Adaptive optimizers like Adam, RMSProp, and AdaGrad are the default choice for training modern deep learning models. They work by dynamically adjusting the learning rate for individual parameters based on historical gradient scales. By dividing the first-moment estimate (the directional signal) by the square root of the second-moment estimate (the uncentered variance), these optimizers perform a form of "sign descent." This ensures that parameters with small, rare gradients receive large enough steps to learn, while frequently updated parameters do not explode.
However, when training models with Differential Privacy (DP), this adaptive mechanism breaks down completely. To guarantee privacy, DP algorithms clip individual gradients and inject random Gaussian noise. As noted by Hilal Asi et al. in Private Adaptive Gradient Methods for Convex Optimization and Tian Li et al. in Private Adaptive Optimization with Side information, this injected noise often drowns out the coordinate-wise gradient variations that adaptive methods rely on. Instead of outperforming standard Stochastic Gradient Descent (SGD), private adaptive optimizers frequently degrade to the point where they perform no better than simple DP-SGD.
The Mathematics of Preconditioner Inflation
Why does adding zero-mean Gaussian noise to gradients systematically ruin an adaptive optimizer? The answer lies in how the second moment is calculated.
Let $g$ be the true clipped gradient, and let $\tilde{g} = g + \eta$ be the private gradient, where $\eta \sim \mathcal{N}(0, \sigma^2)$ represents the injected privacy noise. While the first-moment estimate remains unbiased because $\mathbb{E}[\tilde{g}] = g$, the second-moment estimate tracks $\mathbb{E}[\tilde{g}^2]$. Expanding this expectation reveals a critical mathematical flaw:
$$\mathbb{E}[\tilde{g}^2] = \mathbb{E}[(g + \eta)^2] = \mathbb{E}[g^2] + 2\mathbb{E}[g\eta] + \mathbb{E}[\eta^2] = \mathbb{E}[g^2] + \sigma^2$$
Because the noise is squared, its variance $\sigma^2$ introduces a systematic, persistent positive bias into the second-moment estimate. As analyzed by Qiaoyue Tang et al. in DP-Adam: Correcting DP Bias in Adam's Second Moment Estimation and Duc Dm et al. in FIBER: A Differentially Private Optimizer with Filter-Aware Innovation Bias Correction, this phenomenon is known as preconditioner inflation.
When the noise variance $\sigma^2$ is large, it dominates the denominator of the update rule. In high-noise regimes, Shubhankar Mohapatra et al. demonstrate in The Role of Adaptive Optimizers for Honest Private Hyperparameter Selection that the second-moment estimates stabilize to a virtually constant value across all dimensions. Instead of scaling updates based on the true geometry of the loss surface, the denominator becomes artificially inflated and uniform. This squashes the effective step sizes toward zero, destroying the characteristic sign-descent behavior of adaptive methods and reverting the algorithm to standard gradient descent with momentum.
The Cost of Bias on Heavy-Tailed Data
The practical consequence of preconditioner inflation is particularly severe for models trained on heavy-tailed distributions, such as large language models. In On the Performance of Differentially Private Optimization with Heavy-Tail Class Imbalance, Qiaoyue Tang et al. show that the noise floor masks the true data curvature of rare features or classes.
In non-private training, Adam naturally scales up updates for rare features because their historical second moments are small. Under DP, however, the additive noise variance $\sigma^2$ ensures that the second-moment estimate for an infrequent feature is nearly identical to that of a frequent feature. The optimizer treats rare features as if they are merely noise, suppressing their updates and causing severe performance degradation on underrepresented data.
Restoring Adaptivity via Noise-Variance Subtraction
The most direct solution to preconditioner inflation is to explicitly subtract the known noise variance from the second-moment state. Because the noise multiplier, clipping threshold, and batch size are public hyperparameters, the exact variance $\sigma^2$ is known ahead of time.
In DP-AdamBC: Your DP-Adam Is Actually DP-SGD (Unless You Apply Bias Correction), Qiaoyue Tang et al. formalize this post-processing step. By subtracting $\sigma^2$ from the running second-moment estimate before taking the square root, they eliminate the systematic positive bias without consuming any additional privacy budget:
$$\hat{v}_t = \max(v_t - \sigma^2, 0)$$
This simple correction prevents the update steps from being squashed toward zero. It restores the uniform step-size behavior of sign descent, allowing private Adam to behave like its non-private counterpart and significantly boosting accuracy on language, vision, and graph tasks.
Alternative Paradigms for Noise-Aware Optimization
Beyond direct subtraction, researchers have developed several complementary strategies to protect adaptive optimizers from privacy noise by altering how the preconditioner is constructed or structured:
- Delayed Preconditioning: Instead of calculating the preconditioner at every step using noisy gradients, Tian Li et al. propose $DP^2$ in Differentially Private Adaptive Optimization with Delayed Preconditioners. By averaging gradients over multiple steps before updating the preconditioner, the mechanism smooths out the high-frequency privacy noise, yielding a cleaner estimate of the gradient geometry.
- Preconditioner Flooring and Filtering: In PRISM: Gauge-Invariant Tangent-Space Differentially Private LoRA, Shihao Wang et al. prevent the optimizer from treating injected noise as a valid gradient signal by flooring the preconditioner based on the known noise level. Similarly, the FIBER framework developed by Duc Dm et al. uses second-order innovation-space filtering to dynamically correct the bias introduced by squared noise.
- Public Side Information: Another approach is to sidestep noisy estimation altogether. In Private Adaptive Optimization with Side information, Tian Li et al. use a small amount of non-sensitive public data or structural heuristics to construct the preconditioner before any private noise is added, ensuring the geometric signal remains pristine.
- Rethinking Noise Structures: The interaction between noise design and optimizer dynamics also extends to how privacy mechanisms are structured. For instance, Ryan McKenna demonstrates in Scaling up the Banded Matrix Factorization Mechanism for Large Scale Differentially Private ML that standard error metrics like Root Mean Squared Error can misrepresent what adaptive optimizers actually need, showing that fewer correlation bands can sometimes improve final accuracy. Concurrently, alternative diffusion mechanisms, such as the mean-reverting processes explored by Borja Balle et al. in Privacy Amplification by Mixing and Diffusion Mechanisms, offer a way to optimize the bias-variance trade-off directly within the privacy mechanism itself.
These diverse methods highlight a shared realization across the field: the primary obstacle to private adaptive optimization is not the presence of noise itself, but rather how standard update rules inadvertently distort that noise into a destructive systematic bias. By making optimizers explicitly aware of the noise variance, we can preserve the geometric insights of adaptive learning under strict privacy guarantees.
Key Foundational Work
Private Adaptive Gradient Methods for Convex Optimization (2021) — contributed 1 concept elements to this direction.
Privacy Amplification by Mixing and Diffusion Mechanisms (2019) — contributed 1 concept elements to this direction.