Source-linked AI summary
GIDS: GAN based Intrusion Detection System for In-Vehicle Network
Eunbi Seo, Hyun Min Song, Huy Kang Kim
TL;DR
In-vehicle CAN buses lack authentication and have few known attack signatures, making accurate intrusion detection important for driver safety. GIDS uses GAN-based deep learning, normal-data training, and two discriminators to detect known and unknown attacks. It reports 100% average accuracy for the first discriminator and 98% average accuracy for the second across four attacks.
Problem
CAN buses lack authentication, and vehicle IDSs face limited known attack signatures while false positives can threaten driver safety.
Method
GIDS combines an attack-trained first discriminator with a GAN-trained second discriminator using normal and generated fake CAN data.
Results
100% average accuracy was reported for the first discriminator, while the second achieved 98% average accuracy across four attacks not used in training.
Takeaways & Limitations
GIDS can detect unknown attacks using only normal data and can be combined with the first discriminator to detect both known and unknown attacks.
Takeaways & Limitations
GAN-based IDS still struggles to distinguish normal electronic-component malfunctions from intentional hacker attacks.
Abstract
from arXiv · showhide
A Controller Area Network (CAN) bus in the vehicles is an efficient standard bus enabling communication between all Electronic Control Units (ECU). However, CAN bus is not enough to protect itself because of lack of security features. To detect suspicious network connections effectively, the intrusion detection system (IDS) is strongly required. Unlike the traditional IDS for Internet, there are small number of known attack signatures for vehicle networks. Also, IDS for vehicle requires high accuracy because any false-positive error can seriously affect the safety of the driver. To solve this problem, we propose a novel IDS model for in-vehicle networks, GIDS (GAN based Intrusion Detection System) using deep-learning model, Generative Adversarial Nets. GIDS can learn to detect unknown attacks using only normal data. As experiment result, GIDS shows high detection accuracy for four unknown attacks.
1 Introduction
In-vehicle CAN networks lack authentication, exposing ECUs and driver safety to attacks. GIDS addresses limitations of attack-specific IDS models by using normal data and GAN-based detection for broader, harder-to-manipulate monitoring.
- CAN broadcasts without authentication, allowing attackers to inject packets through external interfaces or OTA modules.Such attacks can cause vehicle malfunctions and threaten driver safety.
- Real-time IDS monitoring is required to detect suspicious ECU-generated network events and respond to known and unknown attacks.
- Existing lightweight IDS models require updates when environments change and may detect only attacks represented during construction.
- GIDS is designed around expandability, effectiveness, and security to address these limitations.It uses only normal data for training and a deep-learning black-box structure that is difficult for attackers to manipulate.
2 Related works
Prior in-vehicle anomaly detection used CAN timing, entropy, sequence, and related features, while GAN research largely focused on image discrimination. This work extends GAN-based IDS to in-vehicle security and reports high performance on CAN data.
- Earlier in-vehicle anomaly detectors examined cyclic-message frequency, low-level communication characteristics, message IDs, entropy, and CAN-message sequences.
- Deep-learning IDS research has addressed security detection, and GANs have expanded beyond image generation into other research areas.
- Few studies had explored GANs for security outside other fields, motivating a GAN-based IDS for in-vehicle CAN security.The authors report high performance and claim expandability, effectiveness, and security for in-vehicle networks.
3 GIDS: GAN based IDS
GIDS converts CAN identifiers into compact one-hot-encoded images and uses GAN-based discriminators to distinguish normal from abnormal traffic. Its design combines known-attack detection with adversarial training intended to detect unknown attacks.
- CAN data representation: GIDS extracts repetitive CAN-ID patterns and converts them into simple images to reduce real-time detection time and improve IDS performance.Using all CAN bits directly could produce complex images unsuitable for real-time detection.
- CAN data representation: One-hot encoding transforms each hexadecimal CAN-ID element into a binary representation with one active bit, producing a CAN-image matrix.
- GAN-based training: GAN training simultaneously uses a generator that captures data distribution and a discriminator that estimates whether samples are real or generated.
- Two-discriminator detection: The first discriminator learns from normal and actual abnormal CAN images, so its detectable attack types are likely limited to those used in training.
- Two-discriminator detection: The second discriminator is adversarially trained with normal CAN images and generator-produced fake images instead of real attack data.
- Two-discriminator detection: GIDS combines both discriminators so the system can target known attacks with attack-trained data and unknown attacks using only normal data.The combined design addresses the lower accuracy that may result from using only the second discriminator.
4 Experiment and Result
The experiments evaluated GIDS on vehicle CAN traffic using defined detection criteria, attack scenarios, data encoding, hyperparameters, and discriminators. The second discriminator detected four attacks absent from training with an average accuracy of 98%.
- Detection rate measures detected abnormal data among all abnormal data, while accuracy measures correctly classified normal and abnormal data.
- Experiment Environment: The dataset was collected from a Hyundai YF Sonata through its OBD-II port using Raspberry Pi3 and included labeled normal and abnormal traffic.
- Attack Design and Dataset: Four CAN-bus attacks were launched: DoS, FUZZY, RPM/GEAR, and GEAR attacks.
- Evaluating One-hot-vector Encoding: One-hot-vector encoding produced simpler CAN images and separated normal from attack data with a clearer classification threshold than binary encoding.The encoded representation also reduced required time and showed better performance than converting binary CAN data directly.
- Hyperparameters: The selected detection threshold was 0.1, input size was 64, and an attack threshold of 1 flagged images containing at least one attack packet.Accuracy increased through input size 64 and tended to decrease afterward; the input size can be changed for different vehicle environments.
- Experiment Result: The first discriminator detected trained attacks well but hardly detected unseen attacks, whereas the second discriminator detected four unseen attacks with an average accuracy of 98%.The second discriminator used random fake data instead of real attack data during training.
5 Conclusion
GIDS combines one-hot CAN-ID encoding, random fake-data training, and two discriminators to detect known and unknown intrusions using normal data. It achieved 100% average accuracy for known attacks and 98% for unknown attacks, while processing 1,954 CAN messages in 0.18 seconds; distinguishing attacks from normal component malfunctions remains challenging.
- GIDS uses one-hot CAN-ID encoding, random fake data, and two discriminators for known and unknown intrusion detection.The first discriminator detects known attack data, while the second detects unknown attack data.
- 100% average accuracy was achieved by the first discriminator, compared with 98% for the second discriminator.
- 0.18 seconds are required to detect about 1,954 CAN messages, with elapsed time maintaining a constant ratio as data volume increases.
- GIDS is described as expandable, effective, and secure for in-vehicle networks.The model can be applied to various vehicle types through new training and hyperparameter adjustment.
- Distinguishing anomalous traffic from normal electronic-component malfunctions versus intentional hacker attacks remains challenging.The authors state that evaluation becomes more precise as more vehicle attack patterns are revealed.