Guaranteeing spoof-resilient multi-robot networks

3 min read

Gil, Kumar and colleagues defend a multi-robot coverage team against a Sybil attack, where one malicious client forges many identities with fake positions to pull the servers away from real clients. Instead of cryptographic keys, each quadrotor server rotates a commodity Wi-Fi card to emulate an antenna array (synthetic aperture radar) and builds a “spatial fingerprint” of each client: received power along each 3D direction. Clients whose fingerprints are too similar are probably the same radio. This gives each client a confidence weight , with a proof that its expectation is near 1 for legitimate clients and near 0 for spoofed ones. The weights drop straight into a coverage controller as per-client importance, and the paper bounds how far the servers can end up from their attack-free positions.

Findings

Numbers are from the RSS 2015 version, which is the copy I read.

  • Two AscTec Hummingbird servers and ten iRobot Create clients, indoors. Spoofed identities reach up to 75% of the legitimate client count in the detection runs.
  • Classification against an RSSI baseline: TPR 96.3% and FPR 3.0% for static spoofers (RSSI: 81.5% and 9.1%); 96.3% and 6.1% for mobile ones; 100% and 3.0% under per-packet power scaling, where RSSI collapses to 74.1% and 27.3%.
  • Multipath helps: two co-aligned clients are separable at 0° offset indoors, against 3° in an anechoic chamber. Free space is the hard case.
  • Coverage, 12 runs in an 8 m by 10 m room: with no defence the servers end up on average 3.77 m from the oracle solution; with weights, 0.02 m. Cost stays near optimal even with spoofers at 300% of the network.

Limits

  • Only clients are adversarial. Adversarial servers are left to future work, and each adversary is assumed to have one antenna.
  • The defence detects identity forgery. A single real robot that reports a false position under its own identity is not caught, since its fingerprint is genuine.
  • The theory leans on assumptions about signal noise that hold best in multipath-rich rooms. The authors flag free space as the challenging setting, which matters outdoors, at sea and in the air.
  • Fingerprints need the server to move its antenna to collect SAR snapshots.

Relevance

Gil et al. use physics as an independent witness against what a robot claims, a robot-team version of the orthogonal monitoring idea in cyber-physical security. The soft weights are a cousin of the trust values used in resilient consensus, and the gap it leaves (a liar with a real identity) is the Byzantine case. See robot security and the blocklist protocol.