A privacy accountant is the part of a differential privacy system that tracks how much privacy budget has been spent over time. As training runs across rounds, batches, or federated updates, each step adds some privacy loss. The accountant combines those steps into a cumulative measure, often reported as values such as epsilon and delta, so teams can see whether the model is still within its allowed limit.
This matters because privacy guarantees are only meaningful if they match the real training run and the model that was released. If the sampling rate, clipping, noise, or number of rounds changes, the final privacy claim changes too. In attacks, a weak or misused accountant can let an over-budget model ship while still showing a reassuring number. In defense, the accountant should gate release, support audit logs, and tie the budget to a specific model version and training history.