Poison AI Model
Poisoning an AI model is like tampering with a recipe book so that a chef unknowingly cooks a dish that tastes wrong or contains hidden ingredients. An attacker changes the internal settings of the AI so that it behaves in unexpected or harmful ways, either by changing how it learns or by directly altering its final instructions.
Model poisoning is a security compromise where an adversary alters an AI model's internal parameters or training pipeline to induce malicious behavior. This can be achieved through direct weight manipulation, data poisoning during training, or adversarial fine-tuning, resulting in a model that performs specific tasks incorrectly or exhibits degraded performance on demand.
Model poisoning refers to the intentional subversion of an AI model's objective function or internal state, resulting in a compromised model that deviates from its intended behavior. Adversaries may achieve this via direct weight manipulation, training set contamination, or malicious fine-tuning. The resulting poisoned model may exhibit targeted misclassification, concept-specific output manipulation, or systemic performance degradation, often while maintaining nominal performance on benign inputs to evade detection.