A Unified XAI-NSAI Analytical Framework for Trustworthy IoT Anomaly Detection

Hugues Leger, Giovani R. AbuaitaharXiv 2026 · arXiv

TL;DR

This paper introduces a unified analytical framework that bridges the gap between Explainable AI (XAI) and Neuro-Symbolic AI (NSAI) for IoT anomaly detection. By evaluating systems across five key dimensions—ranging from architectural design to user-centered support—it provides a structured roadmap for building IoT security solutions that are not just accurate, but also transparent and domain-aware.

What Problem Does This Paper Address?

Internet of Things (IoT) devices are now backbone components of critical infrastructure, from smart power grids to intensive care units. In these high-stakes environments, detecting anomalies (like sensor failures or cyberattacks) is vital. However, modern Deep Learning models are often "black boxes"—they might tell you that something is wrong without explaining why.

While Explainable AI (XAI) helps identify which data features triggered an alert, it often lacks "semantic grounding"—it doesn't know if an alert makes sense according to the laws of physics or medical rules. Conversely, Neuro-Symbolic AI (NSAI) incorporates domain knowledge (rules and logic) but often fails to provide human-friendly explanations. Furthermore, IoT systems face unique hardware constraints (low power, limited memory) that many AI models ignore. The authors address the lack of a cohesive way to evaluate and design systems that balance these competing needs for trust, reasoning, and efficiency.

Methods

The authors developed a unified analytical framework based on a structured review of recent literature (54 papers analyzed, with 16 representative models examined in depth). The core of their contribution is the Five Design Dimensions, which allow researchers to create a "Capability Profile" for any anomaly detection system:

  1. D1: Explanation Scope & Stage: When is the explanation generated (during or after detection) and what does it cover (local alerts vs. global model behavior)?
  2. D2: Symbolic Reasoning Integration: How deeply is domain knowledge (rules, ontologies) baked into the AI?
  3. D3: Architectural Design: Are the components sequential (a pipeline) or tightly integrated?
  4. D4: Deployment Context: Where does the code run? (Centralized cloud vs. resource-constrained edge devices).
  5. D5: User-Centered Explainability: Who is the explanation for? (Technical developers vs. operational staff like doctors or engineers).
The proposed unified XAI-NSAI framework structure.

To make the framework actionable, they proposed a Readiness Scoring system (0-3 scale) to qualitatively assess how well a system supports each dimension.

Datasets

As this is a framework and methodology paper, the authors do not introduce a new benchmark dataset. However, they demonstrate the framework's "design utility" by applying it to a hypothetical Medical IoT (MIoT) system designed for multivariate healthcare time-series. In this illustration, they reference the use of the MIMIC-IV ICU dataset for monitoring vital signs like heart rate and SpO2.

Key Insights and Ideas

  • The Synergy of XAI and NSAI: The authors argue that XAI and NSAI are two sides of the same coin. XAI tells you what influenced a decision (the "what"), while NSAI assesses if that decision is consistent with known rules (the "why").
  • The "Post-Hoc" Trap: A key finding from their literature review is that most current systems rely on "post-hoc" XAI (like SHAP or LIME). These are essentially "afterthoughts" that don't influence the actual detection logic, which can limit their reliability in critical scenarios.
  • Deployment is a Trust Factor: Trust isn't just about accuracy; it's about availability. If an explanation mechanism is too computationally heavy to run on an edge device, it becomes useless in real-time IoT environments.
  • User Gaps: There is a significant lack of "user-adaptive" explanations. Most current tools provide technical feature attributions that are useful for data scientists but potentially confusing for the domain experts (like clinicians) who actually need to make decisions based on the data.

Key Results

The authors applied their framework to three prominent existing studies to highlight "capability gaps":

  • Rathi & Shandilya [26]: Scored high on deployment (D4) because it was built for the edge, but scored 0 on symbolic reasoning (D2) because it lacked domain rules.
  • Bizzarri et al. [19]: Showed deep logic integration (D2/D3) but lacked user-facing explanations (D5) and hadn't been validated for edge deployment (D4).

They then used the framework to guide the design of a neuro-symbolic healthcare system, shown in Figure 3. This design illustrates how to align neural detection (BiLSTM Autoencoders) with physiological rules (e.g., "heart rate cannot jump 50bpm in one second") to provide "clinician-interpretable" feedback.

Fig. 3. Illustrative framework guided design for multivariate healthcare time-series anomaly and medical event detection. The five dimensions of the proposed unified XAI–NSAI framework inform design choices. This is merely a design illustration rather than an empirical evaluation.

Limitations and Future Work

  • Qualitative Nature: The readiness scores (0-3) involve interpretive judgment. While useful for comparison, they are not yet a validated quantitative metric for "trustworthiness."
  • Integration Challenges: The authors note that truly integrating symbolic reasoning into the learning process (rather than just using it as a post-filter) remains a difficult research challenge.
  • Standardization: There is a pressing need for standardized evaluation methods that measure not just accuracy, but also "explanation faithfulness" and "reasoning consistency" across distributed IoT environments.

Summary updated

Abstract

The widespread deployment of Internet of Things (IoT) devices in critical environments makes anomaly detection essential for identifying abnormal behaviors and supporting timely investigation and response. However, accurate detection alone is insufficient. Practical IoT deployments must also provide explanations, incorporate domain knowledge and reasoning, operate under IoT constraints, and support human decision making. Although recently proposed approaches increasingly incorporate Explainable Artificial Intelligence (XAI) or Neuro-Symbolic Artificial Intelligence (NSAI), XAI and NSAI methods are often considered independently. We argue that they should be considered together to better understand their complementary roles in supporting trustworthy IoT anomaly detection. To address this gap, we propose a unified analytical framework organized around five complementary dimensions. The framework provides a structured basis for characterizing existing approaches, identifying capability gaps and design trade-offs, and guiding the design of new anomaly detection systems for different IoT applications. We demonstrate the framework's analytical utility by applying it to three representative studies and illustrate its design utility through its application to a Medical IoT anomaly detection system.
BibTeX
@misc{leger2026unified,
  author = {Hugues Leger and Giovani R. Abuaitah},
  title = {{A Unified XAI-NSAI Analytical Framework for Trustworthy IoT Anomaly Detection}},
  year = {2026},
  eprint = {2610.41424},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.41424},
}
Figure