Models on the Move: Towards Feasible Embedded AI for Intrusion Detection on...
summary
Proposes MULSAM, a multi-dimensional LSTM with self-attention tuned for vehicular CAN bus intrusion detection, trained on five attack types synthesized from the 4TU driving dataset to catch small-batch injection attacks. Reports 98.98% detection accuracy (quantized accelerator at 98.81%) across replay, DoS, flooding, spoofing, and deletion attacks, implements a convolutional-heavy architecture and attention pipeline on an FPGA with 1.88 ms latency, and shows real-time throughput exceeding a saturated CAN bus.
AI-written summary · about
chapters (8)
- 0:00Embedded AI in vehicles
- 2:55ML-based intrusion detection challenges
- 4:04Proposed solution overview
- 6:17CAN bus attack datasets
- 8:03MASAN architecture & FPGA optimization
- 12:46Experimental performance comparison
- 14:53Real-world FPGA testing
- 16:42Conclusion & future work
related talks
12mOn the Real-World Applicability of Automotive CAN Intrusion Detection Systems
14mAutoHack: A Physically Verified Multi-Bus CAN Dataset for Intrusion Detection...
14mCANBERT+: An Optimized Multi-Layer BERT-Based Intrusion Detection for CAN Bus
14mStateful Behavior Inference and Runtime Enforcement for Vehicle Network Security
30mFingerprinting Electronic Control Units for Vehicle Intrusion Detection
12mCARdea: Two-Phase Plausibility-Based Vehicle-to-Vehicle Anomaly Detection...
48mStay Ahead of Adversarial AI in OT/ICS Environments – Mitigating CWE-1039
15mThreat Analysis and Detection in In-Vehicle Infotainment System Leveraging MITRE...