💻 computer science

Enhancing V2X Communication Security Using an MTCNN-LSTM-Based Multilevel Trust Evaluation Framework

This paper proposes a closed-loop V2X security and energy allocation framework that integrates an MTCNN-LSTM-based multilevel trust evaluation model to achieve high-accuracy multi-attack detection and optimize battery consumption in smart grid-enabled electric vehicle networks.

Original authors: Sanjay Kumar Sonker, Bharat Bhushan Sagar, Vibha Kaw Raina, Md Khalid Jamal

Published 2026-09-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sanjay Kumar Sonker, Bharat Bhushan Sagar, Vibha Kaw Raina, Md Khalid Jamal

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a future where electric vehicles do not merely drive on roads but talk to everything around them: to other cars, to traffic lights, to charging stations, and to the power grid itself. This network, known as vehicle-to-everything communication, promises a smarter, more efficient transportation system. However, this constant chatter creates a new vulnerability. Just as a crowded room can be overwhelmed by a single person shouting lies or flooding the air with noise, these digital networks can be hijacked by attackers who pretend to be many cars at once or who simply clog the system with fake requests. If the power grid cannot tell the difference between a real car needing a charge and a malicious program draining resources, the entire system could fail, leaving real drivers stranded and wasting precious energy.

A team of researchers has developed a new way to protect these networks by teaching computers to judge the character of every vehicle in real time. Instead of relying on static rules that hackers can easily bypass, they created a system that watches how vehicles behave over time, looking for subtle signs of trouble. The researchers built a digital model that combines two powerful types of artificial intelligence. One part acts like a sharp eye, scanning for immediate patterns in the data, while the other acts like a memory, remembering how a vehicle has acted in the past to spot changes in its behavior. By feeding this system with real-world data about how electric cars consume energy, move, and communicate, the researchers trained it to recognize when a vehicle is acting normally and when it is part of a cyberattack.

The core of their discovery lies in a concept called "trust," which is calculated not just from a single message, but from a vehicle's entire history. The system looks at nine different clues, such as how much energy a car is using, how fast it is traveling, how often it sends messages, and even the temperature of its battery. If a vehicle suddenly starts sending too many messages, asks for an unrealistic amount of power, or shows signs of a battery temperature that doesn't match its driving, the system flags it as suspicious. This approach allows the network to distinguish between a genuine emergency and a coordinated attack where a single bad actor pretends to be dozens of cars to steal electricity or crash the grid.

In their tests, the researchers simulated a busy network with hundreds of virtual electric vehicles and introduced various types of attacks, including those that flood the system with data and those that create fake identities. The new model proved remarkably effective, correctly identifying malicious behavior in nearly all cases. It achieved a detection rate of over 99 percent, meaning it almost never missed an attack, while also keeping false alarms to a minimum. Perhaps most importantly, the system did not just detect the problem; it solved the resource crisis that often follows. When the model identified a group of fake vehicles trying to drain the charging stations, it automatically cut off their access. This prevented the waste of energy that usually happens during such attacks. In simulations where half the network was under attack, the system managed to keep energy waste to a manageable level, ensuring that the real, trustworthy vehicles could still get the power they needed.

The researchers also found that this method helps protect the physical health of the batteries. By denying charging requests to suspicious vehicles, the system prevents the kind of overcharging that can damage a battery over time. In their simulations, the model successfully balanced the need for energy with the need for security, ensuring that the grid remained stable even when under heavy pressure. The study suggests that by using this layered approach of watching behavior, calculating trust, and making instant decisions, we can build a transportation network that is not only smart but also resilient. While the results come from computer simulations using real-world driving data, they offer a clear path forward for securing the electric future, ensuring that when our cars talk to the grid, they are speaking the truth.

Technical Summary: Enhancing V2X Communication Security Using an MTCNN-LSTM-Based Multilevel Trust Evaluation Framework

1. Problem Statement

The deployment of Vehicle-to-Everything (V2X) infrastructure faces critical challenges in securing communication links, ensuring privacy, and maintaining dependability, particularly within the context of Electric Vehicle (EV) charging and Smart Grid integration. Existing systems suffer from three primary limitations:

Data Limitations: A heavy reliance on synthetic or static datasets (e.g., VeReMi, NS-3 simulations) that lack real-world urban noise, latency, and GPS errors, leading to models that overfit clean data and miss static attacks in real interference.
Energy-Agnostic Detection: Current detection pipelines often ignore battery-specific features such as State of Charge (SoC), energy consumption rates, and battery temperature. This creates vulnerabilities where attackers can exploit charging mechanisms (e.g., fake low SoC for energy fraud) without triggering detection, as there is no established link between misbehavior and battery health.
Inadequate Attack Modeling: Many existing approaches focus on single attack types (e.g., only Sybil) or fail to model hybrid, concurrent attacks (e.g., Sybil combined with DDoS). Furthermore, they lack real-time adaptability to transition from benign to malicious states based on dynamic energy and behavior scores.

These gaps result in systems that cannot effectively mitigate static, dynamic, or hybrid cyber-attacks exploiting bidirectional V2G (Vehicle-to-Grid) channels, leading to vulnerabilities in charging authentication, data integrity, and grid coordination.

