Contractive Projection for Privately Trained Dual-Domain MRI Reconstruction

Fast MRI scanners collect only part of the raw frequency-space data, called k-space, and a neural network fills in the rest. The strongest reconstruction networks run in a loop. Each pass cleans up the image, cleans up its frequency representation, and then forces the result to agree with the k-space lines that were actually measured. Hospitals want to train such networks on patient scans with differential privacy (DP), a formal guarantee that the trained model reveals little about any one patient. The standard way to get that guarantee is to clip each patient's gradient and add Gaussian noise to every update. We argue that this noise does more than lower accuracy. It quietly breaks the property that makes the reconstruction loop settle to a single answer: that each learned block shrinks distances between inputs. When that property fails, running the loop longer amplifies artifacts instead of removing them.

The fix we propose borrows the logic of a classic result on private optimization. There, noisy objectives are projected onto a well-behaved function class so that the downstream solver stays stable. We instead project each privatized weight update onto the set of learned blocks that shrink distances. Because this projection only touches already-privatized quantities, it costs no extra privacy budget. The result is a privately trained dual-domain network whose reconstruction loop provably converges to a unique image at every noise level.

Core Claim

The asymptotic-stability literature has established that noise and iterative dynamics interact badly. Differentially Private Distributed Convex Optimization via Functional Perturbation proves that noise injected into the messages of an iterative solver either vanishes, which breaks privacy, or persists, which kicks the solver off its fixed point forever. Their escape is to perturb the objective once and then post-process it. They smooth it by truncating high-order coefficients and project it onto strongly convex functions with bounded derivatives, so that any standard solver run on the result is stable. This paper supplies the motivating mechanism for our proposal, and its projection step is the template we reuse.

The dual-domain restoration literature has the matching iterative structure but no privacy. DuDoRNet alternates a k-space restoration network, an image-domain network, and hard data-consistency layers that overwrite predicted k-space entries with measured ones. We supply the method's architecture template from this work.

Write one pass of such a network as $x_{k+1} = D(G_\theta(x_k))$. Here $x_k$ is the image estimate at pass $k$, $G_\theta$ is the composed learned image and k-space blocks with weights $\theta$, and $D$ is data consistency. In the single-coil Cartesian case, $D(z) = F^{-1}(My + (I-M)Fz)$, where $F$ is the Fourier transform, $M$ the sampling mask and $y$ the measurements. $D$ is affine, and its linear part is an orthogonal projection, so $D$ never increases distances. If $G_\theta$ has Lipschitz constant $\rho<1$, then the Banach fixed-point theorem gives a unique reconstruction $x^\star$ and geometric convergence:

$$\|x_k - x^\star\| \le \rho^k \|x_0 - x^\star\|.$$

Our hypothesis is that DP-SGD pushes $\rho$ above one. Adding an $m\times n$ Gaussian perturbation with entry scale $\sigma$ raises a weight matrix's largest singular value by roughly $\sigma(\sqrt m+\sqrt n)$, a standard random-matrix fact. Accumulated over many noisy steps, this growth multiplies across layers.

The proposal is to follow each privatized update with a projection, $\theta \leftarrow \Pi_\rho(\theta - \eta\,\tilde g)$. Here $\tilde g$ is the clipped, noised gradient, $\eta$ the step size, and $\Pi_\rho$ caps every layer's operator norm so that the composed block has Lipschitz constant at most $\rho$. The frequency domain makes this projection exact and cheap for convolutions. A convolution's singular values are the per-frequency singular values of its kernel's Fourier transform, so clipping them frequency by frequency is itself a dual-domain operation.

What is new is the target of post-processing. Sample-Efficient Differentially Private Fine-Tuning via Gradient Matrix Denoising post-processes gradients to recover their low-rank structure. We post-process weights to preserve a stability property of the inference-time loop. That paper's "alignment paradox" is our second motivating fact: denoised gradients speed up early training yet reach the same final plateau as plain DP-SGD. Our reading is that the late-stage damage is accumulated random weight drift, which gradient denoising cannot undo but a norm projection can bound.

Research Program

The first stage is diagnostic and could falsify the premise. We train a DuDoRNet-style unrolled network on fastMRI multi-coil data under DP-SGD across a sweep of noise multipliers. The privacy unit is the patient volume. We track the per-layer spectral norms, estimated by power iteration, and the end-to-end Lipschitz constant of $G_\theta$. We then run the trained loop for one, two and four times the number of passes it was trained with, and record whether the iterate residual $\|x_{k+1}-x_k\|$ decays or grows.

The second stage builds the projected trainer and asks three questions. How should $\rho$ be set: fixed, or annealed toward one as training ends? Is projection needed on both the k-space and image branches? Does it compose with gradient denoising and with the double-momentum aggregator of Double Momentum and Error Feedback for Clipping with Fast Rates and Differential Privacy? That aggregator keeps one momentum buffer before clipping and one after noising, and it is the natural choice for a federated extension across sites.

The third stage is analysis. Under the contraction constraint, a fixed-point perturbation bound shows how the final error splits into two parts. One part is the approximation cost of restricting to $\rho$-contractive blocks. The other scales with how far the privatized weights sit from their noise-free counterparts, divided by $1-\rho$. The analysis should predict an interior optimal $\rho$ that moves toward one as the privacy budget $\epsilon$ loosens, which the experiments can check.

The multi-coil case needs care. Data consistency with coil sensitivity maps is nonexpansive only under a suitable weighted norm, so we either use that norm or report empirical Lipschitz estimates alongside the proof for the single-coil case.

