PB-AMD
An optimization algorithm for noise variance scheduling in differentially private learning that combines pathwise adjoint gradients with a mirror-descent entropy update, strictly maintaining the total privacy budget unchanged while satisfying zCDP constraints.
Timeline (last 90 days)
The paper proposes the Privacy-Budget Adjoint Mirror Descent (PB-AMD) algorithm, which combines pathwise adjoint gradients with entropy updates to solve for the noise variance schedule.
In tracking-task experiments, PB-AMD outperforms several baseline allocation strategies.
In the learning-rate restart setting, PB-AMD exhibits better predictive performance than uniform allocation, while the learning-rate allocation performs comparably to it.
Under the condition of reaching the same stationarity tolerance, the schedule optimization time of PB-AMD is lower than that of projected gradient descent.
All Facts (4)
The paper proposes the Privacy-Budget Adjoint Mirror Descent (PB-AMD) algorithm, which combines pathwise adjoint gradients with entropy updates to solve for the noise variance schedule.
50%UnverifiedIn tracking-task experiments, PB-AMD outperforms several baseline allocation strategies.
50%UnverifiedIn the learning-rate restart setting, PB-AMD exhibits better predictive performance than uniform allocation, while the learning-rate allocation performs comparably to it.
50%UnverifiedUnder the condition of reaching the same stationarity tolerance, the schedule optimization time of PB-AMD is lower than that of projected gradient descent.
50%