Decision-Based Attacks

Decision-based attacks are a class of adversarial machine learning attacks where an attacker manipulates input data to cause a machine learning model to make incorrect predictions, based solely on observing the model's final output decisions (e.g., class labels) without access to internal details like gradients or probabilities. These attacks exploit the model's decision boundaries by iteratively perturbing inputs until a misclassification occurs, often using techniques like boundary search or query-based optimization. They are particularly relevant in black-box scenarios where the attacker has limited knowledge of the model architecture or training data.

Also known as: Decision Boundary Attacks, Black-Box Adversarial Attacks, Query-Based Attacks, Decision-Only Attacks, Boundary Attacks

Why learn Decision-Based Attacks?

Developers should learn about decision-based attacks to enhance the security and robustness of machine learning systems, especially in applications like fraud detection, autonomous vehicles, or cybersecurity where adversarial inputs can have serious consequences. Understanding these attacks helps in designing defensive strategies, such as adversarial training or input sanitization, to mitigate risks in real-world deployments where models are exposed to malicious actors. It is crucial for roles involving AI security, model validation, or deployment in sensitive environments.

Learning Resources

• docs: Adversarial Robustness Toolbox Documentation
• tutorial: Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models
• course: Coursera: Adversarial Machine Learning
• video: YouTube: Decision-Based Attacks in Adversarial ML
• book: Book: 'Adversarial Machine Learning' by Yevgeniy Vorobeychik and Murat Kantarcioglu