When ten hospitals jointly train a survival model, two separate locks protect each hospital's model update. Encryption hides the individual update in transit and during aggregation. Differential privacy, which adds calibrated random noise so that no single record changes the output much, protects the combined model the server finally sees. These locks fail in different ways. Encryption can fail retroactively: an adversary who records today's traffic can decrypt it once a quantum computer exists, an attack called "harvest now, decrypt later." Network measurements show this is a present risk, not a hypothetical one, because most servers still force clients back onto quantum-vulnerable key exchange. Medical data must also stay secret for decades. Differential privacy fails differently: in small cohorts, the noise needed is so large relative to the signal that training breaks down.
This note proposes treating the noise as a second line of defense that remains in force after the encryption is gone. In distributed differential privacy, each hospital adds part of the noise itself before encrypting. If the encryption is later broken, an attacker sees each hospital's update with only that hospital's share of the noise. The guarantee left over, which we call the privacy floor, is weaker than the central guarantee by a factor equal to the number of participants. With a thousand phones that factor makes the floor worthless. With ten hospitals it is only ten, and the floor can mean something. The research question is how cheaply the floor can be raised to a useful level, and what that costs in model accuracy on small survival cohorts.
Core Claim
Consider $N$ hospitals. In each of $T$ rounds, every hospital clips its integer-encoded update to $\ell_2$ norm $C$ and adds discrete Gaussian noise of per-coordinate variance $\sigma_\ell^2$; a discrete Gaussian is the integer-valued analogue of the usual Gaussian. Privacy is measured in zero-concentrated differential privacy, where a smaller parameter $\rho$ means stronger privacy and $\rho$ adds up over rounds. Two guarantees then hold together:
$$\rho_{\text{central}} \approx \frac{T C^2}{2 N \sigma_\ell^2}, \qquad \rho_{\text{floor}} = \frac{T C^2}{2 \sigma_\ell^2} = N \cdot \rho_{\text{central}}.$$
The first holds while secure aggregation hides individual updates, because the server sees the summed noise of variance $N\sigma_\ell^2$. The second holds against anyone who later strips the encryption off a single hospital's harvested update. The form follows from the Gaussian mechanism's $\rho = \Delta^2/(2\sigma^2)$ per release, with sensitivity $\Delta = C$. The central form is approximate because a sum of discrete Gaussians is only close to a discrete Gaussian, a gap that The Distributed Discrete Gaussian Mechanism shows is exponentially small.
Our claim is that, for clinical federations with $N \approx 10$, three ingredients together allow a floor near the budgets that small-cohort work already treats as usable ($\varepsilon \approx 5$–$9$) while keeping central utility close to today's trusted-server systems. The first is an explicit noise split chosen for the floor rather than only for the central guarantee. The second is server-side post-processing clipping of the decoded noisy sum. The third is projecting updates into a low-dimensional subspace before noise is added, so the total noise energy grows with the subspace rank rather than the parameter count. Encryption is provided by post-quantum lattice-based secure aggregation. The contribution is the two-regime accounting, the attack-validated floor, and the accuracy/floor trade-off curve. The cryptography is not new.
What Exists and What Is Missing
Three systems sit closest to this idea, and each covers a different piece.
Beskar builds post-quantum secure aggregation from CRYSTALS-Kyber (key exchange) and CRYSTALS-Dilithium (signatures). It offers either local or central differential privacy, and at $\varepsilon = 10$ on MNIST it reports 0.87 accuracy under central privacy versus 0.78 under local privacy. It treats the two as alternatives. It never reports what guarantee survives if the encryption fails, and it does not study small cohorts.
The Distributed Discrete Gaussian Mechanism supplies our noise mechanism. Its contributions are client-side discrete noise, flattening by a randomized Hadamard transform, randomized rounding, and modular secure aggregation at 16 to 18 bits per coordinate. It targets about 1,000 clients per round, where the floor is meaningless, and it runs over classical Diffie-Hellman secure aggregation.
Buffalo shows that Ring-LWE masking (Ring-LWE is a lattice problem believed hard even for quantum computers) works for federated aggregation, but it provides no differential privacy at all.
