Source-linked AI summary
A Survey on Machine Learning-based Misbehavior Detection Systems for 5G and Beyond Vehicular Networks
Abdelwahab Boualouache, Thomas Engel
TL;DR
5G-V2X security remains challenged by authenticated internal attackers and threats that PKI alone cannot address. The paper surveys and classifies ML-based MDSs from security and ML perspectives, then derives recommendations and open issues. Its analysis finds limited support for unseen-attack detection, predominantly traditional and non-updated models, and unresolved dataset, security, and deployment challenges.
Problem
Authenticated internal attackers and strategic threats remain beyond what PKI alone can address, while ML-based MDSs lack comprehensive analysis.
Method
The paper surveys and classifies ML-based MDSs, analyzes them from security and ML perspectives, and derives recommendations and open research issues.
Results
Most systems target specific attacks rather than anomalies or unseen attacks, use traditional ML, and are usually not updated after deployment.
Takeaways & Limitations
Effective ML-based MDS development requires attention to attack coverage, model updating, collaborative learning, standardized datasets, security, and deployment constraints.
Takeaways & Limitations
Existing results are difficult to reproduce and compare because benchmark datasets and standardized V2X attack procedures are absent.
Abstract
from arXiv · showhide
Significant progress has been made towards deploying Vehicle-to-Everything (V2X) technology. Integrating V2X with 5G has enabled ultra-low latency and high-reliability V2X communications. However, while communication performance has enhanced, security and privacy issues have increased. Attacks have become more aggressive, and attackers have become more strategic. Public Key Infrastructure proposed by standardization bodies cannot solely defend against these attacks. Thus, in complementary of that, sophisticated systems should be designed to detect such attacks and attackers. Machine Learning (ML) has recently emerged as a key enabler to secure our future roads. Many V2X Misbehavior Detection Systems (MDSs) have adopted this paradigm. Yet, analyzing these systems is a research gap, and developing effective ML-based MDSs is still an open issue. To this end, this paper present a comprehensive survey and classification of ML-based MDSs. We analyze and discuss them from both security and ML perspectives. Then, we give some learned lessons and recommendations helping in developing, validating, and deploying ML-based MDSs. Finally, we highlight open research and standardization issues with some future directions.
1 INTRODUCTION
5G-V2X improves communication capabilities but remains exposed to strategic internal attacks that cryptographic protections cannot fully address. This survey examines ML-based misbehavior detection systems, organizing prior work and identifying guidance and open issues for their development and deployment.
- Motivation: Authenticated internal attackers can still launch denial-of-service, position, message-dropping, and falsification attacks, motivating MDSs as a complement to PKI.These attacks remain dangerous because attackers are already authenticated members and can resist cryptographic defenses.
- Motivation: ML has become a key enabler for network security, prompting multiple ML-based MDSs for detecting attacks in 5G-V2X.The paper connects ML’s prior success in network security with its expected role in 5G and beyond networks.
- Contributions: The survey fills a research gap by analyzing ML-based MDSs from both security and machine-learning perspectives.It establishes analysis guidelines and discusses existing systems across security and ML characteristics.
- Contributions: The paper surveys and classifies ML-based MDSs, analyzes existing solutions, presents development and deployment recommendations, and highlights open research and standardization issues.These contributions cover taxonomy, analysis, lessons learned, recommendations, and future research priorities.
- Organization: The paper’s roadmap covers 5G-V2X background, an MDS taxonomy, security and ML-oriented analysis, recommendations, open research issues, and conclusions.The background includes architecture, security requirements, attacker models, attacks, and ML-based MDS development and evaluation.
3. ML-based Misbehavior Detection Systems: a taxonomy
This section frames ML-based misbehavior detection within 5G-V2X architecture, security requirements, attacker models, and attack classification.
- Security and attack foundations: The section organizes background into security requirements, attacker models, and attack classification before addressing misbehavior detection.The stated structure covers security services, adversary types, and attack categories.
- Security requirements: 5G-V2X security requires authentication, integrity, availability, confidentiality, non-repudiation, access control, and privacy.Privacy includes identifier, location, and exchanged-data protection.
