Source-linked AI summary
Adaptive Distributed Physical-Layer Authentication and Attack Detection in 6G Non-Terrestrial Networks via Causal Meta-Learning
Parsa Rajabi, Mohammad Reza Abedi, Nader Mokari, Paeiz Azmi, Halim Yanikomeroglu
TL;DR
NTN physical-layer authentication must handle severe channel dynamics, distribution shifts, and limited generalization from single-feature methods. SAFA-MZ combines causal modeling, multi-feature distributed fingerprinting, meta-learning, and conditional TDOA verification. It achieves 92% accuracy and 96% AUC while outperforming baseline and single-feature methods.
Problem
NTN PLA faces severe Doppler shifts, long delays, fast channel variation, pilot contamination, and distribution shifts, while existing methods often rely on single features or generalize poorly to unseen environments.
Method
SAFA-MZ combines SCM, backdoor adjustment, IRM, MAML, multi-feature fingerprints, and two-stage local-to-distributed authentication for adaptive DPLA.
Results
92% accuracy and 96% AUC were achieved by SAFA-MZ, outperforming baseline and single-feature methods.
Takeaways & Limitations
The framework supports environment-invariant authentication features, rapid adaptation with few labeled samples, and reduced overhead by activating distributed verification only when needed.
Abstract
from arXiv · showhide
Physical-layer authentication (PLA) in non-terrestrial networks (NTNs) is challenged by severe Doppler shifts, long delays, and fast channel variations, which cause distribution shifts and degrade conventional learning methods. Existing PLA schemes often rely on single features or generalize poorly to unseen environments. This paper proposes a secure adaptive framework for authentication in multi-zone networks (SAFA-MZ), a causal meta-learning framework for distributed PLA (DPLA) in NTNs. First, we design a multi-feature fingerprint that combines spatial, angular, combiner, subspace, and Doppler-delay features. The fingerprint is adaptive and distributed, as it fuses heterogeneous physical-layer features and measurements from multiple aerial nodes. Second, we formulate a structural causal model (SCM) to capture the relations among design choices, environmental factors, extracted features, and authentication outcomes. Third, we develop a model-agnostic meta-learning (MAML) strategy with invariant risk minimization (IRM) and causal consistency regularization for fast adaptation to unseen NTN environments with few labeled samples. Fourth, we propose a two-stage authentication scheme that performs local recognition and activates time-difference-of-arrival (TDOA) localization with a graph attention (GAT) network only when needed, which reduces backhaul overhead. Simulations show that SAFA-MZ achieves 92% accuracy and 96% AUC, outperforming centralized deep learning and single-feature baselines across diverse environments.
I. INTRODUCTION
The paper addresses distribution shifts, pilot contamination, and dynamic NTN channels with a causal, adaptive, distributed authentication framework. It combines multi-feature fingerprints, meta-learning, and conditional distributed verification for robust authentication with reduced overhead.
- Motivation: Existing NTN PLA methods rely on single features, while time-varying channels, hardware impairments, pilot contamination, and distribution shifts degrade authentication robustness.Deep-learning fingerprints can also be vulnerable to adversarial perturbations and heterogeneous receiver hardware.
- Framework: SAFA-MZ combines SCM, backdoor adjustment, IRM, and MAML to learn environment-invariant features and adapt rapidly to unseen NTN conditions.The framework also estimates conditional average treatment effects of key design variables.
- Authentication: The dual-mode DPLA framework performs local authentication first and activates TDOA-based distributed verification only when local recognition is unreliable.This architecture uses multiple aerial nodes connected to a fusion center and is intended to reduce overhead.
- Fingerprinting: The proposed fingerprint integrates directional, beam, combiner, subspace, Doppler-delay, SINR, and INR features under pilot contamination.The resulting feature vector is interference-aware and designed for dense multi-user NTN uplinks.
- Signal model: The system models uplink observations with time-varying channels, user interference, pilot reuse, and normalized linear MMSE-type combining.The received signal, channel, interference, and noise terms support multi-feature extraction across distributed aerial nodes.
B. Authentication
Local authentication compares each observed feature vector with enrolled user profiles using a Mahalanobis-distance-based soft legitimacy score.
- Local phase: Each node tests the observed feature vector against enrolled user identities and computes a local soft score from Mahalanobis distance.The score supports local recognition before any distributed verification stage.
1) Local phase:
The local phase estimates legitimate feature statistics during enrollment and uses Mahalanobis distance to quantify profile consistency.
- Enrollment: Enrollment estimates the mean vector and covariance matrix of each legitimate user’s feature distribution from collected observations.These statistics define the reference profile used for subsequent authentication.
