TL;DR
This paper introduces a novel two-stage deep learning model for detecting and classifying attacks on in-vehicle CAN-FD communication. By combining Long Short-Term Memory (LSTM) networks with an attention mechanism, the proposed system achieves high accuracy in identifying both anomalous messages and their specific attack types, outperforming existing methods.
What Problem Does This Paper Address?
Modern vehicles rely heavily on in-vehicle networks, particularly the Controller Area Network with Flexible Data (CAN-FD) rate, for communication between Electronic Control Units (ECUs). While CAN-FD offers increased speed and data capacity over traditional CAN, it lacks inherent security features, making it vulnerable to various cyberattacks. These attacks can range from disrupting normal operations (Denial of Service - DoS) to injecting malicious commands, posing significant risks to vehicle safety and functionality. Existing anomaly detection methods often struggle with accuracy, especially in classifying the specific type of attack, and may require extensive prior knowledge of network behavior. This paper aims to address this gap by developing a more robust and accurate intrusion detection system for CAN-FD messages.
Methods
The proposed solution consists of two cascaded deep learning sub-models:
- Anomaly Data Detection Model (ADDM): This model uses stacked LSTM layers to learn long-range dependencies and temporal patterns within CAN-FD frame data. It performs binary classification, distinguishing between normal and anomalous messages. The LSTM architecture is well-suited for sequential data like network traffic, allowing it to capture evolving patterns.
- Anomaly Classification Detection Model (ACDM): This model builds upon ADDM by incorporating a temporal attention mechanism. The attention mechanism allows the model to dynamically focus on the most relevant parts of the input sequence, improving its ability to identify subtle attack characteristics and relationships between messages. This enables multi-class classification, identifying specific attack types such as Replay, DoS, Fuzzing, Spoofing, Scaling, and Ramp attacks.
The overall framework involves data collection, anomaly injection to create an attack dataset, data preprocessing (min-max normalization), and then feeding the processed data into ADDM for initial anomaly detection, followed by ACDM for classification.
Datasets
The proposed method is evaluated on two datasets:
- Real-vehicle dataset: This dataset was collected from a real vehicle with L2 autonomous driving capability. Anomalous frames were generated by the authors based on known attack patterns.
- CAN-FD Intrusion Dataset: This dataset was provided by the Hacking and Countermeasure Research Lab (HCRL) and contains Flooding, Fuzzing, and Malfunction attacks.
Key Insights and Ideas
- The power of attention for CAN-FD: The paper highlights that while LSTMs are good at capturing temporal dependencies, the addition of an attention mechanism significantly enhances the ability to pinpoint critical features within anomalous CAN-FD messages, leading to better classification of specific attack types.
- Two-stage detection is effective: Separating the task into anomaly detection (ADDM) and then classification (ACDM) allows for a more focused and robust approach, improving overall accuracy.
- Real-world data is crucial: The use of a real-vehicle dataset, augmented with synthetically generated attacks, provides a more realistic evaluation than purely simulated environments.
- Beyond simple periodicity: The proposed method moves beyond traditional periodicity-based detection, which can miss dynamically timed attacks, by focusing on the semantic and temporal patterns within the message content.
Key Results
- ADDM Performance: On the real-vehicle dataset, ADDM achieved an accuracy of 0.9995, outperforming the LSTM-based model (0.9900) and HyDL-IDS (0.9935) by approximately 0.50% and 0.41%, respectively.
- ACDM Performance: For multi-class anomaly detection on the real-vehicle dataset, ACDM achieved a maximum accuracy improvement of 1.44% over the LSTM-based model and 1.01% over HyDL-IDS.
- Attack-Specific Improvements: ACDM demonstrated superior performance in detecting Replay, Fuzzing, Spoofing, Scaling, and Ramp attacks, particularly showing a significant reduction in False Negatives (FNs) and False Positives (FPs) compared to baseline models.
- HCRL Dataset: On the HCRL dataset, ACDM achieved comparable or slightly superior results to the baseline models, with perfect scores (1.0000) for DoS and Flooding attacks.
Limitations and Future Work
- Limitations: While the model shows excellent performance, the paper notes that the training times for ACDM and the comparative models are closely aligned, suggesting potential for further optimization in inference speed. The authors also acknowledge that the current approach might not be sufficient for entirely novel attack types that deviate significantly from known patterns.
- Future Work: The authors propose to delve deeper into anomaly detection by considering the physical significance of anomalous signals. They aim to develop a more adaptive framework capable of dynamically adjusting to new types of network threats. This will involve using causal relationships of CAN/CAN-FD messages and graph structures, potentially leveraging transformers for more advanced predictions, classifications, and scoring mechanisms.
Summary updated