running-adversarial-perturbation-suite

What problem does it solve?

This Skill measures how brittle a trained vision or tabular model is to small, adversarially chosen input changes, so you can report a defensible robustness number instead of relying on clean accuracy alone.

Core Features & Use Cases

• Threat-model first evaluation: Requires a declared norm, epsilon with units, attacker access level, and targeted or untargeted goal before any attack runs.
• Canonical attack stack: Runs FGSM, PGD-20, and AutoAttack-standard in increasing strength to produce comparable empirical robustness results.
• Tabular safety checks: Applies feasibility constraints for categorical, integer, monotone, and other domain-limited features so adversarial examples stay realistic.
• Reporting and auditability: Produces clean accuracy on the attacked subset, per-attack robust accuracy, monotone sanity checks, and saved adversarial examples for inspection.

Quick Start

Ask Claude to evaluate your trained model under a declared threat model, run FGSM, PGD-20, and AutoAttack-standard on a held-out subset, and produce the robustness report with saved adversarial examples.