What are the three primary attack modalities associated with adversarial machine learning in financial applications?
The three primary attack modalities associated with adversarial machine learning in financial applications are evasion attacks, poisoning attacks, and model extraction attacks.
Evasion attacks occur during the inference phase, where attackers subtly alter input data to produce incorrect classifications, often targeting fraud detection systems. Poisoning attacks compromise models during the training phase by injecting malicious data points, which can bias future model predictions. Model extraction attacks aim to steal proprietary algorithms by querying API endpoints, allowing attackers to replicate commercial ML models without access to training data.
Key points
- Evasion attacks manipulate input data to evade detection systems.
- Poisoning attacks inject malicious data during training to bias models.
- Model extraction attacks steal algorithms via API queries.
- These attacks exploit vulnerabilities in AI/ML systems.
- Adversarial attacks can be automated and scaled.
- Financial institutions face significant risks from these attack modalities.
Related questions
AI and ML Governance in Financial Services
Richard Gwashy Young
Routledge, Taylor & Francis Group