Model risks

AI Security and Risk Management

Angeline Corvaglia

Founder and Digital Transformation Expert

Monitoring risks over the lifecycle

representation of AI lifecycle

During development

  • Fair and accurate data
  • Learns effectively
  • Harms of deploying flawed system

dot

After deployment

  • Models can change (drift) over time
  • Lead to biased or inaccurate outputs
AI Security and Risk Management

Seven common model risks

  chess board to represent model risks

 

  • Lack of model transparency and interpretability
  • Bias
  • Hallucination
  • Model drift
  • Overfitting
  • Underfitting
  • Data leakage
AI Security and Risk Management

  image representing transparency

 

Transparency

  • How AI models make decisions

Interpretability

  • Why a model produces certain outputs

exclamation point

In sensitive areas like healthcare or finance

  • Decision-making process
  • As important as the decision itself
AI Security and Risk Management

Bias

Unfair preferences or prejudices

  • Biased training data
  • Algorithms
  • Unrepresentative or flawed data
  • Human assumptions

Detected and mitigated for fair and ethical AI systems

AI bias

AI Security and Risk Management

Hallucination

icon representing hallucination

Occur for many reasons:

  • Insufficient training data
  • Incorrect assumptions
  • Biases in the data
AI Security and Risk Management

Model drift

Previously accurate

  • Lose their relevance
  • Accuracy

Model get confused:

  • Different data
  • Different patterns within the data

representation of model drift

AI Security and Risk Management

Effective and accurate learning

representation of learning

Overfitting

  • Learns too many details

light bulb

Underfitting

  • Too simple to learn

light bulb

Data leakage

  • Outside data included in training
AI Security and Risk Management

Explainable AI (XAI)

 

  • Methods and techniques
  • Decisions understandable to humans
  • Explanations for the reasoning behind a model's output

explainable AI representation

AI Security and Risk Management

More robust, fair, and secure AI systems

 

  • Security measures
  • Responsible development
  • Long-term impact

arrow

Should be beneficial for everyone

  representation of robust, fair & secure AI

AI Security and Risk Management

Let's practice!

AI Security and Risk Management

Preparing Video For Download...