NEW Object-detection attacks on Faster R-CNN & YOLOv5

Attack your models before someone else does.

ptAI runs real adversarial attacks — evasion, privacy, model theft, poisoning and object detection — against your own models and data, and turns them into evidence: metrics, adversarial datasets and a pass/fail verdict you can put in a release gate.

Open the dashboard Try the live demo Explore the API

Checking API status… IBM ART 1.20.1 REST API · Swagger · Postman

ptai console

Overview

Attacks · 30 days0

Models under test0

Failing latest verdict3 / 9

Median run time41s

Attack activity Live

Verdicts

Vulnerable 48%At risk 21%Robust 25%Run failed 6%

IDAttackModelResult

od_a2f4d0PGD · object detectionYOLOv5smAP −0.75

j_9b9592PGD · white-boxIDS · Keras−24.0 pp

j_2c44a8CopycatCNNMNIST CNNfidelity 84%

j_d651ceMembership inferenceIDS · KerasAUC 0.51

21adversarial attack algorithms

6attack families, from evasion to backdoors

7model frameworks you can upload

1^APIone REST contract for UI, scripts and CI

IBM ARTTensorFlow · KerasPyTorchscikit-learnXGBoostLightGBMCatBoostONNX

Capabilities

Every way a model gets attacked — in one lab.

Choose the attacker's access — white-, grey- or black-box — and ptAI runs the matching IBM ART attacks, scores the damage and explains it.

Evasion

Imperceptible perturbations that flip predictions — with gradients, through a surrogate, or from decisions alone.

FGSMPGDBIMJSMAC&WDeepFoolAPGDZOOBoundaryHopSkipJump

Membership inference

Does the model leak who was in its training set? Rule-based and black-box attacks with AUC and attacker advantage.

Rule-based MIABlack-box MIA

Model extraction

Steal a working copy from predictions alone and measure how faithfully it reproduces your model.

CopycatCNNfidelityquery budget

Poisoning & backdoors

Inject a few malicious training samples, retrain a clone and measure the attack success rate of the hidden trigger.

BadNetsClean-labelHidden trigger

Robustness pentest

An ε-sweep across FGSM, PGD and BIM with thresholds and a pass/fail verdict — the release-gate view.

ε-sweepverdictSTIX report

Object detection NEW

Make objects vanish, appear or change class on Faster R-CNN and YOLOv5 — with mAP drop, boxes and adversarial patches.

FGSMPGDDPatchRobustDPatchAdv. Patch

Bring your own model

Upload the model you ship — not a demo.

Tabular or image, deep learning or trees. Untrusted formats are only ever loaded inside an isolated, network-less sandbox.

Gradient attacks where the framework allows them, transfer attacks where it doesn't
Saved test targets: model, dataset and preprocessing reused across runs
Compatibility checks before you launch — not a failed job afterwards

TensorFlow · Keras.h5 · .keras

PyTorch.pt · .pth · .safetensors

scikit-learn.joblib · .pkl · .skops

XGBoost · LightGBM.json · .ubj · .txt

CatBoost.cbm

ONNX.onnx · black-box

How it works

From upload to verdict in three steps.

Choose a target

Upload a model and data — or pick a saved target, a built-in model or the COCO detection sample.

Pick the attack

Family, access level and algorithm, with compatibility checked against your model and parameters explained.

Get the evidence

Live progress, metrics and a verdict, adversarial datasets, side-by-side images and a result ZIP — kept in Reports.

Developers

The same engine, as an API.

Everything the dashboard does is one HTTP call away — authenticated with an API key, run as a background job, delivered as a ZIP or JSON. Drop it into CI.

Swagger UI Try every endpoint from the browser. Postman collection One-click import, every attack pre-filled.sign-in required · raw file at /postman API guide Endpoints, parameters, jobs and errors. Interactive demo MNIST and YOLOv5 under attack — explore real results, no sign-in.

# White-box PGD on a tabular model → result ZIP
curl -X POST https://api.pt4ai.k3ylabs.dev/api/attack/evasion/whitebox \
  -H "X-API-Key: $PTAI_KEY" \
  -F data_type=structured -F attack=pgd \
  -F 'params={"eps":0.1,"eps_step":0.01,"max_iter":40}' \
  -F dataset=@test.csv -F model=@model.h5 \
  -F scaler=@scaler.joblib -F encoder=@encoder.joblib \
  -o result.zip

import requests, time

API = "https://api.pt4ai.k3ylabs.dev"
H = {"X-API-Key": "..."}
files = {k: open(f, "rb") for k, f in {"dataset": "test.csv", "model": "model.h5",
         "scaler": "scaler.joblib", "encoder": "encoder.joblib"}.items()}
job = requests.post(f"{API}/api/attack/evasion/blackbox?async=true", headers=H, files=files,
                    data={"data_type": "structured", "attack": "hsj"}).json()
while requests.get(API + job["status_url"], headers=H).json()["status"] != "done":
    time.sleep(5)
open("result.zip", "wb").write(requests.get(API + job["result_url"], headers=H).content)

# PGD on YOLOv5s against two COCO samples → boxes, mAP drop, datasets
curl -X POST https://api.pt4ai.k3ylabs.dev/api/attack/object-detection \
  -H "X-API-Key: $PTAI_KEY" \
  -F model=yolov5s -F attack=pgd \
  -F 'params={"eps":8,"eps_step":2,"max_iter":10}' \
  -F images=000000206487.jpg,000000118209.jpg \
  -o od_result.zip

# models, attacks and sample images
curl https://api.pt4ai.k3ylabs.dev/api/object-detection/catalog -H "X-API-Key: $PTAI_KEY"

Start testing

Find out how your model breaks — today.

Open the dashboard Try the live demo