Ilir Gashi
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Ilir Gashi has a total of 58 co-authors. Here are the top co-authors by number of shared publications:
- Vladimir Stankovic
University of Strathclyde
8 shared publications Michel Cukier
6 shared publications- Peter Popov
City University London
6 shared publications - Alysson Bessani
Universidade de Lisboa
5 shared publications Areej Algaith
4 shared publicationsBertrand Sobesto
A. James Clark Sch. of Eng., Univ. of Maryland, College Park, MD, USA
3 shared publications
Latest Publications
- 8 August 2026
- Published by Springer Nature
- p. 296-311
The Statistical Assessment of Bayes–“sub”Optimal Binary Machine Learning Classifier Risk
- 8 August 2026
- Published by Springer Nature
- p. 296-311
The Statistical Assessment of Bayes–“sub”Optimal Binary Machine Learning Classifier Risk
- 8 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Technology and Society Magazine
- Vol. 44 (2) , 77-84
Privacy-Preserving Federated Learning for Phishing Detection
Machine learning is one of the most prominent technologies used to combat phishing detection; however, the vast amount of data required for training models for detection raises a privacy concern for end users. Gathering email or document data may very well contain private information, and the machine learning models learn from the words and other attributes of these text-based documents. Gathering this information in a centralized location and using it to train models could pose a security risk on all levels of data acquisition, from the transfer of the data to the storage. Federated learning is emerging as a promising alternative to traditionally centralized machine learning for phishing detection. The advantages of federated learning, mainly in privacy and scalability, are weighed against the issue of detection accuracy. Federated learning provides the ability to train models without the transfer of sensitive data, more or less no raw data from the device, and allows the training to be done locally; this eliminates the privacy exposure accompanied by traditional machine learning models that operate in a centralized manner. However, this alone is not enough to comply with privacy regulations, such as General Data Protection Regulation (GDPR) and the European Union (EU) Artificial Intelligence Act (AI Act), and privacy-preserving technology must be used in conjunction to ensure federated learning's compliance with privacy regulations. This article is a dedication to Professor Thomas Chen's aspirations in the field of cybersecurity. This article is dedicated to his memory.
- 8 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Technology and Society Magazine
- Vol. 44 (2) , 77-84
Privacy-Preserving Federated Learning for Phishing Detection
Machine learning is one of the most prominent technologies used to combat phishing detection; however, the vast amount of data required for training models for detection raises a privacy concern for end users. Gathering email or document data may very well contain private information, and the machine learning models learn from the words and other attributes of these text-based documents. Gathering this information in a centralized location and using it to train models could pose a security risk on all levels of data acquisition, from the transfer of the data to the storage. Federated learning is emerging as a promising alternative to traditionally centralized machine learning for phishing detection. The advantages of federated learning, mainly in privacy and scalability, are weighed against the issue of detection accuracy. Federated learning provides the ability to train models without the transfer of sensitive data, more or less no raw data from the device, and allows the training to be done locally; this eliminates the privacy exposure accompanied by traditional machine learning models that operate in a centralized manner. However, this alone is not enough to comply with privacy regulations, such as General Data Protection Regulation (GDPR) and the European Union (EU) Artificial Intelligence Act (AI Act), and privacy-preserving technology must be used in conjunction to ensure federated learning's compliance with privacy regulations. This article is a dedication to Professor Thomas Chen's aspirations in the field of cybersecurity. This article is dedicated to his memory.
- Open Access
- 22 October 2021
- Published by Springer Nature in Empirical Software Engineering
- Vol. 27 (1) , 1-30
Dynamical analysis of diversity in rule-based open source network intrusion detection systems
Diverse layers of defence play an important role in the design of defence-in-depth architectures. The use of Intrusion Detection Systems (IDSs) are ubiquitous in this design. But the selection of the “right” IDSs in various configurations is an important decision that the security architects need to make. Additionally, the ability of these IDSs to adapt to the evolving threat-landscape also needs to be investigated. To help with these decisions, we need rigorous quantitative analysis. In this paper, we present a diversity analysis of open-source IDSs, Snort and Suricata, to help security architects tune/deploy these IDSs. We analyse two types of diversities in these IDSs; configurational diversity and functional diversity. In the configurational diversity analysis, we investigate the diversity in the sets of rules and the Blacklisted IP Addresses (BIPAs) these IDSs use in their configurations. The functional diversity analysis investigates the differences in alerting behaviours of these IDSs when they analyse real network traffic, and how these differences evolve. The configurational diversity experiment utilises snapshots of the rules and BIPAs collected over a period of 5 months, from May to October 2017. The snapshots have been collected for three different off-the-shelf default configurations of the Snort IDS and the Emerging Threats (ET) configuration of the Suricata IDS. The functional diversity investigates the alerting behaviour of these two IDSs for a sample of the real network traffic collected in the same time window. Analysing the differences in these systems allows us to get insights into where the diversity in the behaviour of these systems comes from, how does it evolve and whether this has any effect on the alerting behaviour of these IDSs. This analysis gives insight to security architects on how they can combine and layer these systems in a defence-in-depth deployment.
