When training a machine learning model with Differential Privacy (DP), we typically inject random noise into the gradients during optimization (such as in DP-SGD). This noise is mathematically calibrated to mask the contribution of any single individual's data point. However, as modern deep learning scales to millions or billions of parameters, a catastrophic phenomenon emerges: the utility of the model collapses.
As noted by Chauhan et al. in their work on large speech encoders, the required privacy noise scales with the model's dimensionality, frequently overwhelming the true learning signal. To understand why this happens—and how a new wave of research is overcoming it—we must look at the geometry of high-dimensional spaces.
The Geometry of Privacy Noise
In standard DP mechanisms, noise is added independently to every single coordinate of a gradient vector. If a model has $d$ parameters, we add a $d$-dimensional noise vector $\mathbf{n} \sim \mathcal{N}(0, \sigma^2 \mathbf{I})$.
The length (or $L_2$ norm) of this noise vector behaves deterministically in high dimensions:
$$\|\mathbf{n}\|_2 \approx \sigma \sqrt{d}$$
As the dimension $d$ grows, the total magnitude of the noise injected into the system grows as $\sqrt{d}$ (or even polynomially, as described by Bellet et al.).
Meanwhile, the true gradient signal does not necessarily grow with the dimension. In fact, the informative components of the gradient are often concentrated along a few critical directions. This creates an unfavorable signal-to-noise ratio: the true learning signal gets dispersed across millions of dimensions, while the noise fills the entire ambient space, completely obscuring the signal.
The authors of PILLAR: How to make semi-private learning more effective point out that this full-dimensional noise injection leads to poor convergence because optimization algorithms waste steps fitting the random noise rather than the true data distribution. Furthermore, Habernal demonstrates that ignoring how sensitivity scales with the dimensionality of the latent space can lead to a rapid degradation of actual privacy protections or utility.
Ambient Space (d dimensions)
+---------------------------------------------------+
| . . . . . Noise fills all dimensions . . . . . |
| . . . . . . . . . . . . . . . . . . . . . . . . |
| ___________________________ |
| / True Gradient Subspace / |
| / (k << d dimensions) / |
| / [Signal Lives Here] / |
| /__________________________/ |
| . . . . . . . . . . . . . . . . . . . . . . . . |
+---------------------------------------------------+
To escape this curse of dimensionality, researchers have developed three primary strategies that shift the privacy penalty from the ambient dimension of the model to the structural dimension of the data.
Strategy 1: Subspace Projection via Public Data
If the noise scales with the ambient dimension $d$, the most direct solution is to avoid optimizing in the ambient space altogether. Gradients in deep networks are highly redundant and often reside in a much lower-dimensional subspace.
The authors of Bypassing the Ambient Dimension: Private SGD with Gradient Subspace Identification introduced Projected DP-SGD (PDP-SGD). Instead of adding noise to the full $d$-dimensional gradient, they use a small, non-sensitive public dataset to identify the principal components (the subspace) where the gradients actually live. By projecting the private gradients onto this $k$-dimensional subspace ($k \ll d$) before adding noise, the privacy mechanism only injects noise proportional to $\sqrt{k}$ rather than $\sqrt{d}$.
Similarly, Pinto et al. leverage public data in their PILLAR framework to perform dimensionality reduction via PCA before applying DP. Because the orthogonal noise outside the subspace is discarded, the gradient signal remains intact, allowing large models to converge even under strict privacy budgets.
Strategy 2: Correlated Noise and Effective Rank
Do we always need public data to escape the curse of dimensionality? Not necessarily. Another approach changes how noise is structured over time.
Standard DP-SGD adds independent noise at every single training step. However, Choquette-Choo demonstrates that adding correlated noise across steps provably outperforms independent noise. The intuition here is elegant: independent noise accumulates across all ambient dimensions over the course of training. By introducing temporal correlation (via the $\nu$-DP-FTRL algorithm), the noise error scales with the effective dimension of the data matrix rather than the ambient parameter count.
This aligns with the theoretical findings of Song et al., who showed that for unconstrained Generalized Linear Models, the error of private gradient descent depends on the rank of the data matrix rather than the ambient dimension of the model space. When the underlying signal exists in a low-rank subspace, a properly calibrated privacy mechanism can avoid paying the penalty for overparameterization.
Strategy 3: Sparsification and Coordinate Selection
A third strategy to mitigate dimension-dependent noise scaling is to explicitly restrict the number of parameters updated at any given time. If you only update a small subset of coordinates, you only need to add noise to those coordinates, effectively shrinking $d$.
For instance, Makni et al. developed SPARTA, an optimization framework for sparse fine-tuning. SPARTA uses private gradient information and a row-grouping technique to select a sparse subset of model weights to update. By shrinking the active parameter space, the injected noise is heavily constrained, preventing the accuracy collapse typical of full-parameter private fine-tuning.
On the optimization theory side, Bellet et al. utilize greedy coordinate descent to navigate high-dimensional empirical risk minimization. They show that updating coordinates selectively can mitigate the polynomial noise scaling that plagues standard high-dimensional private optimization, offering a clear path forward for scaling private learning to massive architectures.
Key Foundational Work
Differentially Private Learning with Kernels (2013) — contributed 1 concept elements to this direction.
(Near) Dimension Independent Risk Bounds for Differentially Private Learning (2014) — contributed 1 concept elements to this direction.