Securing Autonomous Fleets with Global-Trained Localized Brains

Securing Autonomous Fleets with Global-Trained Localized Brains



The value of connected everything will be derived from unprecedented data sharing, none the more so than with autonomobiles.  Optimizing traffic patterns, energy consumption and parking utilization with rider comfort and safety; more global data yields a better local experience.  Conversely, concerns about the security  are well documented and rightly so; after all, with human lives strapped in the passenger seat, all hypothetical IoT cyber fears suddenly become visceral. Despite the fact it could be “the great public-health achievement of the 21st century”, we cannot help but fear the idea of a small team of hackers wreaking international havoc on our highways.  After all, the damage caused by any single drunk driver or texting accident is local, limited by the laws of mass and velocity, rather than the global potentiality of electrons run amok with no attributable pilot. A catastrophic wreck with no one to blame.

Furthermore, a parallel meta-threat exists beyond the periphery of immediate physical danger.  A clever hacker knows that the value of the data transmitting between these devices is ever more valuable and pernicious in the long run. Omniscience over everyone’s commuting patterns, time at work, time at home (and away from home), coupled with wearable data and in-home devices creates a full DNA-in-silica of every individual.  Yet society needs the global benefits of energy optimization, time and convenience, so how are security, privacy and utility measured?

There is no silver bullet, and calls to delay deployment of fully autonomous fleets may be justified, but there are some novel approaches that dramatically mitigate these challenges. Just as Apple has touted the use of privacy-preserving data mining to balance the equation of convenience and individual privacy, there are new, industrial cryptographic methods that can address similar challenges for IoT and autonomous vehicles.  It is possible to train machine learning models with private data sets so that no single data point is identified but statistical learning is maintained, including outliers like balls bouncing in the road or black ice conditions.  Just like with humans, the more you experience, the more you know how to react in the future.  Expanding the training materials is absolutely necessary for safety and cybersecurity, particularly to avoid spoofing image classification algorithms. But this is only step one.

Once the model is trained, vetted and deployed in the fleet, the ability for any party to instantaneously override its intelligence must be mitigated.  Each vehicle is thus allowed decision making autonomy and maintains the privacy of new inputs, including location information, while still communicating with the world to optimize global performance.  The meta-model is only updated when a given threshold of the fleet experiences similar inputs that affect their operating behavior, which is a great application for distributed ledger technology like blockchain; however, instead of ecologically disastrous ‘proof-of-work’ algorithms used in Bitcoin, it could be a ‘proof-of-input’ or similar function.  This incidentally opens a great opportunity for shared-asset ownership and tracking as well.

In the hyperbole of IoT Part 1, things were connected just because they could be. In the automotive industry, things must be connected because they should be. The former brought justified scrutiny of utility, privacy and security, while the latter dispels the myth of mutual exclusivity through global learning and local intelligence.