All shield benchmarks are open-source and reproducible from the repository. The v1 → v2 deltas below show that the application-layer hardenings (L1–L6) are not incremental; they reset the operating curve.
synthetic_radar
Synthetic radar with cross-receiver coordinated jamming. Metric: SNR improvement vs Wiener+median baseline.
v1: +0.11 dB (essentially baseline). v2: +4.5 dB — 40× improvement from L1 (continuous weights) + L2 (per-band) + L3 (joint-basket).
ctu13_botnet
CTU-13 botnet detection in network logs. Metric: caught / 3 known bots, with false-positive count.
v1: 1/3 caught + 3 FP. v2: 2/3 caught + 2 FP. The remaining 1/3 is the CTU-13 long-tail bot whose signature is below the cascade-internal floor.
texbat_gps ds4 / ds7
TEXBAT GPS-meaconing benchmark. Metric: caught / 2 meaconed datasets, with false-positive count on clean datasets.
v1: 0/2 caught + 2 FP. v2: 1/2 caught + 0 FP. The new joint-basket score (L3) eliminates the FPs.
Source: shield/benchmarks/