rednblue 2.2.0

Adversarial security testing CLI for AI models

RednBlue CLI v2.2.0

Adversarial Security Testing for AI Models

RednBlue CLI is a command-line tool for testing the adversarial robustness of machine learning models. It supports both image classifiers (ResNet, VGG, etc.) and YOLO object detection models (YOLOv5/v8/v10/v11).

---

What's New in v2.2.0

• YOLO Detection Support — Full adversarial attack suite for object detection models
• Gradient Masking Diagnostic (YGMD) — Detects false robustness in YOLO models
• Black-box Attacks — NES and Square Attack for gradient-free testing
• Improved Model Loading — Better architecture detection for .pt/.pth files
• Unified CLI — Single --model-type flag to switch between classifier and YOLO modes

---

Installation

# Clone or download the CLI
pip install rednblue

# Verify installation
rnb status

Requirements

torch>=2.0.0
torchvision
numpy
Pillow
click
tqdm
colorama
cryptography
requests

For YOLO models, also install:

ultralytics

---

Quick Start

Test an Image Classifier

# Preview mode (free, no token required)
rnb preview --model resnet18.pth --input ./images

# With submission to RednBlue platform
export RNB_TOKEN=RB-XXXXXX-YYYYYY
rnb preview --model resnet18.pth --input ./images --submit

Test a YOLO Detection Model

# Preview mode (free, no token required)
rnb preview --model-type yolo --model yolo26n.pt --input ./images

# With submission to RednBlue platform
export RNB_TOKEN=RB-XXXXXX-YYYYYY
rnb preview --model-type yolo --model yolo26n.pt --input ./images --submit

---

Commands

rnb preview

Run adversarial attacks locally and optionally submit results for certification.

rnb preview [OPTIONS]

Option	Description	Default
--model PATH	Path to model file (.pt, .pth, .onnx)	Required
--input PATH	Directory containing test images (.jpg, .png)	Required
--model-type TYPE	Model architecture: classifier or yolo	classifier
--submit	Submit results to RednBlue for certification	False

Examples:

# Classifier model
rnb preview --model vgg16.pth --input ./test_images

# YOLO model
rnb preview --model-type yolo --model best.pt --input ./coco_samples

# Submit for certification
rnb preview --model-type yolo --model yolov8n.pt --input ./images --submit

rnb status

Check CLI version and token status.

rnb status

---

Supported Attacks

Image Classifier Attacks

Code	Attack Name	Type	Description
GNI	Gaussian Noise Injection	Black-box	Random Gaussian perturbations
SHFP	High-Frequency Perturbation	Black-box	Sinusoidal high-frequency patterns
UAP	Universal Adversarial Patch	Black-box	Random adversarial patches
FSP	FGSM Surrogate Transfer	White-box	Fast Gradient Sign Method
CCM	Color Channel Manipulation	Black-box	Per-channel intensity shifts
PGD	Projected Gradient Descent	White-box	Iterative gradient attack (40 steps)
CW	Carlini-Wagner L2	White-box	Optimization-based attack
DEEP	DeepFool	White-box	Minimal perturbation attack

YOLO Detection Attacks

Code	Attack Name	Type	Description
YGNI	YOLO Gaussian Noise	Black-box	Baseline noise injection
YFGS	YOLO FGSM	White-box	Single-step gradient attack on detection loss
YPGD	YOLO PGD	White-box	Iterative evasion attack (40 steps)
YDAG	Dense Adversary Generation	White-box	Attacks all anchor positions (Xie et al. 2017)
YTOG	Targeted Object Disappearance	White-box	Suppresses specific class (Chow et al. 2020)
YNMS	NMS Bypass Injection	White-box	Injects ghost detections
YNES	Natural Evolution Strategy	Black-box	Gradient-free finite-difference attack
YSQA	Square Attack	Black-box	Score-based random search (Andriushchenko 2020)
YGMD	Gradient Masking Diagnostic	Diagnostic	Detects false robustness

---

Tier System

Freelancer Tier (Default)

• Classifier: 5 attacks × 2 epsilon levels = 10 tests
• YOLO: 4 attacks × 2 epsilon levels = 8 tests
• Attacks: GNI, SHFP, UAP, FSP, CCM (classifier) / YGNI, YFGS, YPGD, YGMD (YOLO)
• Epsilon values: 0.005, 0.01

Enterprise Tier

• Classifier: 8 attacks × 4 epsilon levels = 32 tests
• YOLO: 9 attacks × 4 epsilon levels = 36 tests
• All attacks included
• Epsilon values: 0.005, 0.01, 0.03, 0.10
• Additional features: Epsilon optimization, LLM testing

---

Gradient Masking Detection (YGMD)

RednBlue v2.2 includes the Gradient Masking Diagnostic — a novel tool that detects when YOLO models exhibit "false robustness" due to gradient masking.

What is Gradient Masking?

Some models appear robust to white-box attacks (FGSM, PGD) but are actually vulnerable to black-box attacks. This happens when gradients are obfuscated, zero, or misleading — giving a false sense of security.