2. Methodology

The authors propose a Multilevel Trust CNN-LSTM (MTCNN-LSTM) framework designed to integrate real-time multi-attack detection, dynamic trust evaluation, and battery optimization.

2.1 System Architecture

The framework operates on a layered architecture leveraging Fog Computing:

Edge Layer: EVs and charging stations generate real-time data.
Network Layer: Utilizes Ethernet, 5G, WSN, and C-V2X for communication.
Data Processing Layer: Roadside Units (RSUs) equipped with fog nodes perform packet capture, data cleaning, and trust assessment using the deep learning model.
Management Layer: Handles intrusion response, long-term data storage, and cloud coordination.

2.2 Trust Evaluation Mechanism

The system computes trust scores based on nine telemetry features derived from the "Electric Vehicle Trip Energy Consumption Data" dataset. Trust is categorized into:

Local Trust (Total Trust): A weighted fusion of three real-time metrics:
- Behavior Score: Calculated based on Trip Energy Consumption and Speed.
- Feedback Score: Derived from historical interactions and subjective trust from neighboring vehicles/RSUs.
- Security Score: A dynamic metric updated in real-time based on authentication failure rates, IDS alert frequency, and battery temperature deviations.
Global Trust: The network-wide average of local trust scores, reflecting the overall reliability of entities within the V2X ecosystem.

2.3 The MTCNN-LSTM Model

The core detection engine is a hybrid deep learning model:

CNN (Convolutional Neural Network): Extracts spatial features from the input telemetry data.
LSTM (Long Short-Term Memory): Analyzes temporal dependencies in the time-series data to detect sequential anomalies.
Input Features: The model processes nine features: Security Score, Feedback Score, Behavior Score, Message Frequency, Messages Received, Energy Consumed, SoC, Security Score Difference, and Energy Consumed Rate.
Training Strategy: The model utilizes class-weighted training to address class imbalances between benign and malicious vehicles and employs multi-speed validation.

2.4 Energy Allocation and Battery Optimization

The framework introduces a closed-loop mechanism where trust scores directly influence energy allocation:

Battery Lifespan Factor ( $L_i$ ): A penalty is applied to the battery lifespan factor if overcharging occurs (SoC > 0.9) or if low SoC is detected during movement, linking malicious charging behavior to physical battery degradation.
Energy Waste Calculation: The system quantifies energy waste ( $E_{waste}$ ) resulting from capacity overflow (due to DDoS flooding) and denied energy to malicious vehicles (identified via low trust scores).
Dynamic Allocation: Charging stations prioritize energy distribution to vehicles with trust scores above a specific threshold (e.g., $\tau_i > 0.5$ ), denying requests from low-trust entities to prevent resource exhaustion.

3. Key Contributions

Real-World Data Integration: The study utilizes real-world EV trip logs (SoC, current, temperature) to simulate actual V2X noise, moving beyond purely synthetic datasets.
Battery-Conscious Detection: It is the first to integrate energy consumption rates, SoC, and battery temperature-based behavioral anomalies directly into the trust layer, enabling the detection of energy fraud and battery-related attacks.
Hybrid Attack Modeling: The model simultaneously detects and classifies four scenarios: None, Sybil, DDoS, and hybrid "Both" attacks within a single framework.
Closed-Loop Energy Management: The framework bridges intrusion detection and resource allocation, dynamically optimizing battery lifespan and energy distribution based on multilevel trust scores.

4. Results and Performance

The model was evaluated using a dataset of approximately 10,000 trip records, split into 80% training and 20% testing.

Classification Performance:
- Accuracy: 94.75% (reported in abstract) / 94.73% (reported in tables).
- Weighted F1-Score: 94.83%.
- Attack Detection Rate (ADR): 99.3% (Abstract) / 99.36% (Table 6).
- False Positive Rate: 4.24%, outperforming standalone GRU (5.36%) and LSTM (6.35%) models.
Attack Specifics: The model effectively distinguished between Sybil, DDoS, and hybrid attacks, with high precision in identifying "None" (benign) instances (1.00 precision).
Energy and Battery Impact:
- In simulations with a 30% attack probability, the system limited energy waste to approximately 550 kWh out of a 2000 kWh supply.
- The Average Battery Lifespan Factor (ABLF) was maintained at 98.23%, with an Average Battery Lifespan During Attacks (ABLF/DA) of 97.56%.
- As attack probability increased from 10% to 50%, global trust values decreased linearly (from 0.88 to 0.68), demonstrating the model's sensitivity to network degradation.

5. Significance and Claims

The paper claims that the proposed MTCNN-LSTM framework fundamentally bridges the gap between intrusion detection and resource allocation in Intelligent Transportation Systems. By leveraging real-world EV data and integrating battery health metrics into trust evaluation, the system achieves:

Resilience: Maintaining functionality and secure energy allocation even under high attack loads (up to 50% malicious nodes).
Sustainability: Preventing energy wastage and protecting battery lifespan from malicious overcharging or fraud.
Robustness: Achieving superior detection rates and lower false positives compared to existing state-of-the-art methods (e.g., XGBoost, standard CNN, GRU, and LSTM).

The authors position this work as a step toward secure, sustainable, and intelligent energy ecosystems for next-generation smart cities, with future work planned for edge deployment on 5G-V2X RSUs and integration with federated learning and blockchain.