Proactive Handoff of Secret-Shared Models Under Client Churn

Some federated learning systems never let anyone see the model during training. The weights are split into secret shares, random-looking pieces held by a rotating committee of clients, and only the final model is ever reconstructed. This matters because the intermediate models that ordinary federated learning reveals at every round are what the strongest data-reconstruction attacks feed on. Attacks that link updates across rounds sharpen their reconstructions, and some learn an image prior from the stream of updates alone. Hiding the intermediate models is the main point of secret-shared training.

Cross-device committees churn, though. When a client leaves and hands its piece to a newcomer, the pieces from before and after the handoff still lie on the same hidden polynomial. An attacker who quietly compromises one client per week can then collect enough pieces over time to rebuild a past intermediate model. Once it has that model, it can run the usual gradient-inversion attacks. We propose to show that this mobile-attacker failure is real and measurable in pixels. We then repair it with proactive secret sharing, a cryptographic technique that re-randomizes every piece at each handoff without changing the secret. Finally, we show that the repair is cheap enough for model-sized state when the pieces are batched.

Core Claim

In multi-round secret-shared training with a changing committee, a naive handoff silently turns a $t$-out-of-$n$ privacy guarantee into a guarantee that holds only until the attacker has spent about $(t+1)/c$ epochs corrupting $c$ clients per epoch. After that point, recovered intermediate models let existing inversion attacks reconstruct client data.

Proactive refresh closes this gap. At each handoff the committee jointly adds a random sharing of zero to every piece. In Shamir sharing, each piece $f(i)$ is a point on a random polynomial $f$ whose value at zero, $f(0)$, is a model weight. Refresh replaces it with

$$h(i) = f(i) + g(i), \qquad g(0) = 0,$$

so $h(0)=f(0)$ but the new pieces are independent of the old ones, as in A Novel Proactive Multi-secret Sharing Scheme and DSO: Dependable Signing Overlay.

The new element is not the refresh formula. It is applying refresh to the model and optimizer state of a training run, at the scale of $10^5$ to $10^6$ secrets per round. It also means evaluating the result with gradient-inversion attacks rather than with abstract share counts.

Where This Sits

Two lines of work motivate the proposal.

Beyond Secure Aggregation: Scalable Multi-Round Secure Collaborative Learning (CLOVER) argues that revealing intermediate models causes cumulative leakage. It keeps them secret-shared using Lagrange coding with a degree-reduction step, at linear communication cost. It assumes a fixed committee and a static attacker.

On the attack side, three papers show that the intermediate-model stream is valuable. AGIC links samples recurring across epochs and reports up to 50% higher reconstruction quality. GIML in Gradient Inversion with Generative Image Prior meta-learns a data prior from a sequence of updates. mGAN-AI isolates one client's idiosyncrasies from successive updates. These attacks are the payoff an attacker gets by recovering hidden models, and we reuse them as our measuring instrument.

Two papers supply the method. CHURP gives dynamic-committee proactive sharing with an efficient bivariate zero-sharing step. Its dimension switching tolerates up to $2t$ corruptions during handoff, but it was built only for blockchain key custody and handles one secret at a time. DMM shows that packed secret sharing can pass a long-lived secret state across changing federated committees at $O(1)$ amortized cost per secret. DMM applies this only to differential-privacy noise state and leaves the model exposed.

Our construction carries DMM's packed resharing over to the model weights themselves, adds CHURP-style zero-sharing at each handoff, and keeps CLOVER's training pipeline.

One body of work constrains the design. Continual Leakage-Resilient Dynamic Secret Sharing in the Split-State Model warns that the refresh computation can itself leak. We therefore state a threat model that excludes side channels during refresh and treat leakage during refresh as out of scope.

LIA-SA breaks secure aggregation by sending crafted "fishing" models. It does not apply directly here, because the server never sends a model it can see in the clear. We will state this rather than claim it as a contribution.

Research Program

Analysis of existing protocols. The first step decides whether the problem exists in current protocols. We will analyze CLOVER's degree-reduction step to determine whether re-encoding into fresh lower-degree polynomials already re-randomizes the model pieces each round. If it does, CLOVER is proactive with respect to rounds, and the gap narrows to committee handoff alone. We will write down the joint distribution of pieces held by a mobile attacker across rounds in three cases: naive handoff, where the departing member passes its piece; handoff by resharing without zero-masking; and full proactive handoff.

Construction. We will implement four handoff variants on top of a CLOVER-style training loop:

  • naive transfer of pieces;
  • per-secret Shamir resharing with zero-polynomial refresh;
  • packed resharing in the style of DMM, extended to weights and optimizer moments;
  • packed resharing with CHURP's dimension switching enabled during transition windows.

