TL;DR
Researchers have developed a prototype electric scooter that uses computer vision to detect its driving environment (sidewalks vs. bike lanes) and automatically enforces legal speed limits. Beyond simple automation, the paper details a robust security architecture—utilizing Trusted Execution Environments (TEE) and secure boot—to prevent hackers from tampering with these safety-critical speed governors.
What Problem Does This Paper Address?
Electric scooters (e-scooters) have become a staple of urban transportation, but they present a unique regulatory challenge. Depending on whether a rider is on a sidewalk, a bike path, or a main road, legal speed limits change drastically to ensure pedestrian safety. Currently, enforcement relies entirely on the rider’s discretion.
While automated lane detection and speed limiting seem like the obvious solution, they introduce a significant "cyber-physical" risk. If a scooter’s control system is hacked, an attacker could bypass speed limits, leading to dangerous accidents. Most current motor controllers are designed for performance, not security, lacking the hardware-level protections found in modern smartphones or servers. The authors address the gap between automated safety features and the underlying security architecture required to make those features trustworthy.
Methods
The authors built a prototype using a two-tier computing architecture:
- Main Control Unit (MCU): A Raspberry Pi 4 equipped with a Google Coral Edge TPU. This unit handles the "heavy lifting"—running deep learning models for computer vision to classify the lane type (pedestrian, cycle, or automobile).
- Motor Controller (MC): A modified VESC (Open Source ESC) based on the STM32F405 microcontroller. This handles the real-time physics of motor control.
The Security Stack
To protect the system, the authors implemented a layered defense-in-depth strategy:
- Trusted Execution Environment (TEE): On the MCU, they partitioned functions into "Secure" and "Non-secure" contexts. Sensitive operations like firmware updates and key storage happen in the secure zone.
- Secure Boot & RDP: The system uses Read-out Protection (RDP) to prevent attackers from dumping firmware via debug ports. A Secure Boot process ensures that only cryptographically signed, authentic code runs upon startup.
- Memory Protection Unit (MPU): The motor controller uses an MPU to define "Execute Never" (XN) regions in memory, effectively neutralizing code-injection attacks.
- Encrypted Communication: External data exchange (via NB-IoT) uses the MQTT protocol wrapped in SSL/TLS to prevent man-in-the-middle attacks.
Datasets
The paper does not introduce a new public dataset. The computer vision component for lane detection is based on the authors' prior work (referenced as [4] in the paper), which focuses on the transition of vehicle speeds. The authors note that the specific performance of the CV algorithm is outside the scope of this paper, as their primary focus is the secure design of the hardware and software integration.
Key Insights and Ideas
- Policy Enforcement via Control Loops: The MCU doesn't just tell the motor how fast to go; it sends high-level "policies" ($p_k$) to the motor controller. The MC then integrates these policies into its internal control loop, ensuring that even if a user pins the throttle, the motor refuses to exceed the lane-appropriate limit.
- Smooth Transitions: To prevent jerky, dangerous deceleration when moving from a road to a sidewalk, the system uses a two-second window to logarithmically ramp the speed down.
- The "Insider" vs. "Outsider" Threat Model: The authors distinguish between an outsider (trying to hack via Bluetooth/Cellular) and an insider (someone with physical access to the MCU). Their TEE approach is specifically designed to thwart an attacker who has already gained a foothold in the non-secure part of the operating system.
Key Results
The authors successfully demonstrated a functional prototype that dynamically adjusts its maximum speed based on visual input:
- Pedestrian Lane: The system capped speed at 6 km/h, even when the driver requested maximum power.
- Bike Lane: The limit was increased to 15 km/h.
- Car Lane: While the display showed the capability for higher speeds (up to 45 km/h in some regions), the prototype maintained a safe regulated cap (15 km/h in their specific test configuration) to demonstrate policy enforcement.
Qualitatively, the security features (Secure Boot and RDP) were shown to protect the secret keys required to sign firmware, making it significantly harder for a "tuned" firmware to be installed by an end-user looking to bypass speed caps.
Limitations and Future Work
- CAN Bus Vulnerability: The internal communication between the MCU and the Motor Controller happens over a CAN bus. The authors admit that CAN is inherently insecure and lacks native encryption/authentication. Currently, they assume an attacker cannot attach external hardware to the bus, but they acknowledge this as a weak point.
- Performance Overhead: While the security features add protection, they also add latency and power consumption. The authors plan to measure the exact computational overhead and cost implications of this secure design in future iterations.
- Complex Attack Strategies: The authors note that a highly sophisticated "insider" attacker might still find ways to read sensitive data if the isolation between the MCU and MC isn't perfectly mirrored across both devices.
Summary updated