Two papers motivate the small-cohort setting. Practical Challenges in Differentially-Private Federated Survival Analysis shows that ten simulated hospitals on GBSG, METABRIC and SUPPORT diverge under client-level privacy. It also shows that clipping the already-noisy average at the server (DPFed-post) recovers up to 17% on average at $\varepsilon = 5.4$ and 12% at $\varepsilon = 8.9$. That work assumes a trusted server and plaintext channels. We reuse its post-processing step, which stays valid after secure aggregation because the server only ever touches the decoded noisy sum. Post-Quantum Cryptography Network Instrument motivates the threat. It measured 0.029% post-quantum adoption across 20.5 million SSH connections and found over 83% of servers on pre-2019 software, so silent fallback to classical key exchange is the normal case.
Three papers shape the design.
- PILLAR supplies the subspace idea. It projects features onto a public-data principal subspace so that private sample complexity scales with rank $k$, not ambient dimension, and reports CIFAR-100 accuracy in a 10% data regime at $\varepsilon = 0.7$ rising from 7.5% to 18.3%. It is limited to linear heads on frozen features. We instead project whole model updates, which is a different and untested use of the idea.
- The Discrete Gaussian for Differential Privacy rules out rounding continuous Gaussian noise: rounded noise has heavier tails, and floating-point sampling leaks through low-order bits. We therefore use its exact sampler. - Improving ML Attacks on LWE warns that sparse, low-entropy secrets are recoverable. Any lattice instantiation we test avoids such secrets.
Research Program
The first piece is accounting. We will derive the two-regime guarantee above with all the practical corrections included: inflation of the sensitivity bound from randomized rounding, the gap from summing discrete Gaussians, and composition over all rounds. The composition point matters because a harvesting adversary collects every round, and with ten hospitals participating every round there is no amplification from random subsampling to help. We will also handle partial compromise. If $j$ of the $N$ hospitals collude with a future decryptor, an honest hospital's update is hidden by noise from the $N - j$ others while the masks hold, and by its own noise once they fail.
The second piece is a pipeline in which the floor is a design target. Each hospital trains locally on a deep survival model (DeepSurv and DeepHit, as in the clinical paper). It then projects the model update onto a rank-$k$ subspace, clips to $C$, flattens and discretizes as in the distributed discrete Gaussian paper, and adds noise with $\sigma_\ell$ set by the target floor. Finally it masks the result with pairwise pseudorandom masks whose seeds are agreed through Kyber, following Beskar's offline precomputation. The server unmasks the sum and applies DPFed-post clipping.
Projection lowers the noise norm from roughly $\sigma_\ell\sqrt{d}$ to $\sigma_\ell\sqrt{k}$ for parameter count $d$, which is what makes a larger $\sigma_\ell$ affordable. We will compare a public-data subspace, built from a disjoint public survival cohort, against a random subspace and against no projection. This tells us whether the public data is doing the work or whether dimension reduction alone suffices.
The third piece is an empirical check that the floor is real. We will simulate the harvest-now-decrypt-later outcome by handing an attacker every round's per-hospital noisy update. We will then run membership inference and gradient-inversion attacks, and compare the attacker's success against the success rate implied by $\rho_{\text{floor}}$.
Training overhead enters as a constraint, not a contribution. Client-level clipping needs no per-sample gradients, so the costs to measure are discretization, flattening, masking, and the extra bits per coordinate needed to keep the modulus large enough for the larger noise without wrap-around. We report these as client CPU time and bytes per round.
Baselines
The naive reference point is plaintext federated averaging without noise. It gives the accuracy ceiling, and its floor is infinite, meaning no protection, because harvested updates are raw. Every privacy method is measured as a fraction of its C-index.
The strongest existing small-cohort method is DPFed-post from Practical Challenges in Differentially-Private Federated Survival Analysis. It uses continuous central noise added by a trusted server plus server-side clipping, run at $\varepsilon = 5.4$ and $8.9$. Its harvested-update floor is also infinite, since individual updates carry no noise. The file reports only the aggregate "up to 17%" and "12%" improvements; for GBSG, METABRIC and SUPPORT individually, no reported result was found on Lacuna. We will therefore rerun it on the same ten-hospital splits, with the same models and the same hyperparameter search budget, with tuning done on a public or synthetic proxy rather than on private data.
