Transferability of Adversarial Examples
What You'll Learn
You'll discover why adversarial examples that fool one ML model often succeed against completely different models, creating a dangerous cross-model vulnerability.
The Cross-Model Attack Phenomenon
When you craft an adversarial example to fool Model A, you might expect it only works against that specific model. Surprisingly, the same perturbed input often fools Model B, Model C, and even models with entirely different architectures—CNNs, ResNets, transformers, you name it.
This property is called transferability, and it's one of the most concerning aspects of adversarial machine learning.
Why Does Transferability Happen?
Think of it like optical illusions for humans: different people with different visual systems still fall for the same illusion. Similarly, ML models trained on the same data learn related decision boundaries, even if their internal architectures differ.
Key factors driving transferability:
- Shared training data – Models trained on ImageNet learn similar high-level features (edges, textures, shapes)
- Similar task objectives – Classification models optimize for similar decision surfaces
- Common vulnerabilities – Models exploit similar statistical regularities, creating overlapping weaknesses
- Linear approximations – Many models behave approximately linearly in local regions, making perturbations effective across architectures
The Black-Box Attack Enabler
Transferability is what makes black-box attacks practical. An attacker can:
- Train a substitute model locally
- Generate adversarial examples against their own model
- Transfer those examples to attack your production system—without ever querying it
This transforms adversarial attacks from theoretical curiosities into real-world threats against deployed systems.
Defense Implications
Because adversarial examples transfer, you can't rely on "security through obscurity" of your model architecture. Ensemble defenses that combine diverse models show promise, but even they're vulnerable to attacks optimized for transferability.
Key Takeaway: Adversarial examples crafted for one model reliably fool other models because they exploit shared learned features and decision boundaries, enabling practical black-box attacks without direct access to the target system.