- Attacker models: V2X adversaries are characterized by coverage, activity, system membership, and motivation, including global or local and internal or external attackers.Active attackers alter or inject messages, whereas passive attackers eavesdrop; malicious and rational attackers differ in objectives.
- Attack classification: The attack taxonomy is linked to targeted applications and whether internal or external attackers can launch each attack.Figure 5 provides an overview, while Table 2 specifies application targets and attacker type.
1) Attacks on Authenticity
Authenticity attacks exploit identities and credentials to inject information, obtain access, or conceal malicious nodes in V2X systems.
- Attacks on Authenticity: Sybil attacks exploit multiple identifiers to inject false information or create an illusion of traffic congestion.They particularly concern non-IP-based safety applications that use multiple identifiers for location privacy.
- Attacks on Authenticity: Impersonation or masquerading uses a valid identity to obtain V2X access, steal private information, and launch further attacks.The described process can include probing, port scanning, and application-layer exploits.
- Attacks on Authenticity: Certificate replication hides malicious V2X-node identities by using replicated certificates, which are removed after blacklisting.The attack exploits certificate reuse despite the revocation process.
2) Attacks on Integrity
Integrity attacks manipulate V2X information or communication paths, ranging from false messages and replay to GPS deception and tunneling.
- Attacks on Integrity: False-message injection or alteration sends incorrect position or speed information that may place honest vehicles in dangerous situations.The passage indicates this attack is more likely in non-IP-based applications.
- Attacks on Integrity: Replay attacks retransmit captured messages at different times while presenting them as messages from the original senders.The attack abuses message timing and apparent provenance.
- Attacks on Integrity: GPS spoofing deceives receivers by retransmitting captured signals or transmitting inaccurate GPS signals.The attacker can use real signals captured elsewhere at a different time.
- Attacks on Integrity: Tunneling connects at least two malicious V2X nodes so false data can be injected from another location.The paper treats tunneling as a special case of false-message injection.
3) Attacks on Availability
Availability and privacy attacks disrupt service delivery, forwarding, radio access, or information confidentiality across V2X networks.
- Attacks on Availability: DoS and DDoS attacks prevent normal access to network services and can target both IP-based and non-IP-based V2X applications.Non-IP attacks may increase periodic-message frequency, while IP attacks include UDP and ARP flooding.
- Attacks on Availability: Blackhole and Greyhole attacks stop message dissemination, with Blackhole dropping all messages and Greyhole dropping selected messages.They affect both IP-based and position-based routing protocols.
- Attacks on Availability: Wormhole attacks establish tunnels between malicious nodes to conduct denial-of-service attacks against IP-based routing protocols.The attack resembles tunneling but is described here as disrupting routing.
- Attacks on Availability: Timing attacks intentionally delay message forwarding, posing particular danger to time-sensitive safety applications.The attack targets dissemination and routing protocols.
- Attacks on Availability: Jamming corrupts data or radio access, while eavesdropping, location tracking, and man-in-the-middle attacks compromise information confidentiality or communication privacy.Both ETSI ITS-G5 and C-V2X are described as vulnerable to jamming.
5) Attacks on Non-repudiation
This section situates misbehavior detection within 5G-V2X security, covering faulty and malicious behavior, attack classification, datasets, simulators, and evaluation resources.
- Misbehavior detection systems: ML-based MDSs use both node-centric statistics and data-centric message content to detect faulty or malicious V2X behavior.Faulty nodes generate incorrect data without malicious intent, whereas malicious nodes transmit erroneous messages intentionally.
- Development and evaluation: Development and evaluation of ML-based MDSs involve public datasets, network simulators, and evaluation metrics.
- Datasets: VeReMi and its extension provide simulated misbehavior data spanning position falsification and additional attacks under varied traffic or radio conditions.VeReMi includes five position-falsification types, while the extension adds speed malfunctions, delayed messages, DoS variants, replay, and disruptive attacks.
- Datasets: Security datasets differ in origin and coverage, including honeypot, emulated, testbed, simulated, and trajectory data.CRAWDAD and NGSIM do not originally contain attacks; authors preprocess them by injecting noise, and no listed dataset includes non-repudiation attacks.
