Modern power systems are no longer just massive machines turning in the dark; they are vast, interconnected networks of digital sensors that watch every wire, switch, and generator. These sensors, often called smart meters or process tags, send a constant stream of data to a central control center, where computer models try to spot anything unusual. If a pipe bursts, a generator fails, or a line is cut, the numbers should change in a way that looks different from normal operation. The hope has always been that these computer models could act as an unblinking eye, catching problems the moment they start. But this reliance on a single, central brain creates a dangerous weakness. If an attacker manages to take control of even a few of these sensors, they can feed the central model a steady stream of lies. They can hide a real disaster by making the numbers look normal, or they can create a fake disaster by making normal numbers look dangerous, all without the central system ever knowing which sensors are telling the truth.

Researchers have long known that standard computer models are fragile in this way. Most of these models look at the entire grid as one giant puzzle, trying to find patterns across hundreds of data points at once. The problem is that if an attacker controls just a few pieces of that puzzle, they can twist the whole picture to their liking. A new study from engineers at State Grid Jibei Information & Telecommunication Company and State Grid Jibei Electric Power Company Limited proposes a different way to look at the problem. Instead of trusting a single, all-seeing model, they built a system that treats every sensor with suspicion and checks its story against the stories of its neighbors. They call this system TALLY, and it is designed to answer a specific, critical question: how many sensors can an attacker control before the system stops being trustworthy?

The researchers built their solution around a simple but powerful idea: if you break a group of sensors into smaller, independent teams, a liar in one team cannot easily convince the others to lie. In their design, the system does not ask one giant model to judge every sensor. Instead, for each sensor, it creates a small panel of judges. Each judge looks at a specific, limited group of other sensors to predict what the target sensor should be doing. If a sensor is telling the truth, its reading will match what most of these judges expect. If it is lying, or if it is being forced to lie by an attacker, its reading will stand out from the crowd. The system then uses a voting mechanism. It ignores the most extreme predictions from the judges and focuses on the middle ground, where the honest sensors naturally cluster. This middle ground acts as a shield; even if an attacker controls a few sensors and tries to push the numbers up or down, the honest sensors in the middle will hold their ground, and the system will know that the outliers are the ones to watch.

To test this idea, the team ran their system on three very different real-world scenarios. The first was a hardware testbed that simulated a power plant with a boiler and a turbine, where sensors recorded data every ten seconds. The second was a dataset from a transmission line with relay sensors that monitor the flow of electricity. The third was a simulation of a neighborhood power grid, where smart meters tracked the actual electricity usage of real customers over time. In each case, they pitted their new system against ten other common detection methods, including standard artificial intelligence models that look at all the data at once. They then launched a series of sophisticated attacks, where they took control of a few sensors and tried to trick the systems into missing real attacks or raising false alarms.

The results showed a clear divide between the old way and the new way. The standard models, which try to see the whole grid at once, collapsed almost immediately when the researchers controlled just one or two sensors. On the power plant testbed, a single compromised sensor caused these standard models to raise false alarms nearly every time, effectively blinding the system. The new TALLY system, however, held its ground. When the researchers allowed the system to tolerate a specific number of compromised sensors, it continued to spot real problems while ignoring the lies. On the neighborhood grid simulation, the new system was actually the best detector of all, outperforming the standard models even without any attacks, because it was better at understanding how the meters in a neighborhood influence each other.

Perhaps the most significant finding was not just that the system worked, but that it could provide a mathematical guarantee for specific alarms. The researchers built a certificate into the design that states: if an alarm is certified, it is mathematically impossible for an attacker to have silenced it by controlling only that many sensors. However, the system cannot certify that no attack was hidden if no alarm sounds; if an event does not generate enough flagged nodes to meet the strict certification threshold, it remains uncertified, regardless of whether it was detected or not. In the tests, the new system successfully identified the sensors that were lying, flagging them for maintenance without ever blaming the honest sensors. Even when the attackers tried to hide a real surge in electricity demand by controlling a few meters, the system still raised the alarm, and the researchers could prove that no amount of lying by those few meters could have stopped it.

The study also revealed that the cost of this protection depends on the nature of the power system itself. In the power plant, where sensors often follow their own predictable patterns, adding more layers of protection required a slight trade-off in how quickly the system spotted minor issues. However, on the neighborhood grid, where sensors are highly dependent on each other, the new design was superior in every way, offering both better detection and stronger protection without any loss in performance. The researchers found that the system could be tuned like a dial; operators could decide how many sensors they were willing to tolerate as compromised and adjust the system accordingly. If they set the dial to tolerate four bad sensors, the system would ignore the noise from those four and focus on the rest, ensuring that the honest sensors were never falsely accused.

This work does not claim to solve every problem in grid security. The researchers were careful to note that their system assumes the sensors themselves are not corrupted before they send data, and it does not protect against an attacker who takes over the central computer itself. It also relies on the fact that the attackers cannot control more sensors than the system is designed to tolerate. But for the specific threat of a few rogue sensors feeding lies into a central model, the study offers a concrete, proven solution. It shifts the paradigm from trusting a single, fragile model to trusting a collective, resilient consensus. By breaking the grid into smaller, overlapping groups and letting the honest majority speak for itself, the system ensures that even if an attacker controls a few nodes, the truth of the grid remains visible. The findings suggest that for the future of smart grid security, the answer to "how many sensors can we trust?" is not "all of them," but rather "enough to form a majority," and that this simple shift in thinking can make the difference between a grid that is vulnerable to deception and one that is resilient to it.