The strongest distributed competitor is standard distributed discrete Gaussian noise over Kyber-based secure aggregation, with noise split evenly to meet a central target. Its floor is exactly $N\rho_{\text{central}}$. Its reported reference point is roughly 25.2% validation accuracy on Stack Overflow next-word prediction at $\varepsilon \approx 10$ with 18 bits, from The Distributed Discrete Gaussian Mechanism.
The opposite extreme is Beskar-style local privacy, where floor equals central guarantee. Its reported point is 0.78 MNIST accuracy against 0.87 for central privacy at $\varepsilon = 10$. We will reproduce that number before moving to survival data.
Our method beats these if, at a matched floor, it reaches higher C-index and lower integrated Brier score than both distributed baselines. At matched central $\varepsilon$, it should stay within a stated margin of DPFed-post. We fix that margin in advance at two C-index points. Two ablations isolate the ingredients: removing projection, and replacing the public subspace with a random one.
Why not just…?
Why not just trust post-quantum encryption and use a trusted-server central guarantee? If Kyber is deployed end to end and never fails, the floor is insurance you never claim. But the measurements in Post-Quantum Cryptography Network Instrument show that fallback to classical channels is common, and the file notes side-channel risks in optimized lattice code on small cores. A guarantee that does not rest on the encryption has value precisely for decades-long medical secrecy. The experiments report the accuracy cost of buying this insurance, so readers can judge the price directly.
Why not just use local privacy, as Beskar already offers? Local privacy maximizes the floor, but Beskar's own numbers show a nine-point MNIST drop, and it reports a sharper drop on CIFAR-100. Our method lies on the curve between local and evenly split noise. The experiment that settles the question is a plot of floor $\varepsilon$ against C-index for all three. If local privacy plus DPFed-post already matches our curve, the projection ingredient is unnecessary, and we will say so.
Why not just run standard distributed noise, since with ten hospitals its floor is already only ten times the central budget? This is the most serious threat. If evenly split noise at Kerkouche-level central budgets already yields an acceptable floor, the contribution shrinks to an accounting observation plus attack validation. That observation is still useful, but it is smaller. The decisive check is whether, at central $\varepsilon \approx 5$, the even-split floor is far weaker than what attacks on harvested updates tolerate. If so, raising it requires our utility-recovery ingredients.
Why not let the lattice encryption's own error term serve as the privacy noise? Lattice ciphertexts already contain discrete Gaussian error, and reusing it could save a separate sampler. But reuse does not change the floor: once the masks are removed, what remains is the same per-hospital noise whether one calls it encryption error or privacy noise. It also couples cryptographic parameters to privacy parameters. We include it only as an engineering ablation that measures bytes and CPU time against separate Kyber masks plus an independent sampler.
Decisive Experiments
All survival experiments use GBSG (2,232 patients), METABRIC (1,904) and SUPPORT (8,873), each split across ten simulated hospitals as in the clinical paper. We sweep the target floor $\varepsilon$ from about 4 to 16 and report central $\varepsilon$, C-index, integrated Brier score, harvested-update attack advantage, bits per coordinate, and client overhead. MNIST and CHMNIST with ten clients anchor the results against Beskar's reported numbers.
Three outcomes would decide the proposal. The first is whether projection plus post-processing lets a floor near $\varepsilon \approx 8$ come within two C-index points of trusted-server DPFed-post. The second is whether measured attack success on harvested updates tracks $\rho_{\text{floor}}$, confirming that the guarantee is not vacuous in practice. The third is whether the public subspace beats a random subspace, which tells us whether public data is essential or whether plain dimension reduction is enough. A negative answer on the first point would still be informative: it would show that small clinical federations cannot currently get privacy that outlives their encryption without giving up substantial accuracy.
References
- The Distributed Discrete Gaussian Mechanism for Federated Learning with Secure Aggregation
- Efficient Full-Stack Private Federated Deep Learning With Post-Quantum Security
- Buffalo: A Practical Secure Aggregation Protocol for Buffered Asynchronous Federated Learning
- Practical Challenges in Differentially-Private Federated Survival Analysis of Medical Data
- Post-Quantum Cryptography (PQC) Network Instrument: Measuring PQC Adoption Rates and Identifying Migration Pathways
- PILLAR: How to make semi-private learning more effective
- The Discrete Gaussian for Differential Privacy
- Improving ML Attacks on LWE with Data Repetition and Stepwise Regression