By Naila Azam, George Loukas and Manos Panaousis — Centre for Sustainable Cyber Security, University of Greenwich
July 2026
A new publication by Naila Azam, George Loukas and Manos Panaousis explores the trade-off between AI explainability and privacy preservation in the context of insider threat detection.
The paper investigates how counterfactual explanations can support the transparency of AI-based decision-making while reducing the risk of unnecessarily exposing sensitive information. To address this challenge, the study introduces a privacy-aware framework that combines Differential Privacy with two new proposed metrics, XAIStrength and XAILeakage, to balance explanation fidelity against privacy risk.
The approach was evaluated using the CERT insider threat dataset. The results show that it can substantially reduce information leakage under different privacy requirements while preserving useful explanations.
The findings demonstrate that transparency and privacy do not need to be treated as competing goals, providing a practical step towards safer, more trustworthy and explainable AI for security-critical applications.
This research is part of GANNDALF’s objective of developing explainable and trustworthy AI technologies for cybercrime prevention and investigation, contributing to the project’s work towards more privacy-aware and trustworthy cybersecurity solutions.
Publication details
Title: Balancing Explainability Strength and Privacy in Counterfactual Explanations for Insider Threat Detection
Authors: Naila Azam, George Loukas and Manos Panaousis
Affiliation: Centre for Sustainable Cyber Security, University of Greenwich
Conference: IEEE Conference on Communications and Network Security (IEEE CNS) 2026
Location: Newark, Delaware, USA
Dates: 14–17 September 2026
Read the publication on Zenodo:
https://zenodo.org/records/21650034