Technical Summary: Explainable AI for Next-Generation Wireless Physical Layer

1. Problem Statement

The transition to sixth-generation (6G) mobile networks envisions an "AI-native" architecture where deep neural networks (DNNs) are deeply integrated into the physical (PHY) layer to optimize spectral efficiency, latency, and autonomy. However, the inherent opacity of deep learning models raises critical concerns regarding system reliability, safety, and trustworthiness, particularly in dynamic, time-varying wireless environments.

Current Explainable AI (XAI) techniques, largely developed for computer vision and natural language processing, are often insufficient for wireless communications. They frequently produce raw pixel- or token-level saliency maps that lack "radio semantics" (e.g., channel statistics, modulation patterns, beam states). Furthermore, many XAI methods are computationally intensive, conflicting with the strict latency and resource constraints of real-time PHY processing. There is a lack of systematic frameworks that align XAI techniques with specific wireless learning tasks (e.g., channel estimation, beam selection) and responsible AI goals (e.g., causality, fairness, privacy).

2. Methodology and Framework

This survey adopts a responsibility-oriented perspective to analyze XAI within the wireless PHY layer. The methodology involves:

• Formalization of Goals: Defining a set of responsibility-driven goals for wireless XAI, including trustworthiness, causality, privacy, fairness, transferability, informativeness, and confidence.
• Systematic Taxonomy: Categorizing XAI approaches based on model structure (transparent vs. post-hoc) and learning paradigms (supervised, reinforcement, transfer, and meta-learning).
• Task-Specific Analysis: Mapping representative XAI techniques to specific PHY learning tasks (channel estimation, modulation classification, anti-jamming, beam selection) and MAC/RAN resource control functions.
• Evaluation Criteria: Establishing communication-aware evaluation metrics that go beyond standard fidelity, incorporating stability under channel perturbations, computational cost, and alignment with radio semantics.

3. Key Contributions

A. Responsibility-Driven XAI Goals for Wireless Systems

The paper establishes a conceptual linkage between interpretability and key responsible-AI objectives tailored to wireless contexts:

• Trustworthiness: Ensuring models act reliably under tight latency and imperfect observations, preventing misplaced confidence in fragile explanations.
• Causality: Distinguishing true cause-and-effect relationships (e.g., deep fading vs. interference) from spurious correlations to enhance robustness against domain shifts.
• Privacy: Using explainability to assess and mask sensitive data representations (e.g., user identity in WiFi CSI) to prevent leakage.
• Fairness: Identifying and mitigating biases in resource allocation (e.g., cell-edge user starvation) by exposing proxy variables in decision-making.
• Transferability: Clarifying the boundaries of model applicability across different frequency bands, hardware, and environments.
• Informativeness: Converting explanations into operational signals (e.g., mapping signal degradation to physical causes like antenna misalignment).
• Confidence: Quantifying uncertainty to enable deferral mechanisms (e.g., triggering HARQ re-transmissions in low-SNR regimes).

B. Systematic Taxonomy of XAI Techniques

The survey organizes XAI methods into three primary categories, detailing their applicability to wireless tasks:

1. Transparent/White Box Models: Linear models, Generalized Additive Models (GAMs), and Decision Trees, which offer inherent interpretability.
2. Explanations of NN Behaviors:
   • Local/Global Model Reduction: Techniques like LIME, SHAP, and knowledge distillation to approximate complex DNNs with simpler surrogates.
   • Automatic Rule Extraction: Methods to distill decision processes into Boolean rules (decompositional and pedagogical approaches).
3. Explanations of NN Representations:
   • Layer/Unit/Block Analysis: Using tools like CAM, Grad-CAM, and DeepLIFT to visualize feature maps and neuron activations.
   • Physical Representations: Integrating physics-informed models (e.g., Split-Step Fourier Method) to ground explanations in physical laws.
4. Explanation-Producing Networks: Attention mechanisms and disentangled networks that inherently highlight relevant inputs.
5. Explainable Reinforcement Learning (XRL): Reward decomposition, learning explainable policies (e.g., decision trees, symbolic policies), and post-hoc enhancement for RL agents.

