Random Allocation Sampling
A sampling scheme in differential privacy in which each user uniformly selects k of t steps at random to participate in. Compared with Poisson sampling, the privacy budget is more evenly distributed, but the fixed cardinality breaks independence across steps.
Timeline (last 90 days)
In differentially private random allocation sampling, each user's data is randomly and uniformly selected to be used in k out of t steps.
In random allocation sampling, because each user's number of participating steps is fixed at k, the privacy budget is distributed more evenly across steps, avoiding the tail risk under Poisson sampling that some users may happen to participate in too many steps.
Fixed-cardinality sampling breaks independence across steps, and its privacy loss accounting cannot directly apply the decomposition formula of Poisson sampling, requiring special handling.
Random allocation sampling is widely used in federated learning and distributed optimization, and the efficiency of privacy accounting has always been one of the key factors constraining its deployment.
All Facts (4)
In differentially private random allocation sampling, each user's data is randomly and uniformly selected to be used in k out of t steps.
50%UnverifiedIn random allocation sampling, because each user's number of participating steps is fixed at k, the privacy budget is distributed more evenly across steps, avoiding the tail risk under Poisson sampling that some users may happen to participate in too many steps.
50%UnverifiedFixed-cardinality sampling breaks independence across steps, and its privacy loss accounting cannot directly apply the decomposition formula of Poisson sampling, requiring special handling.
50%UnverifiedRandom allocation sampling is widely used in federated learning and distributed optimization, and the efficiency of privacy accounting has always been one of the key factors constraining its deployment.
50%