- Network simulators: Common simulation tools include OMNeT++, SUMO, NS2/NS3, Veins, and CTUns-5.0 for constructing network or mobility environments.Veins combines OMNeT++ and SUMO to support vehicular-network simulation, while SUMO models city-scale mobility.
1) Traditional learning
Traditional learning covers supervised and unsupervised machine-learning approaches used to classify, organize, reduce, or detect anomalies in V2X data.
- Traditional learning: Traditional learning denotes machine-learning algorithms that are not based on deep learning.
- Traditional learning: Supervised learning trains on labeled data for classification or regression, including classification of safety messages as malicious or normal.Representative classifiers include Naive Bayes, logistic regression, SVM, KNN, random forest, neural networks, boosting, and ensemble methods.
- Traditional learning: Unsupervised learning uses unlabeled data to identify structure through anomaly detection, clustering, or dimensionality reduction.Anomaly detection identifies unexpected items without prior knowledge, while clustering organizes data into groups.
2) Deep learning
Deep learning uses multilayer neural networks for supervised and unsupervised tasks, with CNNs, RNNs, and autoencoders representing key approaches.
- Deep learning: Deep learning is a neural-network-based subset of ML in which deeper architectures can perform well on large datasets with many features and rows.The paper describes applications including computer vision, natural language processing, and machine translation.
- Supervised learning: CNNs learn features from matrix-like inputs through convolution and pooling, whereas RNNs address sequential or time-series data.
- Unsupervised learning: Autoencoders compress inputs into an encoding and reconstruct them, using reconstruction loss to detect anomalies.The architecture contains an encoder, a bottleneck representation, and a decoder.
3) Advanced ML concepts
Advanced ML concepts extend learning across distributed data, sequential decision-making, related tasks, limited labels, and synthetic-data generation, alongside standard MDS metrics.
- Advanced ML concepts: Federated learning builds a global model collaboratively across nodes without sharing their datasets.Nodes compute local model updates over multiple rounds coordinated by an FL server.
- Advanced ML concepts: Reinforcement learning trains agents through rewards and penalties while they learn actions that maximize outcomes over time.
- Advanced ML concepts: Transfer learning reuses knowledge from one related problem, such as DoS detection, for another such as DDoS detection.
- Advanced ML concepts: Semi-supervised learning combines a small labeled dataset with a large unlabeled dataset when labels are difficult to obtain.
- Advanced ML concepts: GANs generate data resembling training data through jointly trained generator and discriminator networks.
- Evaluation metrics: Evaluation metrics include accuracy, precision, recall, F1-score, TPR, TNR, FPR, FNR, ROC curves, and AUC.These metrics quantify detection correctness, attacker identification, honest-vehicle rejection, trade-offs, or class separability.
3 ML-BASED MISBEHAVIOR DETECTION SYS-
The survey classifies ML-based misbehavior detection systems by V2X application type and reviews systems targeting diverse attacks across protocol layers. It summarizes their datasets, learning approaches, detected behaviors, and evaluation practices.
- ML-based MDSs are classified into non-IP safety applications, IP non-safety applications, and systems applicable to both, corresponding mainly to facilities, transport/networking, and physical-layer attacks.
- Non-IP-based applications: The reviewed non-IP systems address position falsification, false information, Sybil, position tracking, and multi-attacks using supervised, unsupervised, reinforcement, and federated learning approaches.
- Non-IP-based applications: Table 4 distinguishes false information involving vehicle dimensions and types, positions, alerts, or road traffic, while three systems do not explicitly specify the detected information type.
- Multi-attacks: The reviewed multi-attack systems cover combinations including position falsification, false information, DoS/DDoS, Sybil, replay, timing, greyhole/blackhole, and impersonation.
- IP-based applications: IP-based systems include unsupervised traffic-flow clustering for DoS detection, while other reviewed systems target greyhole/blackhole, Sybil, wormhole, and multiple attacks.
4 SUMMARY & DISCUSSION
The survey finds that existing ML-based MDSs predominantly detect known attacks with supervised, single-node models trained on datasets. It also identifies limited support for unseen attacks, collaborative privacy-preserving learning, context awareness, model updating, and deployment analysis.
