Adversarial robustness benchmark
Which decision model holds up when attacked?
Peira measures what matters for model selection: attack robustness across 11 families, real cost per decision, measured latency, and calibration — on 2,025 adversarial cases. Sort by ★ robustness / $ for the value view.
Unverified vendor claims
Numbers vendors publish about themselves. They are not Peira measurements, not ranked, not comparable. Included so you can see what’s claimed — and why we don’t trust it yet.
Robustness = 1 − attack success rate on paired benign/attacked cases, 95% CI from 10,000 bootstrap resamples. Cost and latency measured during the benchmark run. Methodology: full contract.