As millions of self-driving cars are expected on the roads by 2025 and autonomous drones continue to generate billions in revenue, ensuring the safety and reliability of these machines has become a key concern for consumers, manufacturers, and regulators. However, the measures needed to safeguard autonomous hardware and software against malfunctions, cyberattacks, and other failures often drive up costs, impacting performance, energy efficiency, and weight, largely due to the use of semiconductor chips.

Researchers from the University of Rochester, Georgia Tech, and the Shenzhen Institute of Artificial Intelligence and Robotics for Society argue that the current trade-off between protection and cost stems from a “one-size-fits-all” approach to safeguarding these systems. In a recent paper published in Communications of the ACM, they propose a new adaptive protection model that tailors security measures to different parts of the system, optimizing reliability without disproportionately increasing costs.

Yuhao Zhu, an associate professor of Computer Science at the University of Rochester, points to Tesla’s Full Self-Driving (FSD) Chips as an example of the current one-size-fits-all strategy. Tesla uses two FSD chips per vehicle to provide redundancy in case one fails, but this approach effectively doubles the cost. Zhu’s team suggests a more nuanced method: applying varying protection levels to different parts of the hardware and software based on their specific vulnerabilities.

“The key idea is to adopt different protection strategies for different components of the system,” Zhu explains. “We can refine these strategies by considering the inherent characteristics of the hardware and software. For instance, the front end of the software stack—responsible for sensing through cameras or LiDAR—requires less protection than the back end, which processes that data and sends critical commands to the vehicle’s mechanical systems.”

According to Zhu, the front end is naturally more fault-tolerant, meaning that low-cost software solutions, like anomaly detection filters, can be sufficient to maintain reliability. However, the back end—responsible for interpreting sensor data and executing decisions—requires more robust safeguards, such as checkpointing to save system states or duplicating critical modules on a chip to prevent system failure.

Zhu and his team are now focusing on addressing vulnerabilities in the latest autonomous systems, which heavily rely on end-to-end neural network AI models. These models, while boosting average performance, complicate failure diagnosis since they operate as a single, opaque system.

“With these massive neural networks, the computations are often so complex that when something goes wrong, it’s nearly impossible to identify the source of the problem,” says Zhu. “This improves performance in general, but makes worst-case scenarios even worse. Our goal is to find ways to mitigate these risks and improve overall system resilience.”

By adopting a vulnerability-adaptive design approach, the team aims to strike a balance between cost, energy efficiency, and reliability in the next generation of autonomous machines.

Source: Communications of the ACM (2024). DOI: 10.1145/3647638