Behavior-Aware Efficient Detection of Malicious EVs in V2G Systems

Ruixiang Wu, Xudong Wang, Tongxin LiarXiv 2024 · OpenReview

Abstract

With the rapid development of electric vehicles (EVs) and vehicle-to-grid (V2G) technology, detecting malicious EV drivers is becoming increasingly important for the reliability and efficiency of smart grids. To address this challenge, machine learning (ML) algorithms are employed to predict user behavior and identify patterns of non-cooperation. However, the ML predictions are often untrusted, which can significantly degrade the performance of existing algorithms. In this paper, we propose a safety-enabled group testing scheme, GTUA, which combines the efficiency of probabilistic group testing with ML predictions and the robustness of combinatorial group testing. We prove that GTUA is $O(d)$-consistent and $O(d\log n)$-robust, striking a near-optimal trade-off. Experiments on synthetic data and case studies based on ACN-Data, a real-world EV charging dataset validate the efficacy of GTUA for efficiently detecting malicious users in V2G systems. Our findings contribute to the growing field of algorithms with predictions and provide insights for incorporating distributional ML advice into algorithmic decision-making in energy and transportation-related systems.

Introduction

Vehicle-to-Grid (V2G) technology is a cornerstone of the future smart grid, allowing EVs to act as mobile batteries that supply power during peak demand. However, this system relies heavily on user cooperation. A "malicious" or non-cooperative user might commit to leaving their car plugged in until 6:00 PM but drive away at 4:00 PM, depriving the grid of expected energy.

Detecting these users is surprisingly difficult. Installing high-precision meters at every single residential charger is prohibitively expensive ($60–$250 per unit) and raises significant privacy concerns regarding user data. The challenge is to identify which specific users are failing to meet their commitments using the fewest possible measurements—ideally by monitoring aggregate power at a substation level.

Methodology: The GTUA Framework

The researchers frame this as a Group Testing problem. Instead of checking every car individually, the system checks a subset (a "group"). If the total power injection from that group deviates from the expected value, the system knows at least one user in that group is non-cooperative.

System architecture of the GTUA framework for malicious EV detection in V2G systems.

The core innovation is GTUA (Group Testing with Untrusted ML Advice). It addresses a common pitfall: while ML can predict user behavior based on history, those predictions are often "untrusted" or noisy. GTUA uses a "meta-testing" approach to balance efficiency and reliability:

  1. Safety Thresholding: It uses a threshold ($\eta$) to split users into two pools based on their ML-predicted probability of being malicious.
  2. Probabilistic Testing (Laminar Algorithm): Users with high-confidence predictions are processed using an algorithm optimized for known probabilities.
  3. Combinatorial Testing (Generalized Binary Splitting): Users with low-confidence or "unreliable" predictions are processed using a robust method that doesn't rely on ML at all.

This dual-path approach ensures that the system doesn't "break" or become wildly inefficient if the ML model encounters a new behavior pattern it hasn't seen before. By treating ML output as a potentially fallible hint, GTUA achieves a near-optimal trade-off between consistency (speed when ML is right) and robustness (reliability when ML is wrong).

Experimental Setup

The study validates GTUA using two primary sources:

  • Synthetic Data: Used to theoretically stress-test the algorithm across varying levels of prediction error (measured via KL-divergence).
  • ACN-DATA: A real-world public dataset of EV charging sessions from Caltech. The researchers used Gaussian Mixture Models (GMMs) to learn "normal" behavior (arrival times, durations, and deviations) and then simulated malicious departures to see if GTUA could catch them.

Key Results

In real-world simulations using ACN-DATA, GTUA reduced the number of required tests by 40% to 80% compared to individual monitoring.

Comparison of GTUA against baseline algorithms across varying levels of prediction error.

As shown in Figure 2, GTUA maintains a stable, low number of tests even as the ML prediction error (Pseudo KL-Divergence) increases. In contrast, the performance of standard probabilistic algorithms like the Laminar Algorithm (LA) degrades sharply when predictions are inaccurate. GTUA's combinatorial "safety net" ensures it remains competitive with Generalized Binary Splitting (GBS) even in worst-case scenarios.

Mathematically, the authors prove that the algorithm is $O(d)$-consistent and $O(d \log n)$-robust. This means it scales efficiently even as the total number of drivers ($n$) grows much larger than the number of malicious drivers ($d$).

Conclusion and Future Work

GTUA provides a robust solution for identifying non-cooperative EVs while minimizing infrastructure costs and preserving privacy through sparse testing. While the results are promising, future work could explore non-adaptive settings for grid hardware that may not support real-time adaptive testing, or more advanced deep learning models to further tighten the consistency of detection. Additionally, addressing strategic users who might intentionally alter their behavior to evade detection remains a key area for future research.

Summary updated

Abstract

With the rapid development of electric vehicles (EVs) and vehicle-to-grid (V2G) technology, detecting malicious EV drivers is becoming increasingly important for the reliability and efficiency of smart grids. To address this challenge, machine learning (ML) algorithms are employed to predict user behavior and identify patterns of non-cooperation. However, the ML predictions are often untrusted, which can significantly degrade the performance of existing algorithms. In this paper, we propose a safety-enabled group testing scheme, \ouralg, which combines the efficiency of probabilistic group testing with ML predictions and the robustness of combinatorial group testing. We prove that \ouralg is $O(d)$-consistent and $O(d\log n)$-robust, striking a near-optimal trade-off. Experiments on synthetic data and case studies based on \textsc{ACN-Data}, a real-world EV charging dataset validate the efficacy of \ouralg for efficiently detecting malicious users in V2G systems. Our findings contribute to the growing field of algorithms with predictions and provide insights for incorporating distributional ML advice into algorithmic decision-making in energy and transportation-related systems.
BibTeX
@misc{wu2024behavior,
  author = {Ruixiang Wu and Xudong Wang and Tongxin Li},
  title = {{Behavior-Aware Efficient Detection of Malicious EVs in V2G Systems}},
  year = {2024},
  eprint = {2411.06113},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.06113},
}
Figure