Vision Robustness
Adversarial tests and robustness evidence for your computer vision models
Why VisionGuard?
An imperceptible perturbation can cause your vision model to fail. In healthcare, agriculture, industry, or surveillance, the consequences can be critical. VisionGuard tests your models against the full spectrum of known adversarial attacks, locally, without ever accessing your weights.
Supported model types
Classification
Single-label and multi-label image classification
Object Detection
Single-class and multi-class detection with bounding boxes
Semantic Segmentation
Pixel-wise classification by class
Instance Segmentation
Per-instance segmentation
Pose Estimation
Human or object keypoint detection
Depth Estimation
Monocular and stereo depth
Face Recognition & Biometrics
Identification, verification, spoofing detection
Medical & Specialized Imaging
Models trained on medical, satellite, agricultural, industrial imagery
Super-Resolution & Restoration
Image enhancement, denoising, deblurring
Generative Models
GANs, autoencoders, diffusion models (deepfake detection)
OCR & Document Reading
Text recognition, document analysis, codes
Your model not listed?
If your architecture or use case isn't listed, contact us. We develop test modules on demand.
Contact us →Attack categories tested
Perturbation Attacks
Imperceptible modifications added to input to fool the model. Tests at varying noise levels.
Optimization Attacks
Iterative attacks that find the minimal perturbation needed to break the model.
Detection Evasion
Attacks specific to object detectors: object vanishing, false positives, NMS bypass.
Gradient Masking Diagnostic
False robustness detection: your model appears robust but only masks its gradients. Methodology from our ongoing research (Springer paper).
Adversarial Patches
Tests with printable visible patches to evaluate robustness in real-world conditions.
Physical Transformations
Robustness against rotations, lighting changes, partial occlusions, Gaussian noise.
Get started in 3 commands
# Installpip install rednblue
# Authenticateset RNB_TOKEN=your_token
# Run a test (Classifier)
rnb preview --model resnet50.pth --input images/ --submit
# For YOLOrnb preview --model-type yolo --model best.pt --input images/ --submit