Author: Aanavi Dalmia

1. Summary

This report completes the single-path mix model and extends the Bayesian attribution framework built around it into a statistically reproducible evaluation tool supporting multiple latency laws through a common implementation. A uniform mechanism draw and a continent-calibrated geographic draw, built from real network ping data, are evaluated as two configurations of the same framework rather than as separate implementations. The work was undertaken to determine whether the anonymity properties of the mix model depend on the specific latency law assumed, or whether they hold more generally once real network geography is imposed in place of an idealised mechanism.

The attribution attack was evaluated against both latency laws under a matched sender-link scale, and the stake-inference attack was evaluated across seven Pareto shape values, each read directly against that shape's own closed-form expected maximum stake. Two findings follow. First, geographic latency imposes a hard, closed-form ceiling on attribution accuracy — a consequence of the model's discreteness rather than of any particular parameter choice — and the measured sweep sits well below this ceiling throughout, confirming that real network geography leaks substantially less sender information than an idealised, freely-tunable mechanism. Second, stake concentration, not the cover mechanism alone, governs how identifiable the largest staker is, with the effect consistent across every Pareto shape tested.

Together these results validate both halves of the single-path mix model against closed-form theory rather than simulation alone, and they validate the underlying evaluation framework itself — its multi-latency-law design and its statistically reproducible methodology — as the baseline against which the next stage of this work, an implementation of the Dandelion++ propagation protocol, will be compared.

2. Parameters

The experiments below share the following configuration unless stated otherwise.

Parameter Symbol Value
Nodes N 200 (attribution), 1000 (stake inference)
Mix hops k 3
Sender set size (fixed cover) $ S_i
Sender hold (mean) $1/λ_S$ 0.25 s
Mix delay (mean, swept) $1/λ_M$ [1.00,0.300,0.100,0.0300,0.0100]s
Uniform sender-link range $U(0,d_{\max}⁡)$ $d_{\max}⁡=0.25 ⇒σ_d≈0.0722$
Geographic first-mix continent $c^∗$ Europe (pinned)
Geographic sender-link spread $σ_d$ ≈0.0303 s
Geographic block-relay delay $δproc$ 0.20 s
Attribution repetitions — 50
Stake-inference epoch length T 20,000 and 4000 for attribution
Stake-inference repetitions — 60
Pareto shapes tested (stake) kstake 4/3
Cover probabilities tested (stake) pS [0.001,0.003,0.01,0.03,0.1,0.3,0.5,0.7]

The uniform sender-link range and the geographic continent pinning are chosen so that the two models' sender-link spreads σdσd sit within the same order of magnitude, so that the comparison drawn in Section 6 reflects the latency law itself rather than an incidental difference of scale.

3. Methodology

3.1 AC-path structure

The uniform and geographic latency laws are implemented as alternative configurations of a single Bayesian attribution framework, sharing the same AC-path structure, likelihood, and posterior computation; only the deterministic link-latency draw differs between them. Both configurations route every genuine message down a single shared path before broadcast:

image.png

Each candidate sender accrues a deterministic latency

image.png

plus an independent random hold.

image.png

The two configurations differ only in how the deterministic component is drawn; the attack, its likelihood, and every downstream measure are shared, unmodified code.

Uniform law. $d_i^S,d_r^R∼U(0,d_{\max}⁡)$, drawn independently per node; $D_M=(k−1)d$, a scalar representing homogeneous middle links.

Geographic law. Every link is a real, continent-calibrated one-way ping,