- Distance measure: Mahalanobis distance normalizes feature comparisons using the spread and correlation structure of legitimate samples.This makes the comparison scale-invariant and accounts for inter-feature correlations.
- Decision score: A small Mahalanobis distance indicates consistency with the enrolled profile, whereas a large distance suggests an attacker or identity mismatch.The distance is mapped to a local soft score using a node-specific threshold.
2) Distributed Verification and Attack Detection:
The framework first tests local consistency against enrolled profiles, then escalates suspicious transmissions to distributed verification. The fusion center estimates location consistency and combines evidence from the most reliable nodes using a graph-based mechanism.
- Each node compares an observed feature vector with enrolled legitimate-user profiles before distributed verification.
- Observations exceeding the local acceptance threshold are flagged as suspicious and trigger adaptive distributed verification.
- The fusion center uses synchronized TOA measurements to form TDOA observations and estimate the transmitter position.
- Two location-consistency scores measure estimated-position stability and agreement with the claimed or best-matching identity.
- The fusion center selects the L most reliable nodes, constructs a delay-, Doppler-, and geometry-aware graph, and applies attention-based readout.
C. Environment and Design Variables for NTN
The paper separates NTN environmental context from controllable design variables and models how both affect extracted features and authentication outcomes. Its causal framework supports adjustment for confounding and analysis of environment-dependent treatment effects.
- NTN context includes received SNR, angular spread, Rician factor, node altitude, propagation geometry, mobility, pilot reuse, and ADC resolution.
- Environmental factors influence propagation, Doppler sensitivity, extracted physical-layer features, and authentication performance.
- Treatments represent controllable configurations such as ADC resolution, the top-L budget, and orbit or platform type.
- The SCM treats environment and treatment as causes of features and authentication outcomes, with both confounding the feature-to-outcome relationship.
- Identifying the feature effect requires conditioning on both environment and treatment, while the treatment effect uses environmental variables for backdoor adjustment.
- CATE analysis characterizes how a design’s effectiveness varies with environmental factors such as elevation angle and node velocity.
B. Meta-Learning with Causal Invariance for NTN
Meta-learning organizes each environment–node pair as a task and trains a shared authentication model across tasks. IRM and doubly robust causal-consistency regularization encourage representations and interventional responses that remain stable across environments.
- Each environment–node pair defines a meta-learning task with separate training and validation datasets.
- A shared model performs task-specific inner-loop adaptation and is updated using validation losses across tasks.
- The IRM penalty encourages the feature extractor to learn environment-invariant causal representations.
- A doubly robust consistency regularizer encourages stable interventional responses across environments and observing nodes.
- The final meta-training objective combines task learning with invariant-risk and causal-consistency regularization strengths controlled by λIRM and λDR.
C. Counterfactual Design Policy for NTN
The counterfactual policy evaluates controllable design configurations from logged data, while attack detection fuses evidence from a selected subset of reliable nodes. The graph captures pairwise delay, Doppler, and geometric consistency while limiting computation when many nodes participate.
- A policy π maps each observed environment to a controllable treatment and aims to maximize interventional authentication performance.
- The doubly robust off-policy estimator evaluates policy performance from logged data without live deployment, reducing operational risk for HAPS and satellite platforms.
- The fusion center scores local evidence and selects the L most reliable nodes for graph construction.
- The selected nodes form a complete graph whose edge features encode pairwise delay, Doppler, and geometric differences.
- Graph attention transforms node features through message passing and attention-based readout into a fused representation augmented with triangulation and identity-location consistency.
- The graph captures delay, Doppler, and geometric consistency in dynamic NTN environments while balancing scalability and expressiveness.
E. Verification Score and ROC Metrics
The ROC analysis evaluates authentication by varying the threshold for accepting transmissions, distinguishing legitimate users from attackers through TPR, FPR, and AUC.
- The ROC curve shows the tradeoff between correctly accepting legitimate transmissions and wrongly accepting attacker transmissions as threshold τFC changes.
- TPR measures correctly accepted legitimate transmissions, while FPR measures wrongly accepted attacker transmissions.
- TP, FN, FP, and TN count correct or incorrect acceptance and rejection outcomes for legitimate and attacker transmissions.
- AUC summarizes the area under the ROC curve, with the absolute value ensuring a nonnegative area.
F. Summary of the Algorithms
The framework combines causal estimation, causally regularized meta-learning, and staged authentication, while its dominant computational costs remain scalable under bounded implementation parameters.
- Summary of the Algorithms: The framework has four stages: data collection, causal estimation, meta-learning, and authentication.The authentication stage produces the final FC score and decision.