- Security and Privacy oriented summary: Most proposed MDSs detect specific attacks rather than anomalies, and their ability to detect unseen attacks remains uncertain because evaluations use well-known attacks.
- Security and Privacy oriented summary: Only a few systems use collaborative learning, most of them federated learning, while context awareness appears in only two works and privacy preservation is associated with collaborative systems.
- ML-Oriented summary: All existing works are validated with datasets, with only one combining datasets and simulations; most studies also omit inference location and post-deployment model updating.
- ML-Oriented summary: Most systems use supervised learning for classification, whereas unsupervised tasks are mostly anomaly detection; only one reviewed system combines supervised and unsupervised learning.
- ML-Oriented summary: Non-IP MDS datasets are generally generated with network simulators, while most IP-based MDSs use public datasets produced from computer-network testbeds, especially VeReMi-related datasets for non-IP systems.
- ML-Oriented summary: Most studies use traditional ML algorithms, although recent work increasingly adopts deep learning, and evaluation metrics vary across systems.
5 LESSONS LEARNED AND RECOMMENDATIONS
The survey recommends clearer claims, broader evaluation practices, and explicit attention to deployment, updating, privacy, and security. It identifies dataset dependence, weak comparability, and vulnerability to adversarial attacks and poisoning as major concerns.
- ML-based MDS claims should be narrowed to the attacks represented in training datasets, and holistic frameworks should combine systems to cover existing and unseen attacks.
- The absence of benchmark datasets and unified metrics makes systems difficult to reproduce and compare, especially when authors generate their own datasets.
- Deployment should be evaluated through inference location, model size, and inference time because deployment affects detection rate and feasibility as well as security and privacy.
- Model-update mechanisms should be defined to prevent accuracy loss over time, with collaborative MDSs offering privacy-preserving update opportunities and motivating online or reinforcement learning.
- ML-based MDSs, including collaborative systems, remain exposed to adversarial and poisoning attacks, while only one reviewed work considers MDS security.
6 OPEN RESEARCH ISSUES
The survey identifies unresolved challenges spanning datasets, zero-day detection, context awareness, distributed security, deployment, incentives, and standardization for ML-based MDSs.
- 6.1 Datasets: benchmarking and reproducibility: Benchmark datasets remain difficult to establish because attack scenarios lack standardization, simulators omit realistic parameters, and computer-network testbeds omit vehicle mobility.The survey also calls for common evaluation metrics, clearer ML parameter reporting, and public source code to improve reproducibility.
- 6.2 Zero-day attacks: Zero-day detection remains open because supervised systems target known attacks, while hybrid human-in-the-loop frameworks are only a promising step toward automatic zero-touch detection.The paper identifies automatic zero-day detection under the zero-touch paradigm as the ultimate goal.
- 6.3 Context-awareness: Existing MDSs mostly use direct parameters, overlooking indirect context such as channel status, temperature, and speed that can affect detection accuracy.The survey notes that limited context-aware systems do not include these indirect parameters in ML training.
- 6.4 Security and Privacy: Centralized MDSs create single-point-of-failure and privacy risks, while federated learning reduces central data storage but retains a global-server failure point and introduces additional security issues.The survey proposes blockchain as a possible way to secure ML-based MDSs, while noting unresolved design problems.
- 6.5 Deployment and Incentives: ML-based MDS deployment must account for V2X hardware, software, model size, processing resources, component placement, early detection, rapid reaction, and incentives for node managers.The paper treats deployment as an open issue requiring both research and industry effort, with game-theoretic incentive models already emerging.
- 6.6 Standardization: Standardization opportunities include attack scenarios, benchmark datasets, validation KPIs, evaluation metrics, validation procedures, and stakeholder plugtests for ML-based MDSs.The paper observes that ongoing MDS standardization has not yet emphasized ML sufficiently.
7 CONCLUSION
The paper surveys and classifies ML-based MDSs for 5G-V2X, analyzes them from security and ML perspectives, and identifies guidance and open issues for effective systems.
- 7 CONCLUSION: The survey classifies ML-based MDSs for 5G-V2X, analyzes their security and ML perspectives, and presents lessons, recommendations, and open research and standardization issues.These contributions target the development of effective ML-based MDSs.