AAISM Domain 2 Practice Test

Manipulating embedding representations to distort a model’s output is about tampering with the internal numerical mapping the model uses to understand inputs. Embeddings turn discrete items like words or users into dense vectors that the neural network relies on to compute predictions. If someone alters these vectors—by changing the embedding table, injecting crafted embeddings, or otherwise skewing how inputs are embedded—the downstream computations drift, leading to incorrect or biased results. This specific form of attack targets the representation layer itself, rather than the broader training data or the model parameters. That’s why the term embedding tampering fits best. Data poisoning would involve corrupting training data to affect performance more generally, model tuning is just adjusting parameters to change behavior, and data sharding is about how data is stored or distributed and isn’t about manipulating the model’s outputs.