Author: Aanavi

Summary

Built the Bayesian sender-attribution attack alongside the stake-inference attack, and built the three measures that grade it: Top-1 de-anonymisation, mean true-sender posterior, and posterior entropy. Both attacks now run through the same simulation pipeline. Stake inference estimates how much stake each node holds from participation frequency over an epoch; attribution estimates which candidate sender sent one specific observed broadcast, from message timing.

Equations

Deterministic delay: Every candidate's link latency is known to the adversary:

image.png

For candidate sender i, observed receiver r: sender-to-mixer link, mix-chain total, exit-to-receiver link.

Random hold: Unobserved, comprising the sender's own queuing delay plus k mixing holds:

image.png

Observed delay:  y=t+μ(L,r)+Z, where L (the true sender) is unknown to the adversary.

Residual:  For each candidate i∈St

image.png

Negative residuals are inadmissible and eliminated.

Posterior:  Residuals scored against the density of Z (hypoexponential — a sum of k+1 exponentials, generally unequal rates), uniform prior:

image.png

MAP estimate:

image.png

Signal-to-noise ratio:

image.png

At this project's operating point (k=3, mean mixing delay $1/λ_M=1s,$ mean sender hold $1/λS=0.25 s$):