AI0-001 AI Security Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Homomorphic encryption

  Why it's wrong here

  Homomorphic encryption permits computation on ciphertext, yet training requires repeated gradient exchange and aggregation across sources, which this alone does not orchestrate. It is tempting because encrypted data stays encrypted during processing, and would be correct when a single party must compute over another's encrypted data.

• ✗

  Secure multi-party computation

  Why it's wrong here

  Secure multi-party computation lets parties jointly compute a function over combined inputs, but it does not itself coordinate iterative model training across many decentralised data holders. It is tempting because it genuinely hides raw inputs during computation, and would be correct for a small, fixed set of parties computing an agreed function.

• ✗

  Differential privacy

  Why it's wrong here

  Differential privacy adds calibrated noise to outputs or gradients to bound individual disclosure, but it does not keep raw data at each source or coordinate distributed training. It is tempting because it is a genuine privacy guarantee, and would be correct when publishing aggregate statistics or model outputs that must resist inference about individuals.

• ✓

  Federated learning

  Why this is correct

  Federated learning trains a shared model by exchanging only parameter updates, such as gradients or weights, between decentralised devices and a coordinating server. Raw records remain on each source, satisfying the stem's constraint that data never leaves its origin. This differs from differential privacy, which adds noise, and homomorphic encryption, which computes on ciphertext.