- 22 October 2021
- Published by Springer Nature in Empirical Software Engineering
- Vol. 27 (1) , 1-30
Open AccessDynamical analysis of diversity in rule-based open source network intrusion detection systems
Diverse layers of defence play an important role in the design of defence-in-depth architectures. The use of Intrusion Detection Systems (IDSs) are ubiquitous in this design. But the selection of the “right” IDSs in various configurations is an important decision that the security architects need to make. Additionally, the ability of these IDSs to adapt to the evolving threat-landscape also needs to be investigated. To help with these decisions, we need rigorous quantitative analysis. In this paper, we present a diversity analysis of open-source IDSs, Snort and Suricata, to help security architects tune/deploy these IDSs. We analyse two types of diversities in these IDSs; configurational diversity and functional diversity. In the configurational diversity analysis, we investigate the diversity in the sets of rules and the Blacklisted IP Addresses (BIPAs) these IDSs use in their configurations. The functional diversity analysis investigates the differences in alerting behaviours of these IDSs when they analyse real network traffic, and how these differences evolve. The configurational diversity experiment utilises snapshots of the rules and BIPAs collected over a period of 5 months, from May to October 2017. The snapshots have been collected for three different off-the-shelf default configurations of the Snort IDS and the Emerging Threats (ET) configuration of the Suricata IDS. The functional diversity investigates the alerting behaviour of these two IDSs for a sample of the real network traffic collected in the same time window. Analysing the differences in these systems allows us to get insights into where the diversity in the behaviour of these systems comes from, how does it evolve and whether this has any effect on the alerting behaviour of these IDSs. This analysis gives insight to security architects on how they can combine and layer these systems in a defence-in-depth deployment.
- 1 September 2021
- Published by Elsevier in Computers & Security
- Vol. 108
Waste not: Using diverse neural networks from hyperparameter search for improved malware detection
- 1 September 2021
- Published by Elsevier in Computers & Security
- Vol. 108
Waste not: Using diverse neural networks from hyperparameter search for improved malware detection
- 6 August 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. PP (15455971) , 1
Predicting the Discovery Pattern of Publically Known Exploited Vulnerabilities
Vulnerabilities with publically known exploits typically form 2-7% of all vulnerabilities reported for a given software version. With a smaller number of known exploited vulnerabilities compared with the total number of vulnerabilities, it is more difficult to model and predict when a vulnerability with a known exploit will be reported. In this paper, we introduce an approach for predicting the discovery pattern of publically known exploited vulnerabilities using all publically known vulnerabilities reported for a given software. Eight commonly used vulnerability discovery models (VDMs) and one neural network model (NNM) were utilized to evaluate the prediction capability of our approach. We compared their predictions results with the scenario when only exploited vulnerabilities were used for prediction. Our results show that, in terms of prediction accuracy, out of eight software we analyzed, our approach led to more accurate results in seven cases. Only in one case, the accuracy of our approach was worse by 1.6%.
- 6 August 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. PP (15455971) , 1
Predicting the Discovery Pattern of Publically Known Exploited Vulnerabilities
Vulnerabilities with publically known exploits typically form 2-7% of all vulnerabilities reported for a given software version. With a smaller number of known exploited vulnerabilities compared with the total number of vulnerabilities, it is more difficult to model and predict when a vulnerability with a known exploit will be reported. In this paper, we introduce an approach for predicting the discovery pattern of publically known exploited vulnerabilities using all publically known vulnerabilities reported for a given software. Eight commonly used vulnerability discovery models (VDMs) and one neural network model (NNM) were utilized to evaluate the prediction capability of our approach. We compared their predictions results with the scenario when only exploited vulnerabilities were used for prediction. Our results show that, in terms of prediction accuracy, out of eight software we analyzed, our approach led to more accurate results in seven cases. Only in one case, the accuracy of our approach was worse by 1.6%.
