Kaiser784's Blog
  • whoami
  • Ingress
  • 90 Day High Frequency
    • Log-1
    • Log-2
    • Log-3
    • Conclusion
  • Making a Boot2root machine with docker
  • Adversarial ML
    • Practical Defenses against Adversarial ML
  • Certifications
    • eJPT
      • Cheatsheet/Notes
        • Enumeration
        • Web Attacks
        • System Attacks
        • Network Attacks
    • Dante ProLabs (HackTheBox)
  • Writeups
    • Google InternetCTF writeup
    • Lakera Gandalf LLM Security
    • OSPG Writeups
    • Matrix - AI Security Challenge by Repello writeup
    • WithSecure AI Challenge - My LLM Doctor writeup
    • WIZ Security Challenges
      • The BIG IAM Challenge writeup
      • K8S LAN Party Writeup
      • EKS Cluster Games Writeup
      • Prompt Airlines Writeup
  • Misc
    • Paul Kadali
    • Redirection
Powered by GitBook
On this page
  • WTF is Adversarial ML ???
  • Attacks
  • Training-Time
  • Data Poisoning
  • Byzantine
  • Decision-Time
  • Evasion Attacks
  • Oracle Attacks
  • Statistical Attack Vectors

Was this helpful?

Adversarial ML

PreviousMaking a Boot2root machine with dockerNextPractical Defenses against Adversarial ML

Last updated 3 months ago

Was this helpful?

WTF is Adversarial ML ???

General terms before we dive in to the attacks

Bias in ML is when a model favors or discriminates against certain groups or outcomes due to flaws in the training data. Think of it like a teacher grading students unfairly because they have preconceived notions (e.g., “students from School X always get low scores”). In adversarial ML, attackers exploit or create this unfairness to harm the model’s credibility or manipulate its predictions.

Related Tactics

Drift happens when a model’s performance worsens over time because the real world changes, and the model doesn’t adapt. There are two types:

  1. Data Drift

    Data drift occurs when the input data the model sees in production is different from the training data. Imagine training a weather model on summer data, then deploying it in winter suddenly, it struggles because snow and ice weren’t in the training data.

  2. Concept Drift

    Concept drift happens when the relationship between inputs and outputs changes. For example, a model trained to predict if someone will buy a product based on their age and income might fail if a new trend (e.g., TikTok challenges) suddenly drives purchases.

Related Tactics

Data Drift related Adversarial Tactics

Concept Drift related Adversarial Tactics

Attacks

Training-Time

Data Poisoning

Byzantine

Decision-Time

Evasion Attacks

Oracle Attacks

Statistical Attack Vectors

Page cover image