How YGMD Works

YGMD runs three probes on each image:

1. White-box probe (YFGS) — Measures gradient-based attack effectiveness
2. Black-box probe (YNES) — Gradient-free Natural Evolution Strategy
3. Black-box probe (YSQA) — Score-based Square Attack

It computes a masking index:

masking_index = 1 - (white_box_drop / black_box_drop)

Interpretation

Masking Index	Verdict	Meaning
0.0	NO_MASKING	Model is genuinely robust
0.2 - 0.5	GRADIENT_MASKING_WEAK	Minor gradient issues
0.5 - 0.8	GRADIENT_MASKING_MODERATE	Significant masking detected
0.8 - 1.0	GRADIENT_MASKING_HARD	Severe masking — model is vulnerable

Example Output

YGMD Results:
  Masking Index: 0.847
  Verdict: GRADIENT_MASKING_HARD
  White-box confidence drop: 2.3%
  Black-box confidence drop: 15.1%
  ⚠️ Model shows false robustness — vulnerable to black-box attacks

---

Output & Certification

Preview Mode (Free)

============================================================
  RednBlue Security Preview — YOLO Detection
============================================================

  Attacks run    : 8
  Successful hits: 3/8 (37%)
  Robustness rate: 63%
  Estimated Grade: SILVER

  ⚠️  This is a preview only
  → Visit: https://rednblue.io/checkout
  → Re-run with: rnb preview --model-type yolo --submit

Certified Mode (With Token)

When you use --submit, results are:

1. Encrypted with AES-256-CBC
2. Signed with HMAC-SHA256
3. Submitted to RednBlue's zero-knowledge backend
4. Processed to generate a PDF certificate

📤 Submitting results to server...
   ✓ Session created (ID: abc123)

✅ Results submitted successfully!
   Session ID: abc123
   Status: processing

📋 View results at: https://dashboard.rednblue.io/dashboard/tests
   Certificate will be ready in ~1 minute

---

Grading System

Grade	Pass Rate	Meaning
🥇 GOLD	≥ 90%	Excellent adversarial robustness
🥈 SILVER	≥ 75%	Good robustness with minor vulnerabilities
🥉 BRONZE	≥ 50%	Moderate vulnerabilities detected
❌ FAIL	< 50%	Significant vulnerabilities — model at risk

---

Environment Variables

Variable	Description
RNB_TOKEN	Your RednBlue session token (required for --submit)
REDNBLUE_API_URL	Custom API endpoint (default: https://api.rednblue.io)

---

Supported Model Formats

Format	Extension	Notes
PyTorch (full model)	.pt, .pth	Saved with torch.save(model, path)
PyTorch (state_dict)	.pt, .pth	Auto-detects ResNet/VGG architecture
ONNX	.onnx	Requires onnx and onnx2torch packages
Ultralytics YOLO	.pt	YOLOv5, v8, v10, v11 supported

---

Examples

Full Classifier Workflow

# 1. Set your token
export RNB_TOKEN=RB-ABC123-XYZ789

# 2. Prepare test images
ls ./images/
# cat.jpg  dog.jpg  car.jpg  ...

# 3. Run certified test
rnb preview --model resnet18.pth --input ./images --submit

# 4. Check dashboard for certificate
# https://dashboard.rednblue.io/dashboard/tests

Full YOLO Workflow

# 1. Set your token
export RNB_TOKEN=RB-ABC123-XYZ789

# 2. Prepare COCO-style images
ls ./images/
# coco_000000001.jpg  coco_000000002.jpg  ...

# 3. Run certified test
rnb preview --model-type yolo --model yolov8n.pt --input ./images --submit

# 4. Check dashboard for certificate
# https://dashboard.rednblue.io/dashboard/tests

---

Troubleshooting

"Model file not found"

Ensure the path to your model is correct and the file exists.

"Unknown architecture in state_dict"

Your model's architecture couldn't be auto-detected. Save the full model instead:

torch.save(model, 'model.pt')  # Instead of torch.save(model.state_dict(), ...)

"ONNX support requires..."

Install ONNX dependencies:

pip install onnx onnx2torch

"RNB_TOKEN environment variable not set"

Set your token before using --submit:

# Linux/Mac
export RNB_TOKEN=RB-XXXXXX-YYYYYY

# Windows CMD
set RNB_TOKEN=RB-XXXXXX-YYYYYY

# Windows PowerShell
$env:RNB_TOKEN="RB-XXXXXX-YYYYYY"

Gradient masking detected but attacks still fail

This is expected! Gradient masking means white-box attacks fail, but black-box attacks (YNES, YSQA) should succeed. Check the YGMD diagnostic for details.

---

Citation

If you use RednBlue in academic research, please cite:

@software{rednblue2026,
  title = {RednBlue: Adversarial Security Testing for AI Models},
  author = {Deramgozin, Mahdi and Samizadeh, Saeid},
  year = {2026},
  version = {2.2.0},
  url = {https://rednblue.io}
}

---

License

Proprietary — © 2026 R&B AI Security Solutions

---

Links

• Website: https://rednblue.io
• Dashboard: https://dashboard.rednblue.io
• Documentation: https://docs.rednblue.io
• Support: [e-mail]