Confidence Boosts Trust-Based Resilience in Cooperative Multi-Robot Systems

2 min read

Ballotta and colleagues make consensus resilient by using “trust observations” from the wireless channel, a value in [0, 1] per message that says how likely the sender is legitimate, so detection is decoupled from the task data. Each legitimate robot updates its state as a mix of its own initial value, weighted by a confidence parameter λ_t, and a trust-weighted average of its neighbours. λ_t decays to zero, so robots lean on their own starting point while they are still unsure whom to trust and hand over to their neighbours as evidence accumulates. Unlike earlier work, robots do not have to wait through a fixed observation window before cooperating.

Findings

  • The protocol reaches consensus with arbitrarily many malicious robots, given that the legitimate robots’ graph is connected (a primitive weight matrix), trust observations are informative on average (mean above 1/2 for legitimate senders, below 1/2 for malicious ones), states are bounded, and λ_t goes to zero.
  • They bound the deviation from the attack-free consensus and the expected convergence rate. Fast decay of λ_t gives quick agreement on a value polluted by early misclassifications. Slow decay gives near-nominal agreement after a long time.
  • Simulation: two platoons of five vehicles merge while 3 malicious vehicles broadcast a constant speed, on a 2-nearest-neighbour topology. Trust values came from Beta(1.5, 1) for legitimate and Beta(0.75, 1) for malicious senders, both close to 1/2 on average.
  • Against the earlier observation-window method with a window of 50 steps set a little short of the unknown time when classification becomes correct, the new protocol with λ_t = e^(−0.05t) had a much smaller deviation without slowing convergence much.

Limits

  • Needs a physical-layer trust signal that the attacker cannot fake. This targets spoofed or impersonated robots. A legitimate robot with a broken sensor passes the channel test and is fully trusted (my reading).
  • Scalar consensus in simulation only.
  • Legitimate robots must stay connected among themselves. With sparser graphs one malicious vehicle could split the platoon (authors).

Relevance

Builds on the channel-based trust idea in gil-2017-spoof-resilient. The method catches the lying robot and ignores the broken one, a gap discussed in the resilient consensus and Byzantine faults notes. See also robot security.