New AI Spots Power Grid Hacks — And Shrugs Off Tampering

Power grids are increasingly digital, which means hackers have more ways in. Attacks on electrical networks aren't hypothetical — knocking out or spoofing grid controls can cause blackouts, damage equipment, and disrupt hospitals, heating, and businesses. So utilities rely on intrusion detection systems: software that watches network traffic and sensor readings for signs of an attack. The problem is that the best of these systems are heavy, power-hungry AI models that are hard to run on the small computers sitting inside substations.

A team of researchers in Australia proposes a leaner recipe. They combine two AI approaches: a spiking neural network (an AI that mimics how brain cells fire in pulses, using very little energy) that turns sensor data into a compact summary, plus a lightweight classifier called XGBoost that makes the final call. Because the brain-like part is trained once and then frozen, only the small classifier needs periodic retraining — cheap and quick.

The results: up to 99.9% accuracy spotting attacks on one real power-system dataset and 95% on another, beating the standalone alternatives. Just as important, the system resists "machine unlearning" attacks — where an adversary slips bad or corrupted data into retraining to blind the detector. Under that kind of poisoning, the new model lost only 0.9% of its accuracy, and only broke down once 70% of the data was corrupted, versus 60% for conventional models.

The catch: this is a first draft on public datasets, not a tested product inside a live utility. Real grids are messier, and attackers adapt. Still, the direction matters — cheaper, tougher security for the infrastructure everything else depends on.