The refresh period is a tunable parameter, from every round to every $k$ rounds.

Measuring leakage. We will build an end-to-end pipeline. A simulated mobile attacker corrupts $c$ committee members per epoch under a configurable churn rate. It reconstructs any intermediate model for which it holds $t+1$ consistent pieces. It then runs AGIC-style multi-round inversion, plus a pixel-matching attack, on the recovered model differences. The outputs are reconstruction PSNR and SSIM as functions of epochs of corruption.

Baselines

Plain secure aggregation (the naive reference). Masked summation in the style of vFedSec reveals every round's aggregate model to the server. It is the default deployment, so it sets the ceiling on leakage. We will run it on identical data, models and rounds, and attack the revealed models with the same inversion attacks. The file gives no reported inversion number for this exact configuration on Lacuna. Relevant reference points from other papers are:

We will reproduce these attacks ourselves rather than compare across papers.

Secret-shared training with naive handoff. This is CLOVER as published, plus the simplest way of handling churn. It is the strongest existing competitor on privacy under a static attacker. No reported result was found on Lacuna for CLOVER's accuracy or leakage on CIFAR-10. Beating it means showing that under mobile corruption its intermediate models become recoverable while ours do not, at matched final accuracy.

DMM-style distributed differential privacy. This is the strongest existing method that already uses packed resharing in federated learning. On SO-NWP it reports accuracy 5 to 6 percentage points above the earlier distributed mechanism, DDG, and parity with central differential privacy, at about 5.7 to 25 MB of communication per client (DMM). We will run DMM at several privacy budgets. We will tune both methods with the same learning-rate sweep and round budget. We will then compare inversion success on DMM's public intermediate models with inversion success against our hidden models, at equal final accuracy.

Our proposal beats DMM if, at DMM's accuracy, our reconstructions stay at the level of the attack's no-information prior while DMM's do not at moderate budgets. It also must keep per-client handoff traffic within the same order of magnitude as DMM's 5.7 to 25 MB band.

Ablations. Three ablations isolate the key ingredient:

  • Removing zero-masking while keeping packed resharing tests whether masking, rather than resharing alone, prevents accumulation.
  • Disabling dimension switching tests whether the transition window is where leakage concentrates.
  • Sweeping the refresh period traces the tradeoff between cost and leakage.

Why not just…?

Why not just re-run key setup every few rounds, as Efficient Vertical Federated Learning with Secure Aggregation recommends? Re-running setup refreshes masking seeds, which nobody needs to preserve. A secret-shared model is a secret that no single party knows and that must survive the refresh. Sharing it afresh would require someone to reconstruct it first, which defeats the purpose. Preserving a secret while replacing all its pieces is exactly what proactive resharing does. The real question is cost, which the packed-versus-unpacked ablation answers.

Why not just add differential privacy and publish intermediate models? This is a genuine threat. If calibrated noise makes each released model safe, hiding it adds little. The evidence for a utility tax is mixed:

The DMM baseline settles this by comparing the two at matched accuracy. If DMM at a useful budget already holds inversion to the no-information level, our contribution reduces to the regime of loose or no budgets, and we will say so.

Why not just trust that degree reduction already re-randomizes the pieces? This is possible, and it is the first thing we check. If CLOVER's re-encoding is implicitly proactive across rounds, the remaining vulnerability is handoff under churn. The naive-handoff baseline isolates that case directly.

Proposed Experiments

Setting. Following CLOVER's evaluated configuration, we train a two-layer network on frozen VGG features for CIFAR-10 and on Fashion-MNIST, and add FEMNIST for realistic per-writer partitioning. Committees have $n$ between 20 and 100, threshold $t=n/3$, churn of 5 to 30% per epoch, and $c$ from 1 to 3 corruptions per epoch.

First decisive check. With naive handoff, attacker model recovery should occur at about $(t+1)/c$ epochs. From then on, AGIC-style inversion on the recovered model differences should approach the PSNR achieved against plain secure aggregation. With proactive handoff, recovery should never occur, and reconstruction quality should stay flat at the prior-only level for the whole run.

Second decisive check. Packed proactive handoff should move the full weight and optimizer state across a committee change within a small multiple of DMM's per-client traffic. Final accuracy should be identical to an unrefreshed run, since refresh does not change the secret.

Failure conditions. Either of two outcomes would sink the proposal. Naive handoff might not lead to recovery in practice, for example because degree reduction already refreshes and churn is always handled by full resharing. Alternatively, DMM at matched accuracy might already defeat the inversion attacks.