Techniques that manipulate AI models by introducing deceptive inputs to cause incorrect outputs.
Deliberate, often imperceptible, modifications to input data - images, text, or audio - that exploit vulnerabilities in an AI model’s decision boundaries. Such attacks highlight black-box system weaknesses and drive the need for proactive defenses: adversarial-training (injecting crafted examples during training), input-sanitization layers, and ongoing “red-team” penetration tests.
Security researchers place tiny, artful stickers on a stop sign so that a self-driving car’s vision system misreads it as “Speed Limit 45.” The automaker responds by integrating adversarial-example detectors and hardening the model with randomized input preprocessing.
