Conversational XAI Outperforms Dashboards for UAV Intrusion Detection Trust
A recent study published on arXiv (2608.10434) investigates the effectiveness of conversational explainable AI (XAI) versus dashboard-based XAI for detecting intrusions in unmanned aerial vehicle (UAV) networks. This research tackles the interpretability issues associated with machine learning-driven intrusion detection systems (IDS), which, while efficient, often function as 'black boxes' due to the complexity of high-dimensional cyber-physical data. Traditional static dashboards may not effectively convey intricate feature relationships to users. To address this, the authors introduce a conversational XAI interface utilizing large language models (LLMs) for real-time inquiries. In a controlled setting, participants engaged in post-incident audits using either the conversational tool or a conventional XAI dashboard, assessing operator comprehension, trust, and dependence. The results suggest that the conversational interface was viewed as more beneficial, indicating promising advantages of LLM-based explanations in security operations for UAV networks.
Key facts
- Study from arXiv:2608.10434 compares conversational XAI with dashboard XAI for UAV intrusion detection.
- Machine learning-based IDS are effective but lack interpretability due to black-box nature.
- High-dimensional multimodal cyber-physical data poses interpretability challenges.
- Static dashboards may struggle to present complex relationships among features.
- Proposed conversational XAI interface powered by large language models (LLMs).
- Controlled experiment with participants evaluated operator understanding, trust, and reliance.
- Tasks involved post-incident auditing.
- Conversational interface was perceived as more useful than the dashboard.
Entities
Institutions
- arXiv