First-hand information for everyone
Scientists have shown that examining how many triangles a node participates in—its local clustering coefficient—lets them spot malicious bots in networks that normally follow geometric rules. In the studied model, ordinary nodes are placed randomly on a high‑dimensional torus and connected if they are close. Bots ignore the geometry and connect independently with the same overall edge probability.
A node’s degree alone cannot reveal a bot, because its degree distribution is the same under both scenarios. However, the expected clustering of a pure geometric node is a known constant C_d that depends only on the geometry’s dimension. Bots, with random edges, have a much smaller expected clustering p_n. The researchers proved that if a node has at least half the expected degree and a clustering below a threshold τ (e.g., τ=C_d/2), then with high probability the node is a bot.
The test scans every node and rejects the null hypothesis when a node meets these degree‑and‑clustering criteria. The authors derived a bound showing that for any set of k_n bot nodes, as long as the average degree μ_n grows faster than log n, the test recovers all bots with probability tending to one—uniformly over all possible bot labels.
The procedure’s guarantees rely on knowing the dimension‑dependent constant C_d and the edge probability p_n. If either value is misspecified, the scan rule may fail to localize the botnet. The paper suggests that consistent estimators of C_d from the graph itself can be used, but establishing finite‑sample calibration remains an open problem.
These results provide a principled way to separate geometry‑driven connectivity from random bot behaviour using only local network statistics, and highlight the importance of accurate parameter estimation for reliable detection.