Everything you need to test model robustness
From a single API call or CSV upload to a scheduled scan pipeline — Ekushield covers the full adversarial testing workflow.
FGSM, PGD, and boundary attacks
Run gradient-based and black-box attacks against classifiers, regressors, and vision models. Sweep the perturbation budget (ε) and get an accuracy-vs-ε curve, not just a pass/fail.
Critical epsilon
Every scan reports the smallest perturbation budget that flips the model's decision on a majority of held-out inputs — a single, comparable number your team can track release over release.
Poisoning & provenance checks
Flag retraining pipelines that ingest unverified data sources, and estimate how much a small poisoned slice could move your decision boundary.
Prompt-injection & jailbreak probing
Run a curated, versioned suite of adversarial prompts against your system prompt and safety classifier, and track your bypass rate over time.
Confidence-oracle detection
Scan your deployed prediction API for response fields (raw probabilities, logits, verbose errors) that make gradient estimation easier for an attacker.
Exportable, audit-ready reports
Every scan is stored with a timestamp, dataset fingerprint, and methodology note — ready to hand to a security reviewer or auditor.
How a scan actually works
1. Bring your data
Upload a labeled CSV, connect a model endpoint, or answer a guided questionnaire about your deployment.
2. We attack it
Ekushield trains or wraps a comparable model and runs a gradient-sign perturbation sweep entirely inside your browser session or your own infrastructure — your raw data never has to leave your control.
3. You get a score
A clean-accuracy baseline, a critical-epsilon rating, and ranked hardening recommendations land in your dashboard in under two minutes.