PART 6 · MODULE 5Protecting AI models
Defense Mechanisms
Strategies to defend AI systems
Loading module...
Strategies to defend AI systems
Attacks are only half the story. The real goal of AI security is building systems that can withstand adversarial inputs. No single defense is perfect, but layering multiple defenses makes attacks significantly harder.
Defense in Depth: Layers of Protection
Train the model on adversarial examples alongside normal data. The model learns to recognize and resist attacks. Like vaccinating the immune system by exposing it to weakened versions of a virus.
Clean inputs before the model sees them. Techniques like JPEG compression, spatial smoothing, or bit-depth reduction can remove adversarial perturbations while keeping the image recognizable to humans.
Mathematical guarantees that no perturbation below a certain size can fool the model. Instead of hoping a defense works, you can prove it. Computationally expensive but theoretically sound.
Add random noise to inputs before classification and average the results. An attacker cannot optimize against randomness because the defense is different every time. Trading accuracy for robustness.
Instead of making the model robust, train a separate detector to spot adversarial inputs before they reach the main model. Flag suspicious inputs for human review.
Every new defense inspires a new attack. Adversarial training was beaten by stronger iterative attacks. Preprocessing defenses were circumvented by attacks that account for the preprocessing. Detection models were fooled by adaptive adversaries.
This does not mean defense is hopeless. It means that no single technique is enough. The best approach is defense in depth— multiple overlapping layers that make attacks progressively harder and more expensive.
Defending AI requires multiple layers — no single defense is enough. Combine adversarial training, input validation, detection, and monitoring. Make attacks expensive and unreliable, even if you cannot make them impossible.
For hands-on practice, check out the AI Security course on GitHub.