C. Application to Wireless PHY and MAC/RAN Layers

The paper provides a comprehensive review of how these techniques apply to specific wireless scenarios:

• Supervised Learning (PHY):
  • Channel Estimation: Using perturbation-based noise masks (e.g., XAI-CHEST) to identify relevant subcarriers; applying SHAP and DeepLIFT to unfolded SBL or LDAMP networks to trace contributions of iterative steps.
  • Modulation Classification: Utilizing Concept Bottleneck models to map raw signals to physical concepts; applying Integrated Gradients and CKA (Centered Kernel Alignment) to analyze temporal dependencies and fused representations in LSTM/Transformer models.
  • Anti-Jamming: Using Grad-CAM to visualize time-frequency patches contributing to jamming detection.
  • Beam Selection: Applying Grad-CAM++ and SHAP to attribute beam choices to specific angle-delay regions or sensor modalities.
• Reinforcement Learning (PHY & MAC):
  • Resource Control: Using reward decomposition (drQ) to expose trade-offs in Modulation and Coding Scheme (MCS) selection; employing symbolic policy distillation (DSP) to convert DRL policies into human-readable rules for beam selection and scheduling.
  • Scheduling: Using Dot-to-Dot and Reward Processing to trace state-action trajectories and temporal dependencies in DRQN-based schedulers.
• Transfer and Meta-Learning:
  • Analyzing how representations shift during transfer (using CKA and t-SNE) and how meta-learners adapt to new tasks (e.g., few-shot AMC, rapid channel adaptation).

D. Standards and Projects

The survey reviews existing standards (IEEE P2976, ETSI, EU AI Act) and ongoing industrial projects (TEFL, SliceOps, XAInomaly, EXPLORA) that integrate XAI into 6G network slicing, anomaly detection, and RAN control.

4. Results and Findings

• No Universal Solution: No single XAI technique is universally suitable for all PHY tasks. Attribution-based methods (SHAP, LIME) are effective for structured inputs like CSI tensors, while representation analysis is better for understanding invariances in transfer learning.
• Radio Semantics are Critical: Explanations must be mapped to domain-specific semantics (e.g., pilots, subcarriers, beam indices) rather than generic feature importance to be useful for communication engineers.
• Trade-offs Exist: There is a non-monotonic trade-off between explainability and communication utility (spectral efficiency, latency). Adding explainability blocks can reduce model capacity or introduce overhead, though it can also improve performance by enabling timely debugging.
• Current Gaps: Most existing efforts apply vision/language XAI tools directly to wireless tasks without adapting to radio constraints. There is insufficient work on cross-layer explanation consistency and explaining emerging LLM/Agentic AI systems in wireless contexts.

5. Significance and Future Directions

The paper argues that for 6G to be truly "AI-native," explainability must evolve from a post-hoc audit tool to an integral component of the design process. The significance of this work lies in:

• Bridging the Gap: Connecting abstract XAI theories with the concrete, constrained realities of wireless physical layer operations.
• Enabling Trust: Providing the necessary transparency for stakeholders (engineers, regulators, operators) to trust AI-driven decisions in safety-critical applications like autonomous driving and remote surgery.
• Guiding Future Research: Identifying open challenges, including:
  • Explainability-Performance Tradeoffs: Theoretical limits of utility vs. interpretability.
  • Customized XAI: Developing algorithms specifically for I/Q samples and CSI tensors.
  • Cross-Layer Consistency: Aligning explanations across PHY, MAC, and network layers.
  • LLM and Agentic AI: Addressing the unique challenges of explaining large language models and autonomous agents in real-time wireless control loops.

The authors conclude that future XAI for wireless systems must be lightweight, real-time capable, and rigorously evaluated against communication-specific metrics to ensure responsible and trustworthy deployment.