AI Trustworthiness Challenges and Opportunities Related to IIoT
Figure 2
learning. It’s because Q-Bert was easily
accessible and could be implemented on a
platform other than the original hardware
for which it was intended, that it was
possible to analyze it using AI. Use of new
hardware enabled the performance
necessary to run many iterations of learning.
Had the Q-Bert ROM been locked away from
prying eyes, then this ‘attack’ would not
have been feasible since it could not then
have been used to recreate a system to
enable learning. Similarly, if the details of a
flight system cannot be replicated, then it
would be more difficult if not impossible to
find exploits in the system through machine
learning.
A New Arms Race!
What about bad actors? It’s likely this is just
the beginning of yet another arms race, but
there are things that can be done to mitigate
risk.
AI systems rely on the ability to test millions
of strategies in a short period of time to find
what works and what doesn’t. Although
keeping intellectual property out of the
hands of competitors can prevent reverse-
engineering,
the
game
example
demonstrates that machine learning can be
effective simply with access to a human-
machine interface. This may be expensive,
such as with an actual airline flight simulator,
but not out of reach of nation-state or other
actors.
It is inevitable that AI will become more
commonplace as a part of testing and quality
assurance regimes due to its many benefits.
Organizations should take steps to ensure
that their IP remains available to their own
Such attacks are possible because the
systems can be obtained to use in machine
- 79 -
June 2019