Baselines

The non-private upper reference is the same unrolled dual-domain architecture trained without DP. It is anchored by the best published hybrid image/k-space result on this benchmark, SSIM around 0.964 on the multi-coil 4x track, reported in Results of the 2020 fastMRI Challenge for a 200M-parameter model. Our models will be far smaller, so we reproduce a matched-size non-private model rather than claim parity with that number. We found no reported result on Lacuna for DP-trained MRI reconstruction of any kind, so every private number in this study is new and must be compared against our own reproductions. DuDoRNet's 32.51 dB PSNR at 5x Cartesian acceleration is on an in-house T1/T2 dataset we cannot access, so it is context only.

The naive reference is the zero-filled inverse Fourier reconstruction, which needs no training and has no privacy cost. Any private model that fails to beat it clearly at a given $\epsilon$ is useless at that budget.

The direct competitor is plain DP-SGD on the same unrolled network. It uses the same clipping norm and noise multiplier and the same accounting, and its learning rate and schedule get the same tuning budget, including aggressive cosine or step decay.

The strongest existing method is DP-SGD with post-hoc gradient denoising following Sample-Efficient Differentially Private Fine-Tuning via Gradient Matrix Denoising, using its $\kappa > 1.02$ threshold on reshaped convolution gradients. It is the strongest because it is also free post-processing, reports a 100% speedup on QNLI with under 1% overhead, and attacks the same noise that we blame. A second strong competitor is Differentially Private Sharpness-Aware Training. It reuses the previous privatized gradient to seek flat regions at no extra privacy cost, and reports 96.00% versus 95.15% for DP-SGD on MNIST at $\epsilon=1$. Flatness is an alternative route to noise robustness.

A feedforward baseline removes the loop entirely: a non-recurrent dual-domain network of matched parameter count, trained with DP-SGD. It tests whether recurrence is worth protecting at all.

The ablations isolate the key ingredient. One projects only once at the end of training instead of every step. One replaces projection with a differentiable spectral-norm reparameterization, so the constraint is learned rather than enforced. The last applies projection to one branch only.

All methods share data splits, privacy accounting, total training steps, and a fixed number of hyperparameter trials. The tuning is itself done on public or held-out non-patient data, to avoid leaking budget. Beating the strongest baseline means two things. At training depth and equal $\epsilon$, the method must improve SSIM by more than the seed-to-seed spread. And when passes are extended to four times training depth, it must show no degradation while the baselines degrade.

Why not just…?

Why not just train with DP-SGD and decay the learning rate hard at the end? A shrinking step size limits late drift, but it does not undo spectral-norm growth already accumulated. Differentially Private Learning Needs Better Features (or Much More Data) and Online Differentially Private Conformal Prediction report that small, fragile learning rates are already what keeps private training stable, at the cost of learning. The tuned-schedule baseline settles this directly. If its Lipschitz estimates stay below one and its extended-pass behavior is stable, our premise is wrong for this setting.

Why not just denoise the gradients with random matrix theory? This is the real threat. The alignment paradox reported in Sample-Efficient Differentially Private Fine-Tuning via Gradient Matrix Denoising suggests it will not fix late-stage damage, but that paper studied language models, not convolutional loops. We run it alone and combined with projection. If it alone keeps $\rho<1$ and matches our accuracy, the contribution shrinks to a diagnostic. If the two are complementary, with denoising speeding early training and projection capping the floor, the combined method is the result.

Why not just drop the loop and use a feedforward network? Without a loop there is no fixed point to lose. But DuDoRNet found that recurrent k-space/image alternation handles non-local aliasing better than single-pass models. The loop also lets compute adapt at test time. The matched feedforward DP baseline measures whether that advantage survives privacy noise. If it does not, we report that recurrence is not worth defending under DP.

Why not just add noise to k-space instead of the training process? Per Differentially Private Distributed Convex Optimization via Functional Perturbation, noise that does not vanish leaves a permanent floor. The fastMRI challenge already reports over-smoothing and hallucinated structures at 8x acceleration without any added noise. Input noise also protects a different quantity, one scan at inference time, which is outside this proposal's threat model.

Proposed Experiments

We use fastMRI multi-coil knee and brain at 4x and 8x acceleration. Privacy budgets are $\epsilon \in \{1, 3, 8\}$, and results are reported as SSIM, PSNR and normalized mean squared error, with three seeds. The federated extension splits training data by acquisition site and field strength into simulated institutions, and compares Clip21-SGD2M with and without projection. Its held-out vendor evaluation mirrors the challenge's transfer track, where cross-vendor quality drops were reported but no number was given.

Stability is measured two ways. We estimate the Lipschitz constant of $G_\theta$, and we plot iterate residuals and SSIM against pass count up to four times training depth. As a hallucination proxy, we measure how much the reconstruction changes under small perturbations of the measured k-space lines, restricted to unmeasured frequencies.

Three checks decide the outcome. First, does plain DP-SGD actually push $\rho$ above one, and does it do so more strongly as the noise multiplier grows? A no here ends the stability story. Second, do plain-DP models get worse with extended passes while projected models plateau? This is the signature the theory predicts. Third, at training depth and equal $\epsilon$, does projection beat the best of tuned DP-SGD, gradient denoising and DP-SAT by more than seed variance, with the gap widening at smaller $\epsilon$? If the third check fails but the first two hold, the honest claim is narrower: projection makes private recurrent reconstruction safe to run longer without making it more accurate. That is still clinically meaningful given the hallucination findings, but it is a smaller result.