A new Wi-Fi based system called Argus can identify specific people just from how their bodies bend ordinary wireless signals, no camera or phone required.
Researchers built Argus, a passive Wi-Fi sensing system that identifies people using Channel State Information (CSI), the routine signal-quality data any Wi-Fi router already collects. Instead of tracking gait or motion like earlier wireless-ID systems, Argus turns short slices of CSI into compact statistical snapshots called statgrams, then feeds them to a lightweight Transformer model that reads patches of that data the way other Transformers read words. On a 154-person dataset with a strict test split, it hit 78.88% top-1 accuracy from just 6 seconds of data, climbing to 84.85% after stitching together 19 overlapping windows over a full minute. Top-5 accuracy reached 99.26%, meaning the correct match was almost always in Argus's shortlist even when its single best guess was wrong.
The bigger claim here is efficiency, not just accuracy. Compared with a raw-CSI Transformer baseline, Argus needed 4.4 times fewer FLOPs per window for a 7.75-point accuracy gain, and on a separate multi-user benchmark called WiMANS it matched specialized per-room models within 1.23 percentage points while using 27 times less compute. That kind of efficiency is what would let device-free identification run on cheap, low-power hardware instead of a server rack.
The authors are upfront that open-set rejection, telling a stranger from anyone in the training set, and moving Argus to a new room remain unsolved problems, which is exactly the gap between a promising benchmark and something you'd want tracking who walks into a building.