Cyber Sentinel: Safeguarding Autonomous Vehicle Supply Chains against Cyber Backdoors and Hardware Faults
To secure the deep neural network (DNN) of autonomous vehicles (AV) against backdoor attacks, researchers will develop a system-aware verification technique (SAVE) inspired by software fuzzing.
Project Investigators
Rationale and Background
A decentralized and distributed supply chain is developing software and hardware modules for AVs, which provides such benefits as faster time-to-market and lower operating and maintenance costs.
However, such a complex distributed supply chain introduces security threats from both software and hardware development perspectives.
In an attack, a backdoored DNN behaves normally with clean inputs, but if a trigger is presented, the input will be misclassified as a target. Safeguarding hardware is equally pivotal.
By bridging the gap between traditional defenses and the intricate demands of the physical realm, researchers seek to ensure the safety and reliability of AVs in real-world scenarios.
Methodology
Researchers plan a Physical Contextual Analysis for Object Detection (PCAOD) system to address backdoor challenges by integrating contextual understanding with object detection capability through:
- Contextual Object Detection.
- Anomaly Detection.
- Feedback Loop with Historical Data.
- Backdoor Trigger Verification.
- Multi-modal Verification.
The proposed methodology bridges the physical-digital dichotomy and addresses the stealth and complexity of potential physical triggers.
In addition to defending the backdoors in the deployed AI/ML models from untrusted supply chains, securing and validating the hardware components is equally pivotal.
Researchers will introduce a system-aware hardware verification technique (SAVE), a hardware fuzzing technique for AVs, which is inspired by software fuzzing. The fuzzing technique comprises input stimuli generation and feeding the input to the design under test (DUT).
Projected Outcomes
Researchers will introduce an open-sourced hardware verification technique for AVs that can detect functional bugs and security exploits included in the supply chain by untrusted vendors. These frameworks will be released for community use.
The technique will be tested on driving simulators available at George Mason University and the University of Virginia.
Researchers will also use the Simulink tool to model AVs’ hardware modules. Based on these models, researchers will execute the SAVE for functional and security evaluation.
The deliverable of this project includes the PCAOD-enabled DNN technique and SAVE framework.