# Scalable Hierarchical AI-Blockchain Framework for Real-Time Anomaly Detection in Large-Scale Autonomous Vehicle Networks

Crypto.world

> Autonomous vehicles must spot attacks on their sensors and communications in under 10 milliseconds to stay safe, but sharing what they learn with the rest of the fleet requires consensus protocols…

**arXiv**: 2511.12648

**Published**: 2025-11-16

**Authors**: Rathin Chandra Shit, Sharmila Subudhi

## AI summary

## Overview

**Research area:** Cybersecurity for autonomous vehicle (AV) networks, combining edge-based machine learning anomaly detection, Byzantine-robust federated learning, and permissioned blockchain coordination.

**Technical level:** Intermediate to Advanced. The paper assumes familiarity with federated learning aggregation, Byzantine fault tolerance, PBFT consensus, and differential privacy, though the three-tier design itself is explained in accessible terms.

**Scope:** The paper proposes and evaluates HAVEN (Hierarchical Autonomous Vehicle Enhanced Network), a three-tier architecture that separates millisecond-scale local threat detection from slower regional and global security coordination, tested on the nuScenes dataset in simulations with 100 to 1000 vehicles.

**Publication details:** arXiv:2511.12648v1 [cs.CR], 16 November 2025, by Rathin Chandra Shit (International Institute of Information Technology, Bhubaneswar) and Sharmila Subudhi (Maharaja Sriram Chandra Bhanja Deo University, Baripada).

## What This Paper Is About

Autonomous vehicles must spot attacks on their sensors and communications in under 10 milliseconds to stay safe, but sharing what they learn with the rest of the fleet requires consensus protocols that take seconds. Existing systems force a choice: fast but isolated local detection, or slow but network-wide intelligence. The paper's goal is to build one architecture that delivers both by splitting security work across three processing tiers, each matched to the latency its task can tolerate.

## Key Contributions

1. **A three-tier hierarchical architecture with explicit latency separation.** Tier 1 handles local anomaly detection at sub-10 ms latency on automotive-grade edge hardware, Tier 2 handles regional collaborative learning through federated learning rounds, and Tier 3 handles global, tamper-resistant threat coordination through blockchain consensus bounded at 3 to 5 seconds.
2. **A lightweight multimodal ensemble detector with uncertainty quantification.** Tier 1 fuses LiDAR, camera, GPS, and IMU features over 50-timestep windows (500 ms of data at 100 Hz) using Random Forest, SVM, and LSTM classifiers, combined through adaptive softmax weighting based on each model's historical validation accuracy, with a dual-threshold decision rule (anomaly score above 0.7 and confidence above 0.8).
3. **A Byzantine-robust, privacy-preserving federated learning protocol.** Regional model updates are aggregated with a trimmed-mean rule designed to tolerate up to 30% compromised participants (parameterized as beta = 0.3), and threat signatures are shared under Laplace-mechanism differential privacy with a budget of epsilon = 1.0.
4. **Selective blockchain logging to cut consensus overhead.** Only events meeting at least one criterion are written on-chain: severity above 0.85, cross-regional frequency above 3, or consensus confidence above 0.9. The authors state this reduces transaction volume by approximately 95% relative to comprehensive logging, with PBFT used for consensus.

## Main Findings

- **Detection performance:** Reported accuracy of 94% and F1-score of 92% across multimodal sensor data, with detection latency below 10 ms.
- **Byzantine resilience:** Fault tolerance was validated experimentally with 20% of nodes compromised, against a design target of tolerating up to 30%.
- **Scalability:** Simulations spanned 100 to 1000 vehicles and covered sensor spoofing, jamming, and adversarial model poisoning attacks.
- **Blockchain overhead:** Selective event logging reduced storage overhead and transaction volume substantially (approximately 95% fewer transactions than comprehensive logging, as stated for the logging rule).
- **Privacy:** Differential privacy was maintained at epsilon = 1.0 while preserving threat-pattern recognition, per the authors.
- **Comparison to prior work:** The paper positions these results against earlier systems that reported response times above 150 ms (Duan et al.), 200 to 300 ms (Khanmohammadi and Azmi), consensus above 5 seconds (Cebe et al., Block4Forensic), and coordination latencies above 500 ms (Al-Hawawreh and Hossain).
- **Not reported in the supplied text:** The content provided ends within Section 4.4.2, before the experimental section. Detailed per-attack results, throughput measurements, ablation studies, baseline comparison tables, dataset split sizes, and per-tier latency breakdowns are therefore not available here and should be read from the full paper.

## Methodology in Plain English

The researchers start from an optimization framing: minimize a weighted sum of detection error, latency cost, and privacy loss, subject to three hard constraints, namely a response time of at most 10 ms per vehicle, detection accuracy of at least 0.94, and a compromised-node fraction no greater than 0.3. Because no single mechanism satisfies all three at once, they divide the work.

