AI RESEARCH ANALYSIS
Comparing XAI Methods for Identifying Critical Features and Expert Contributions in Multi-Expert Edge Computing Attack Detection
Authored by Minh Hoang -Nguyen, Trong-Minh Hoang, and Thi Trang Linh Le. Published on 26 November 2025. DOI: 10.1145/3785520.3785524.
Executive Impact at a Glance
Key findings and quantified insights demonstrating the direct relevance and potential of this research for your enterprise AI initiatives.
0%
Contribution from Top-Performing Expert
0%
Maximum Similarity between Gradient-Based Methods
0
Consistently Influential Core Features
0
Diverse XAI Techniques Applied
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Overview & System Architecture
XAI Methodologies Compared
Feature & Expert Impact Analysis
Metric-Based XAI Evaluation
Unveiling Multi-Expert IDS in Edge Computing
In edge computing environments, Intrusion Detection Systems (IDS) are vital, but their resource limitations necessitate effective feature selection. Multi-expert systems enhance robustness, yet their decision-making can be opaque. This study addresses this by comparing Explainable AI (XAI) methods—SHAP, LIME, Integrated Gradients, and DeepLIFT—to clarify feature importance and expert contributions in a multi-expert feature selection system for IDS. We evaluate these methods using quantitative metrics like Hamming Distance, Jaccard Index, Spearman Rank, Pearson Correlation, and Cosine Similarity to assess consistency and differences in explanations.
Enterprise Process Flow
IoT Device
→
Dataset
→
Preprocessing Data
→
Multi-expert system for feature selection
→
STACKING
→
Evaluation and comparison of explainability methods
→
Identification of influential features and base experts in the decision of the ensemble system
Diving into Explainable AI Methodologies
This study selected four representative explanation methods: SHAP, LIME, DeepLIFT, and Integrated Gradients. This selection ensures diversity in approaches, covering both model-agnostic techniques (LIME, SHAP) and model-specific methods for deep neural networks (DeepLIFT, Integrated Gradients). These methods allow for both local explanations of individual feature contributions and global synthesis of feature importance across the entire system, essential for understanding multi-expert decision-making in edge computing IDS.
| Method | Type | Core Mechanism | Key Advantage | Observed Limitation |
|---|---|---|---|---|
| SHAP | Model-agnostic | Shapley values (game theory) | Fairly allocates contribution, unified approach | Limitations adapting to temporal variations, distinct feature sets in results |
| LIME | Model-agnostic | Local linear approximation | Fast, simple local explanations | Instability with heterogeneous data, sensitive to noise |
| DeepLIFT | Model-specific (DNN) | Backpropagates activation differences | Overcomes zero gradients, clear explanation values | Divergence from Integrated Gradients despite family, reliance on baseline |
| Integrated Gradients | Model-specific (DNN) | Path integral of gradients | Addresses zero gradients, axiomatic properties | Prone to bias with single baseline, divergence from DeepLIFT |
Identifying Critical Features and Leading Experts
Our analysis revealed clear patterns in how experts contribute and which features are most influential in the multi-expert feature selection system. Understanding these impacts is crucial for optimizing IDS performance in resource-constrained edge environments.
SelectKBest
Dominant Expert in Multi-Expert System
With stable influence values ranging from 0.25 to 0.27 across all four explanation methods, SelectKBest consistently plays a central role in the final decision-making process. This dominance suggests potential for system optimization through prioritizing or re-weighting core experts.
Variance, TCP Flags, HTTPS, Traffic Volume
Core Influential Features for Attack Detection
These features (e.g., ack_flag_number, rst_count, Tot size, Tot sum) consistently emerged as highly influential across multiple XAI methods. Their robustness indicates their critical role in distinguishing between normal and malicious traffic in edge computing environments.
Optimizing Edge IDS: A Case for Prioritized Experts and Features
The clear stratification of expert influence and the robust identification of core features provide a strategic roadmap for enhancing Intrusion Detection Systems (IDS) on edge computing platforms. By emphasizing outputs from high-impact experts like SelectKBest and focusing on data from critical features such as traffic variability (Variance) and TCP flags, system designers can optimize resource allocation. This targeted approach allows for improved detection accuracy and reduced computational overhead, which are vital for the resource-constrained nature of edge devices. Leveraging diverse XAI techniques also enables cross-validation and bias reduction, leading to more reliable and transparent IDS decisions. This study confirms that strategic use of explainability enhances both the performance and interpretability of multi-expert systems.
Quantitative Evaluation of XAI Method Consistency
The consistency and similarity of XAI methods were quantitatively assessed using Hamming Distance, Jaccard Index, Spearman Rank, Pearson Correlation, and Cosine Similarity. These metrics provide a multi-dimensional view of how different explanation methods align or diverge in their feature explanations.
| XAI Pair | Hamming Distance (Feature Set Disagreement) | Jaccard Index (Feature Set Overlap) | Spearman Rank (Ranking Consistency) | Pearson Correlation (Attribution Value Linearity) | Cosine Similarity (Directional Alignment) |
|---|---|---|---|---|---|
| SHAP vs LIME | Moderate (0.32) | Moderate (0.43) | Moderate (0.50) | Moderate (0.30) | Moderate (0.62) |
| SHAP vs Gradient-Based (Avg) | Moderate-High (0.48-0.56) | Low (0.18-0.25) | Low (-0.01 to 0.15) | Low (0.06-0.18) | Moderate (0.60-0.64) |
| LIME vs Gradient-Based (Avg) | Moderate-High (0.40-0.48) | Low (0.25-0.33) | Low (0.00-0.17) | Low (0.16-0.34) | High (0.72-0.77) |
| Integrated Gradients vs DeepLIFT | Very Low (0.08) | Very High (0.82) | Near Perfect (0.96) | Near Perfect (0.96) | Near Perfect (0.99) |
Estimate Your AI ROI
Understand the potential financial impact of implementing advanced AI solutions in your enterprise. Adjust the parameters below to see estimated annual savings and reclaimed human hours.
Estimated Annual Savings
$0
Annual Hours Reclaimed
0
Your AI Implementation Roadmap
A phased approach to integrate these advanced AI insights into your enterprise operations, ensuring a smooth transition and measurable impact.
Phase 1: Discovery & Strategy Alignment
Initial consultations to understand your specific business challenges, data landscape, and strategic objectives. We define key performance indicators and align AI solutions with your overarching enterprise goals.
Phase 2: Pilot Program & Proof of Concept
Develop and implement a targeted pilot program based on the research insights. This includes data preparation, model training, and initial deployment within a controlled environment to demonstrate tangible value and gather feedback.
Phase 3: Full-Scale Integration & Optimization
Expand the successful pilot to full enterprise-wide integration, ensuring scalability, security, and seamless workflow. Continuous monitoring and iterative optimization cycles are established for sustained performance and evolving needs.
Phase 4: Performance Monitoring & Future-Proofing
Establish robust monitoring frameworks and reporting mechanisms. Provide ongoing support, training, and strategic advisory to adapt to new technologies and market demands, securing your long-term AI advantage.
Ready to Transform Your Enterprise with AI?
Leverage cutting-edge research to build intelligent, explainable, and efficient systems. Book a complimentary consultation with our AI strategists.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat