Which term describes the manipulation of data inputs to degrade the AI model's accuracy?

• A. Data poisoning
• B. Model drift
• C. Prompt injection
• D. Adversarial Inference

Answer: (A)

Data poisoning describes the act of tampering with the data used to train or update a model, with the goal of reducing its accuracy. By injecting mislabeled samples or corrupted data into the training process, an attacker shifts the model’s learned patterns, causing degraded performance on legitimate inputs. This directly targets the model’s knowledge base, making it the most fitting term for degrading accuracy through data manipulation. Other concepts involve different problems: model drift stems from natural changes in data distribution over time and isn’t necessarily a deliberate attack on training data; prompt injection targets behavior within a live session by manipulating prompts rather than the training data; and adversarial inference refers to attempts to fool a model at inference time, which is about exploiting responses rather than degrading training accuracy.