisions available in python programming and their role in developing secure applications. Herein the analysis is done based on security mechanisms such as validation of inputs, authentication and authorization, libraries available for cryptography, management of dependency tools, static code analysis tools and secure web frameworks. The paper highlights the common security challenges faced in python programming and emphasis on the importance of secure coding practices and comparison of vulnerability detection tools implemented in the past research works. The finding proclaims that python provides several built-in tools and mechanisms which when properly implemented can notably improve reliability and security of software systems. Keywords: Python Security, Secure Coding Practices, Vulnerability Detection, Machine Learning Based Security Analysis, Static Code Analysis, Cryptography in Python, Software Security

Static code analysis is a very important stage in the development and implementation of software, and it needs to be used to obtain a better code. The most complicated in static analysis is the analysis of source code, and further analysis. Analysis of the source code can be done multiple times and expand the set of necessary grammar. The main goal was and remains to provide a solution to reduce the time needed for a global reassessment of static analysis at code level after a change.The main problem when using static code analysis is to build an abstract semantic graph, because each software solution is provided with a separate data warehouse. The proposed solution proposes to use as a database repository using graphs. Thus, the storage mechanisms of the created abstract semantic graph have been simplified, which in turn, in addition to increasing the clarity of the information we store, provides convenient ways to further work with the stored information.The developed solution can transform a rather large source repository into a graphic representation and maintain it later. It is found that the approach is suitable for carrying out code compliance checks and for performing static analysis tests on a graphic representation. This approach also uses advanced file-level detailing, accelerating static analysis. Based on my measurements, the frameworks are fast enough to help their users quickly change the repository of codes.This article confirms the thesis about the possibility of storing an abstract semantic graph in a graph database, and after refinement, if it contains sufficient transformations and requests for language processing, can become a complete transport for communication between various static analysis tools that usually perform one of two the functions of verification either for quality or for vulnerability, thereby making a unified creation of an abstract semantic graph.As an improvement it is necessary to consider the possibility of incremental analysis – the analysis of changes in the code, in order to minimize resource costs for a rather resource intensive operation of the structure of the abstract semantic graph.Manuscript received 09.06.2020

The prevalence and significance of web services in our daily lives make it imperative to ensure that they are – as much as possible – free from vulnerabilities. However, developing a complex piece of software free from any security vulnerabilities is hard, if not impossible. One way to progress towards achieving this holy grail is by using static code analysis tools to root out any common or known vulnerabilities that may accidentally be introduced during the development process. Static code analysis tools have significantly contributed to addressing the problem above, but are imperfect. It is conceivable that static code analysis can be improved by using AI-powered tools, which have recently increased in popularity. However, there is still very little work in analysing both types of tools’ effectiveness, and this is a research gap that our paper aims to fill. We carried out a study involving 11 static code analysers, and one AI-powered chatbot named ChatGPT, to assess their effectiveness in detecting 92 vulnerabilities representing the top 10 known vulnerability categories in web applications, as classified by OWASP. We particularly focused on PHP vulnerabilities since it is one of the most widely used languages in web applications. However, it has few security mechanisms to help its software developers. We found that the success rate of ChatGPT in terms of finding security vulnerabilities in PHP is around 62-68%. At the same time, the best traditional static code analyser tested has a success rate of 32%. Even combining several traditional static code analysers (with the best features on certain aspects of detection) would only achieve a rate of 53%, which is still significantly lower than ChatGPT’s success rate. Nonetheless, ChatGPT has a very high false positive rate of 91%. In comparison, the worst false positive rate of any traditional static code analyser is 82%. These findings highlight the promising potential of ChatGPT for improving the static code analysis process but reveal certain caveats (especially regarding accuracy) in its current state. Our findings suggest that one interesting possibility to explore in future works would be to pick the best of both worlds by combining traditional static code analysers with ChatGPT to find security vulnerabilities more effectively.

The work described in this paper aims at developing a machine learning based tool for automatic identification of vulnerabilities on programs (source, high level code), that uses an abstract syntax tree representation. It is based on FastScan, using code2seq approach. Fastscan is a recently developed system aimed capable of detecting vulnerabilities in source code using machine learning techniques. Nevertheless, FastScan is not able of identifying the vulnerability type. In the presented work the main goal is to go further and develop a method to identify specific types of vulnerabilities. As will be shown, the goal will be achieved by optimizing the model’s hyperparameters, changing the method of preprocessing the input data and developing an architecture that brings together multiple models to predict different specific vulnerabilities. The preliminary results obtained from the training stage, are very promising. The best f1 metric obtained is 93% resulting in a precision of 90% and accuracy of 85%, according to the performed tests and regarding a trained model to predict vulnerabilities of the injection type.

In Engineering and Scientific computing, a significant number of software systems that are being commonly used by the Engineering community are developed and deployed as Open Source Systems (OSS). This collaborative development model has been widely accepted in both academia and industry. As these systems are typically developed by developers who are experts in their fields (e.g., Engineering), some of those programmers may be lacking an essential skillset today's software quality and engineering concepts (e.g., source code maintainability and analyzability, source code vulnerability, parallelization, etc.). An empirical study is presented that examines some of the source code quality aspects and programming practices (e.g., jumps statements, usage of recursion) from software engineering perspective in ten open source Engineering software systems. The studied systems are written in the C/C++ programming languages and comprised of over ten million lines of code, and present numerous tasks of the engineering community, including computer-aided design, computational fluid dynamics, and finite element analysis. Static analysis methods are applied to each system to determine the number of jump statements (Goto and Break statements) and functions that are recursively invoked. Functions that contain direct or indirect recursive calls are determined and counted. The results show that number of the used jump statements and functions that have recursive calls are significant and thus pose more of a roadblock to effectively and automatically maintain, comprehend, analyze, and re-engineer systems. Lastly, this leads to inefficiencies in a software engineer's ability to accomplish these common tasks. The study showed that issues in source code stemming from the lack of coding standards have increased over time and are prevalent.

The increasing complexity of software applications and the necessity for minimizing software vulnerabilities has given rise to the use of machine learning techniques that can identify software vulnerabilities in source code. However, many of these techniques lack the accuracy needed for industrial practice. The contribution of this work is the novel use of an Abstract Syntax Tree Neural Network (ASTNN) to identify and classify software vulnerabilities in the Common Weakness Enumeration (CWE) types. We make two fundamental claims in this work. First, the use of an ASTNN performs better than prior machine learning neural network architectures. Second, the benchmark data set commonly used for machine learning vulnerability classification is flawed for this use. To illustrate these claims, we describe our ASTNN architecture and evaluate it with more than 44,000 test cases across 29 CWEs in the NIST Juliet Test Suite data set. Results show a minimum of 88% accuracy across all CWEs.