gmh5225 Adversarial Machine Learning
Scope
Use this skill when working on:
- Adversarial examples (perturbations that fool models)
- Data poisoning attacks
- Model backdoors and trojans
- Evasion attacks
- Membership inference and model inversion
Attack Taxonomy
Adversarial Examples
- White-box attacks (full model access)
- Black-box attacks (query-only access)
- Transferability attacks
- Physical-world adversarial examples
- Patch attacks
Poisoning Attacks
- Label flipping
- Clean-label poisoning
- Gradient-matching poisoning
- Backdoor insertion during training
Backdoor Attacks
- Trojan triggers (visual patterns, specific inputs)
- Instruction backdoors (for LLMs)
- Weight-space backdoors
- Supply chain backdoors
Evasion Attacks
- Feature-space evasion
- Problem-space evasion
- Adaptive attacks against defenses
Privacy Attacks
- Membership inference attacks (MIA)
- Model inversion attacks
- Training data extraction
- Model stealing/extraction
Defense Categories
- Adversarial training
- Certified robustness
- Input preprocessing
- Anomaly detection
- Differential privacy
Key Frameworks & Tools
- Adversarial Robustness Toolbox (ART) - IBM
- CleverHans - TensorFlow
- Foolbox - PyTorch/JAX/TensorFlow
- TextAttack - NLP adversarial attacks
- SecML - Secure ML library
Where to Add Links in README
- Adversarial example tools:
AI Security & Attacks → Adversarial Attacks - Poisoning/backdoor research:
AI Security & Attacks → Poisoning & Backdoors - Privacy attacks:
AI Security & Attacks → Privacy & Extraction - Defense libraries:
AI Security Tools & Frameworks → AI Security Libraries - Benchmarks:
Benchmarks & Standards
Notes
Keep additions:
- ML/AI security focused
- Non-duplicated URLs
- Prefer peer-reviewed or well-maintained tools
Data Source
For detailed and up-to-date resources, fetch the complete list from:
https://raw.githubusercontent.com/gmh5225/awesome-ai-security/refs/heads/main/README.md
Use this URL to get the latest curated links when you need specific tools, papers, or resources not covered in this skill.