A fly-brain detector for factory control attacks

Factories run on small controllers called PLCs. I tested a detector inspired by the fruit fly brain, a sparse code with a one-pass memory, on real industrial control data with 50 attacks.

On healthy hardware it lost to a standard method, PCA, with a score of 0.81 against 0.88.

Then I set 20 percent of the stored numbers to zero, like failing memory. A hyperdimensional detector lost nothing (0.793 to 0.794). The fly code lost 0.03, PCA lost 0.04, and a neural network autoencoder lost 0.08. With its alarm threshold set on day one, the autoencoder was in alarm 58 percent of the time at only 5 percent faults.

So the brain-inspired part did not win, and the fault tolerance result held up. It is a pilot, and a paper with my research advisor is in progress.