Browsing by Author "Usman Javed Butt"

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    Using XAI Techniques to Detect Targeted Data Poisoning Attacks on Healthcare Applications of Machine Learning Systems
    (Springer Cham, 2025) Eyad Dhaher Megdadi; Usman Javed Butt
    This research study explores the application of Explainable Artificial Intelligence (XAI) methods for detecting targeted data poisoning attacks in healthcare machine learning systems. As machine learning becomes increasingly integrated into critical fields like healthcare, the integrity and security of training data have become paramount concerns. Data poisoning attacks, which manipulate training datasets to influence model behaviour, pose a significant threat to the reliability and effectiveness of these systems. Our study presents a novel approach that leverages XAI techniques, particularly focusing on global explanations of selected features, to identify signs of data manipulation. We propose a method of monitoring the impact level of carefully chosen features as an indicator of potential poisoning, using predetermined thresholds to trigger warnings when unusual patterns are detected. The research methodology involves applying global explanation method, to measure and monitor features impact in healthcare datasets, then explore the effectiveness of this approach using a case study on hypothyroid diagnosis, where data poisoning could lead to delayed treatment with potentially life-threatening consequences. Research findings suggest that XAI techniques can provide valuable insights into the behaviour of machine learning models, enabling more effective detection of subtle, targeted poisoning attacks. However, we also acknowledge limitations, including the need for some prior knowledge of potential attack goals and the risk of false positives or negatives.