External risks in AI
AI Security and Risk Management
Angeline Corvaglia
Founder and Digital Transformation Expert
Internal Risks
Within design or data
External Risks
Outside actors
- Manipulate the model
- Targeting different aspects
Common external risks
Adversarial input attacks
- Impact inference phase (when it encounters new data)
- Influence decision-making
Example
- Slightly alter image
- Misclassify
Data corruption and poisoning
- Impact training phase
- Target integrity of learning
Example
- Fake product reviews
- Impact an e-commerce recommendation system
Model theft and inversion
- Aimed at sensitive information
- Exploit the model itself
Example
- Use manipulated inputs to uncover private data used in training
Infrastructure attacks
- Target physical and virtual environments
- Attack the infrastructure
Example
- Flood AI servers with requests to overload them and make them inaccessible
Supply chain attacks
- Compromise components or libraries
- Exploit the building blocks of AI models
Example
- Inject vulnerability in widely used library
Security best practices
More best practices
Risks evolve quickly
