@Article{iasc.2026.088003,
AUTHOR = {Anil Carie, Sanjeev Kumar Makala, Satish Anamalamudi, Shaik Teena, Awadhesh Dixit, Pandu Sowkuntla, Bhaskar Marapelli},
TITLE = {Hierarchical Trust Management for C-V2X Networks: A Game-Theoretic Edge-Cloud Architecture with LLM-Driven Calibration},
JOURNAL = {Intelligent Automation \& Soft Computing},
VOLUME = {41},
YEAR = {2026},
NUMBER = {1},
PAGES = {139--166},
URL = {http://www.techscience.com/iasc/v41n1/68854},
ISSN = {2326-005X},
ABSTRACT = {Connected Vehicle-to-Everything (C-V2X) networks rely on Basic Safety Messages (BSMs) for cooperative awareness, but authenticated vehicles can still transmit falsified kinematic data—an insider attack that cryptographic authentication cannot prevent. Existing misbehavior detection systems (MDSs) achieve high detection rates on static benchmarks yet provide no formal guarantee that honest behaviour is a rational vehicle’s dominant strategy. We present a four-layer hierarchical trust architecture for 5G New Radio (NR) C-V2X that integrates edge AI detection, cloud large language model (LLM)-driven weight calibration, game-theoretic conviction, and ledger-anchored payoff tracking. A Nash equilibrium gate, calibrated with a wide margin above the empirically observed honest-vehicle suspicion ceiling, achieves zero observed false positives across 1584 honest phase-3 instances (one-sided 95% Clopper–Pearson upper bound: 0.19% per instance), without requiring any assumption about the honest-score distribution’s shape. A clawback mechanism makes the expected payoff of attacking strictly negative for all tested temptation levels. Evaluation on NS3 with 3GPP Rel-18 channel models across nine attack types—including rational adversaries with heterogeneous temptation <math id="mml-ieqn-1"><msub><mi>T</mi><mi>i</mi></msub><mo>∼</mo><mi>U</mi><mo stretchy="false">(</mo><mn>0.5</mn><mo>,</mo><mn>2.5</mn><mo stretchy="false">)</mo></math>—yields a detection rate (DR) of <math id="mml-ieqn-2"><mn>86.3</mn><mi mathvariant="normal">%</mi></math> with false-positive rate (FPR) of <math id="mml-ieqn-3"><mn>0</mn><mi mathvariant="normal">%</mi></math> observed across 1584 honest phase-3 instances. This zero-FPR result holds for the evaluated single-UAV configuration at <math id="mml-ieqn-4"><mi>N</mi><mo>≤</mo><mn>50</mn></math> vehicles per zone; at <math id="mml-ieqn-5"><mi>N</mi><mo>=</mo><mn>100</mn></math>, mean FPR rises to <math id="mml-ieqn-6"><mn>4.9</mn><mi mathvariant="normal">%</mi></math> (range <math id="mml-ieqn-7"><mn>0</mn></math>%–<math id="mml-ieqn-8"><mn>13.5</mn><mi mathvariant="normal">%</mi></math> across seeds) due to coverage-boundary effects, and multi-UAV sectorisation to extend this range remains unvalidated futurework. All 95 rational attackers (<math id="mml-ieqn-9"><mi>N</mi><mo>=</mo><mn>30</mn></math>, seeds 1, 2, 3, 5, 6) chose cooperation, producing a honesty premium of <math id="mml-ieqn-10"><mi mathvariant="normal">Π</mi><mo>=</mo><mo>+</mo><mn>39.2</mn></math> Trust Credits (positive in every seed, range <math id="mml-ieqn-11"><mo>+</mo><mn>16.4</mn></math> to <math id="mml-ieqn-12"><mo>+</mo><mn>55.4</mn></math> TC). A five-arm ablation isolates the Nash gate’s and clawback’s independent contributions to this outcome. Investigating cloud LLM calibration’s contribution to detection rate, we identified and corrected a learning-rate implementation fault that had suppressed nearly all of the LLM’s recommended adjustments throughout the original evaluation; under the corrected configuration, calibration produces a statistically significant detection-rate improvement, confirmed with a live cloud-unreachable control (<math id="mml-ieqn-13"><mo>+</mo><mn>5.3</mn></math> percentage points, <math id="mml-ieqn-14"><mi>p</mi><mo>=</mo><mn>0.034</mn></math>), reported alongside the original finding for transparency. To our knowledge, this is the first V2X misbehavior detection system to provide formal incentive compatibility under heterogeneous temptation for economically-rational attackers under the specified payoff model; this claim does not extend to attackers pursuing non-economic objectives, collusion, Sybil behaviour, or calibration-pipeline poisoning, which remain open problems.},
DOI = {10.32604/iasc.2026.088003}
}