- 1 November 2019
- Published by Elsevier in Computers & Security
- Vol. 87
Vulnerability prediction capability: A comparison between vulnerability discovery models and neural network models
- 1 November 2019
- Published by Elsevier in Computers & Security
- Vol. 87
Vulnerability prediction capability: A comparison between vulnerability discovery models and neural network models
- 1 December 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 269-278
Detecting Malicious Web Scraping Activity: A Study with Diverse Detectors
We present results on the use of diverse monitoring tools for the detection of malicious web scraping activity. We have carried out an analysis of a real dataset of Apache HTTP Access logs for an e-commerce application provided by a large multinational IT provider for the global travel and tourism industry. Two tools have been used to detect scraping activities based on the HTTP requests: a commercial tool, and an in-house tool called Arcane. We show the benefits that can be achieved through the use of both systems, in terms of overall sensitivity and specificity, and we discuss the potential sources of diversity between the tool's alert patterns.
- 1 December 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 269-278
Detecting Malicious Web Scraping Activity: A Study with Diverse Detectors
We present results on the use of diverse monitoring tools for the detection of malicious web scraping activity. We have carried out an analysis of a real dataset of Apache HTTP Access logs for an e-commerce application provided by a large multinational IT provider for the global travel and tourism industry. Two tools have been used to detect scraping activities based on the HTTP requests: a commercial tool, and an in-house tool called Arcane. We show the benefits that can be achieved through the use of both systems, in terms of overall sensitivity and specificity, and we discuss the potential sources of diversity between the tool's alert patterns.
- Open Access
- 25 September 2018
- Published by Springer Nature in Computing
- Vol. 101 (2) , 139-160
Cluster-based vulnerability assessment of operating systems and web browsers
- 25 September 2018
- Published by Springer Nature in Computing
- Vol. 101 (2) , 139-160
Open AccessCluster-based vulnerability assessment of operating systems and web browsers
- 1 September 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 57-64
Finding SQL Injection and Cross Site Scripting Vulnerabilities with Diverse Static Analysis Tools
The use of Static Analysis Tools (SATs) is mandatory when developing secure software and searching for vulnerabilities in legacy software. However, the performance of the various SATs concerning the detection of vulnerabilities and false alarm rate is usually unknown and depends on many factors. The simultaneous use of several tools should increase the detection capabilities, but also the number of false alarms. In this paper, we study the problem of combining several SATs to best meet the developer needs. We present results of analyzing the performance of diverse static analysis tools, based on a previously published dataset that resulted from the use of five diverse SATs to find two types of vulnerabilities, namely SQL Injections (SQLi) and Cross-Site Scripting (XSS), in 132 plugins of the WordPress Content Management System (CMS). We present the results based on well-established measures for binary classifiers, namely sensitivity and specificity for all possible diverse combinations that can be constructed using these 5 SAT tools. We then provide empirically supported guidance on which combinations of SAT tools provide the most benefits for detecting vulnerabilities with low false positive rates.
- 1 September 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 57-64
Finding SQL Injection and Cross Site Scripting Vulnerabilities with Diverse Static Analysis Tools
The use of Static Analysis Tools (SATs) is mandatory when developing secure software and searching for vulnerabilities in legacy software. However, the performance of the various SATs concerning the detection of vulnerabilities and false alarm rate is usually unknown and depends on many factors. The simultaneous use of several tools should increase the detection capabilities, but also the number of false alarms. In this paper, we study the problem of combining several SATs to best meet the developer needs. We present results of analyzing the performance of diverse static analysis tools, based on a previously published dataset that resulted from the use of five diverse SATs to find two types of vulnerabilities, namely SQL Injections (SQLi) and Cross-Site Scripting (XSS), in 132 plugins of the WordPress Content Management System (CMS). We present the results based on well-established measures for binary classifiers, namely sensitivity and specificity for all possible diverse combinations that can be constructed using these 5 SAT tools. We then provide empirically supported guidance on which combinations of SAT tools provide the most benefits for detecting vulnerabilities with low false positive rates.
- 17 August 2018
- Published by Springer Nature
- p. 267-281
Diversity in Open Source Intrusion Detection Systems
- 17 August 2018
- Published by Springer Nature
- p. 267-281
Diversity in Open Source Intrusion Detection Systems