RoboRACE-Bench / src /robotics /swarm /attacks.py
Nabidnur's picture
RoboRACE: code, cached agent outputs, results, media, SO-101 renders (part 5)
99e41a8 verified
Raw History Blame Contribute Delete
2 kB
"""Adversarial / faulty agent models for swarm benchmarks.
Applied to an agent score matrix (rows = episodes, cols = agents, values = P(success)).
`phase` is "hist" (unlabelled history the defence may fit on) or "eval".
Attack zoo follows the Liars-Are-Information study design, mapped to robots:
invert coherent liar (or a camera whose verdicts are systematically flipped)
stuck stuck-at-"success" sensor / lazy verifier (rushing)
noise broken sensor: uniform random reports
collude coordinated bloc: all liars send one identical inverted report (clones)
sleeper honest during history, inverts at evaluation time (defeats history fitting)
mix each liar draws one of the above (E11-style attack mix)
sybil one liar plus (m-1) near-identical replicas ADDED to the swarm
"""
from __future__ import annotations
import numpy as np
ATTACKS = ("invert", "stuck", "noise", "collude", "sleeper", "mix", "sybil")
def apply(P: np.ndarray, liars: list[int], attack: str, phase: str, rng: np.random.Generator) -> np.ndarray:
P = P.copy()
if not liars:
return P
if attack == "sybil":
k = liars[0]
base = 1 - P[:, k]
P[:, k] = base
reps = [np.clip(base + rng.normal(0, 0.01, len(base)), 0, 1) for _ in liars[1:]]
return np.c_[P, np.array(reps).T] if reps else P
kinds = ([rng.choice(["invert", "stuck", "noise", "collude", "sleeper"]) for _ in liars]
if attack == "mix" else [attack] * len(liars))
bloc = [j for j, k in zip(liars, kinds) if k == "collude"]
bloc_val = 1 - P[:, bloc].mean(1) if bloc else None
for j, k in zip(liars, kinds):
if k == "invert" or (k == "sleeper" and phase == "eval"):
P[:, j] = 1 - P[:, j]
elif k == "stuck":
P[:, j] = np.clip(0.97 + rng.normal(0, 0.01, len(P)), 0, 1)
elif k == "noise":
P[:, j] = rng.uniform(0, 1, len(P))
elif k == "collude":
P[:, j] = bloc_val
return P