2016-nasr-game
findings extracted from this paper
-
Internet connectivity is the primary determinant of RAD attack strength across nation-state censors: China (573 ASes, 858 ring ASes) achieves a censorship metric of 0.277 under profile T1, while Syria (4 ASes, 5 ring ASes) achieves only 0.101 with the same decoy budget. Venezuela, despite fewer total ASes than Saudi Arabia (44 vs. 107), achieves a higher censorship metric (0.210 vs. 0.197) owing to its disproportionately large ring AS count (835 vs. 176), confirming that ring AS count predicts RAD capability better than raw AS count.
-
Optimal RAD by a QoS-cautious wealthy Chinese censor (profile T1, F/ρ₀ = 5×10^6) forces 10.8% of routes onto non-valley-free (NVF) paths and 1.2% onto less-preferred routes, while still leaving 16.3% of routes traversing decoy ASes—zero routes become unreachable at this budget. The NVF and less-preferred-route fractions rise and then fall as decoy budget increases, as further RAD routing gains diminish past a crossover point.
-
The game-theoretic optimal decoy placement (ε-Nash equilibrium via best-response dynamics against an optimal RAD adversary) achieves a censorship metric of 0.2 at budget ratio F/ρ₀ = 10^8, versus 0.42 for the best prior heuristic ('sorted' placement from Houmansadr et al. [14]) under the same budget—a 2× improvement in censorship resistance per dollar. Prior comparisons used ad hoc RAD deployments rather than the optimal adversary, understating the benefit of principled placement.
-
Game-theoretic simulation shows that a QoS-cautious, wealthy Chinese censor (profile T1/T4) cannot reduce decoy-accessible routes below ~27% (censorship metric ≈ 0.277) via the RAD attack regardless of budget. An irrational censor can achieve a censorship metric of 1.000 but only by making 70.3% of all Internet routes unreachable to Chinese users—a collateral-damage threshold that constrains rational nation-state censors in practice.
-
In the autonomous (non-centrally-funded) deployment model, the decoy service fee γ (ratio of decoy revenue to transit revenue per MB) is the primary lever for censorship resistance: for China with profile T1, γ = 5 leaves 9.6% of routes usable for circumvention after optimal RAD, compared to 16.3% under the centrally-funded model at budget ratio F/ρ₀ = 5×10^6. Higher fees compensate ASes for RAD-induced transit revenue loss and sustain participation, but the autonomous model delivers roughly half the censorship resistance of a centrally-funded deployment at comparable incentive levels.