- Summary of the Algorithms: Causal estimates guide MAML training through invariance and doubly robust consistency regularizers for unseen-environment evaluation.These regularizers encourage reliance on stable causal relationships rather than environment-specific correlations.
- Summary of the Algorithms: Algorithm 1 estimates propensity scores, outcome models, doubly robust means, average treatment effects, and conditional average treatment effects.
- Summary of the Algorithms: Algorithm 2 repeatedly adapts parameters, evaluates validation losses, computes IRM and doubly robust regularizers, and updates the initialization.
- Computational Complexity: O(Ns|K|M log M) is the local-processing complexity under LoS-dominant conditions with bounded subspace rank.In rich scattering, the corresponding expression becomes O(Ns|K|M 2).
- Computational Complexity: O(niterNs(CML + |T |)) is the meta-learning and causal-estimation training complexity, approximately linear in niterNs when dimensions and treatment count are bounded.
- Computational Complexity: O(|Ku|M log M) is the dominant online complexity for homogeneous node dimensions and bounded subspace rank.Conditional distributed verification and TDOA localization add lower-order terms when their graph sizes and model dimensions remain small.
- Computational Complexity: The proposed framework adds offline causal-estimation and meta-learning computation, while online cost remains dominated by scalable local processing.The comparison is against a local-only deep-learning baseline without causal adaptation, graph fusion, or TDOA attack detection.
V. SIMULATION RESULTS
The simulations generate multi-environment HAPS and satellite tasks, train causal meta-learning in a few-shot setting, and evaluate local and distributed verification under specified NTN conditions.
- Data Generation: The dataset covers HAPS and satellite scenarios, using five training scenarios and one testing scenario for user-ID recognition.Each scenario contains ten users, each performing ten movements.
- Data Generation: The uplink uses Rician fading at fc = 2 GHz, with environmental variables sampled from the ranges in Table II.
- Data Generation: The experiments use 100 enrollment snapshots per legitimate user, with a two-dimensional reference subspace defined by the two dominant eigenvectors.
- Meta-Learning Setup: The causal meta-learning framework trains across |E| = 5 environments using 5 support and 5 query samples per user in a few-shot setting.Training uses 160 outer epochs and 20 episodes per epoch.
- Distributed Verification: Distributed verification uses |K| = 4 aerial nodes, each with Mk = 16 antennas at 20 km altitude and velocity 15 m/s.
- Evaluation Setup: The FC classifier is a two-layer MLP with hidden units (64, 32), sigmoid output, threshold τFC = 0.5, and 10 Monte Carlo runs.The SNR sweep spans −10 to 30 dB.
4) Test Scenario:
The test evaluates SAFA-MZ on unseen environment combinations using few-shot adaptation and compares fusion strategies and feature importance across NTN conditions. SAFA-MZ achieves 92% accuracy and 96% AUC while feature relevance varies with propagation, platform, interference, system conditions, and attacker capability.
- Evaluation setup: 10 support samples per user adapt the meta-learner, followed by testing on 100 query samples; reported metrics are AUC and authentication accuracy.
- Fusion comparison: The proposed method achieves the highest AUC by combining all eight features, causal relationships, and meta-learning for unseen NTN conditions.
- Feature importance: Feature importance depends on scenario: spatial and geometry features are broadly reliable, while interference-aware features matter more in dense or pilot-contaminated access.
- Platform effects: LEO links emphasize delay–Doppler features, GEO links emphasize spatial features, and HAPS links emphasize angular and subspace features.
- Attacker effects: Far attackers make delay–Doppler features effective, whereas nearby or capable attackers shift reliance toward spatial, combiner, subspace, and SINR features.
- Ablation analysis: Feature-removal analysis finds no universally dominant feature, and some spatial or interference features are redundant under particular conditions.
- Overall performance: 92% accuracy and 96% AUC are reported for SAFA-MZ, outperforming baseline and single-feature methods.
APPENDIX A PROOFS
The appendix establishes identification and consistency results for causal estimators under standard assumptions. It also shows double robustness and characterizes conditions under which the invariant-risk objective vanishes.
- Identification: Under consistency, conditional exchangeability, and positivity, the interventional mean is identified by adjustment over Z_k,u.
- Estimating function: The population estimating function has expectation equal to the interventional mean, establishing its target parameter.
- Consistency: With positivity, bounded moments, and standard regularity conditions, the sample estimator is consistent.
- Double robustness: The estimator is doubly robust, remaining consistent if either the propensity model or outcome model is correctly specified.
- Mediation: The appendix decomposes a treatment effect into natural indirect and direct effects.
- Invariant risk: If a representation makes w = 1 stationary for every environment, node, and user risk, the invariant-risk objective equals zero.