Sparsity-Preserving Differentially Private Training of Large Embedding Models
This addresses privacy concerns in recommendation systems and language applications, offering an incremental improvement over DP-SGD by preserving sparsity for more efficient training.
The paper tackles the problem of gradient sparsity loss in differentially private training of large embedding models, presenting DP-FEST and DP-AdaFEST algorithms that achieve a 10^6× reduction in gradient size while maintaining comparable accuracy on real-world datasets.
As the use of large embedding models in recommendation systems and language applications increases, concerns over user data privacy have also risen. DP-SGD, a training algorithm that combines differential privacy with stochastic gradient descent, has been the workhorse in protecting user privacy without compromising model accuracy by much. However, applying DP-SGD naively to embedding models can destroy gradient sparsity, leading to reduced training efficiency. To address this issue, we present two new algorithms, DP-FEST and DP-AdaFEST, that preserve gradient sparsity during private training of large embedding models. Our algorithms achieve substantial reductions ($10^6 \times$) in gradient size, while maintaining comparable levels of accuracy, on benchmark real-world datasets.