TLDR: This research introduces a novel Agentic Artificial Intelligence (AAI) framework for the Internet of Electric Vehicles (IoEV) to address cyber-attacks, unreliable battery predictions, and opaque decision-making. The framework uses specialized AI agents for cyber-threat detection at charging stations, real-time battery State of Charge (SoC) estimation, and State of Health (SoH) anomaly detection, all coordinated through an explainable reasoning layer. It also includes a three-agent pipeline for user-centric assistance, leveraging large language models (LLMs) for intent interpretation and task optimization. The system is built on a five-layered architecture with strong privacy safeguards and has been validated through experiments showing significant improvements in security and prediction accuracy.
The world of electric vehicles (EVs) is rapidly evolving, moving towards a connected ecosystem known as the Internet of Electric Vehicles (IoEV). This vision promises seamless interaction between EVs, charging stations, and grid services, paving the way for greener and more sustainable transportation. However, this interconnectedness also brings significant challenges, including vulnerabilities to cyber-attacks, difficulties in accurately predicting battery states, and complex decision-making processes that can erode trust and performance.
To tackle these critical issues, researchers at L3i – La Rochelle University have introduced a groundbreaking Agentic Artificial Intelligence (AAI) framework specifically designed for the IoEV. This innovative framework employs specialized AI agents that work together to provide autonomous threat mitigation, robust data analytics, and transparent decision support, aiming to make the IoEV ecosystem more secure, reliable, and understandable.
A Novel Agentic AI Architecture for IoEV
The core of this research is an AAI architecture that features dedicated agents for various crucial tasks. These include agents for detecting and responding to cyber threats at charging stations, real-time estimation of a battery’s State of Charge (SoC), and identifying anomalies in a battery’s State of Health (SoH). All these agents are coordinated through a shared, explainable reasoning layer, ensuring that decisions are not only effective but also transparent.
The framework focuses on several key areas:
- Interpretable Threat Mitigation: Developing mechanisms that can proactively identify and neutralize attacks on both physical charging points and the learning components within the system.
- Resilient Battery Models: Proposing SoC and SoH models that continuously learn and adapt, even in the face of adversarial inputs, to produce accurate, uncertainty-aware forecasts with explanations that humans can easily understand.
- User-Centric Assistance: Implementing a unique three-agent pipeline where each agent uses advanced language models (LLMs) and dynamic tool invocation to interpret user intent, contextualize tasks, and execute formal optimizations for personalized assistance.
How the Agentic IoEV Framework Works
The system is built on a five-layered architecture, leveraging edge computing and 5G connectivity for secure and intelligent decision-making:
- IoEV Layer: Collects data from vehicles and charging stations, such as voltage, temperature, and traffic.
- Network Layer: Uses 5G and Software-Defined Networking (SDN) for fast and reliable communication.
- Edge Computing Layer: Hosts autonomous agents and supports distributed, low-latency processing, allowing models to adapt to new data incrementally.
- Agentic AI Layer: Deploys task-specific agents for cybersecurity, battery analytics, and user services, all with context-aware and explainable reasoning.
- Application Layer: Provides an interface for users and operators, offering transparency, feedback, and adaptive control.
Privacy and secure data handling are paramount. The framework employs measures like encrypted data transmission, minimal data retention, agent isolation, and federated learning with differential privacy to protect sensitive information.
The Agentic Workflow in Detail
The system’s workflow begins with an Embedded Intent Recognizer in the EV or charging station, which interprets user queries or system alerts. This information is then routed to specialized agents:
- Safety and Security Agent (SSA): This agent, augmented with domain-specific knowledge and tools, handles safety and security requests. It invokes diagnostic tools for EV battery safety or charging station security, uses explainability tools like SHAP or LIME to interpret results, and then generates user-friendly explanations for drivers or operators.
- Multi-Agent Coordination and Support: For high-level user requests, a three-agent pipeline is activated:
- Personalized Support Agent (PSA): Translates natural language requests (e.g., “charge my EV as cheaply as possible”) into a formal optimization problem skeleton, identifying necessary parameters.
- Contextualizer Agent (CA): Enriches the problem skeleton by mapping abstract parameters to real-world values, using pre-trained models to forecast electricity prices, charging station demand, and user behavior patterns. This agent accesses data transiently and securely.
- Solver Agent (SA): Takes the fully contextualized problem and uses external numerical solvers (e.g., from scipy.optimize, cvxpy) to find accurate solutions, ensuring physical and quality-of-service constraints are met.
Also Read:
- Large Language Models Transform Security Operations Centers: A Survey of Capabilities and Future Directions
- AI-Powered Solutions for Flexible Automotive Architectures
Validation and Future Outlook
The framework was rigorously validated through comprehensive experiments across diverse IoEV scenarios. The results demonstrated significant improvements in security and prediction accuracy for battery state diagnosis and cyber-attack detection. For instance, models achieved high accuracy in detecting EVCS attacks and estimating battery health, while the explainability features provided clear, human-understandable insights into the model’s decisions.
The research paper, titled “Towards Trustworthy Agentic IoEV: AI Agents for Explainable Cyberthreat Mitigation and State Analytics” by Meryem Malak Dif, Mouhamed Amine Bouchiha, Abdelaziz Amara Korba, and Yacine Ghamri-Doudane, was accepted at the 50th Annual IEEE Conference on Local Computer Networks (LCN’25). You can read the full paper here: Research Paper.
Looking ahead, future research will delve into dynamic agent collaboration strategies, real-world field trials, and integration with emerging Vehicle-to-Grid (V2G) standards. A crucial area will also be investigating the adversarial resilience of LLM-driven agents against threats like prompt injection and adversarial intent manipulation, further strengthening the security of autonomous agents in safety-critical IoEV applications.