Tier 1 runs on the vehicle. Raw sensor streams are converted into compact statistical features (for example point density and height variance for LiDAR, brightness and contrast for cameras, position and orientation for GPS/IMU). Three models score each sliding window, their votes are weighted by past reliability, and a confidence estimate derived from disagreement among the models gates the final decision. If both thresholds are crossed, the vehicle executes an immediate safety response on its own and sends a hashed threat signature to its regional coordinator, so safety never depends on connectivity.

Tier 2 runs regionally, over clusters of 100 to 500 vehicles with coordination rounds every 30 seconds. Each vehicle trains locally, sends only the difference between its updated weights and the global weights, and the coordinator applies a trimmed mean that discards the most extreme updates before averaging. Laplace noise is added to the aggregate to protect individual contributions. This is what provides Byzantine robustness without exposing raw sensor data.

Tier 3 runs globally on a consortium blockchain. Because writing every event on-chain would be too slow, only high-severity, frequently recurring, or high-consensus-confidence events are logged, and smart contracts automate responses. PBFT provides the consensus, with a stated 3 to 5 second delay that is acceptable for audit and intelligence sharing but would be unacceptable for braking decisions.

Evaluation uses the real-world nuScenes autonomous driving dataset and large-scale simulations varying fleet size and attack type. The simulation setup, hardware details, and full results tables fall in sections of the paper not included in the supplied text.

## Why This Matters

**Impact on research:** The paper reframes the AV security problem as a workload-placement question rather than a single-algorithm question, arguing that real-time detection, collaborative learning, and auditable logging should not compete for the same processing path. It also directly targets a gap the authors identify in prior work: blockchain-based automotive proposals have mostly been demonstrated on proof-of-concept deployments of fewer than 100 vehicles, while this work reports simulations up to 1000.

**Real-world applications:**

- Fleet-level security operations centers for robotaxi or logistics fleets, where per-vehicle detection must roll up into shared threat intelligence without exposing customer data.
- Tamper-evident incident logging for forensic investigation and regulatory audit of autonomous driving systems.
- Cross-manufacturer or cross-region threat intelligence sharing, where selective on-chain logging provides an auditable but low-volume record.
- Urban-scale V2X deployments where jamming, spoofing, and model poisoning campaigns need to be correlated across many vehicles in near real time.

**Industry relevance:** The authors explicitly flag that academic proposals often ignore automotive-grade hardware constraints, real-world network conditions, and standards such as ISO 26262 and ISO 21434. HAVEN's tiering and its stated intra-vehicle, sub-200 ms vehicle-to-coordinator, and 3 to 5 second global latency budgets map onto how automotive security architectures are actually specified, which makes the design easier for practitioners to reason about.

## Future Directions

- **Hardware and on-road validation.** The evaluation is dataset-driven and simulated. Testing on actual automotive-grade ECUs, under thermal and power limits, and in live vehicle networks is the obvious next step the design implies.
- **Pushing the Byzantine bound.** The architecture is designed for up to 30% compromised participants but was validated at 20%. Behavior between those points, and under coordinated rather than random compromise, remains an open question.
- **Tuning the trade-off weights and thresholds.** The latency and privacy weights in the objective, the dual decision thresholds (0.7 and 0.8), the trim fraction (0.3), and the privacy budget (1.0) are fixed by empirical validation; adaptive or learned settings are not explored in the provided text.
- **Governance and standards alignment.** The paper identifies regulatory compliance as a motivating requirement but leaves open how consortium membership, validator trust, and audit obligations would be governed across manufacturers and jurisdictions.

## Target Audience

Researchers and graduate students working on autonomous vehicle security, edge AI, federated learning, or blockchain for cyber-physical systems will get the most from this paper, particularly those interested in how to partition security workloads across latency tiers. Automotive security architects, fleet operators, and standards practitioners evaluating whether blockchain or federated learning can fit within real-time safety budgets will also find the tiering argument and the selective-logging rule directly useful. Readers seeking detailed benchmark tables should consult the full paper, since the experimental sections are outside the excerpt summarized here.

## Authors’ abstract

The security of autonomous vehicle networks is facing major challenges, owing to the complexity of sensor integration, real-time performance demands, and distributed communication protocols that expose vast attack surfaces around both individual and network-wide safety. Existing security schemes are unable to provide sub-10 ms (milliseconds) anomaly detection and distributed coordination of large-scale networks of vehicles within an acceptable safety/privacy framework. This…

[The full abstract on arXiv](https://arxiv.org/abs/2511.12648)

[Read the original paper](https://arxiv.org/abs/2511.12648)

---

Canonical: https://crypto.world/paper/scalable-hierarchical-ai-blockchain-framework-for-real-time-anomaly-2511.12648
