Targeted vs Untargeted Attacks
What you'll learn: The critical distinction between adversarial attacks that force a specific wrong prediction versus those that simply cause any misclassification.
The Core Distinction
When crafting adversarial examples, attackers have two fundamentally different goals that define their strategy:
Untargeted attacks aim to make the model predict anything except the correct class. The attacker doesn't care what the model outputs—as long as it's wrong. Think of this as vandalizing a street sign so it becomes unreadable; you don't care what drivers think it says, just that they misread it.
Targeted attacks aim to make the model predict a specific incorrect class chosen by the attacker. This is much more constrained and typically harder. It's like carefully altering that street sign so drivers specifically think it says "Turn Right" instead of "Stop"—you need precise control over the outcome.
Why This Matters
The distinction has major practical implications:
- Security severity: Targeted attacks are often more dangerous. Making a spam filter classify one message as "not spam" is targeted; making it misclassify randomly is untargeted.
- Attack difficulty: Targeted attacks generally require larger perturbations or more sophisticated techniques because you're constraining the output space.
- Defense strategies: Some defenses work better against one type than the other, so understanding the threat model is crucial.
In a white-box or black-box setting (from the threat model you learned previously), attackers choose their strategy based on their goals and available information.
Key Takeaway: Untargeted attacks seek any misclassification, while targeted attacks force a specific wrong output—targeted attacks are typically harder to execute but potentially more dangerous in real-world scenarios.