Hospitals that want to train a shared model predicting which patients with mild cognitive impairment will progress to Alzheimer's disease face a security problem. Existing encrypted aggregation schemes hide each site's model update from the coordinating server. They also assume every site follows the protocol. A single misbehaving site can submit an encrypted update that is malformed or enormously scaled, and nobody can see it until it has already corrupted the global model. The standard fix is to attach a zero-knowledge proof that the update is well-formed. Proving that such a proof keeps the protocol secure requires an imaginary "simulator" that can pull each cheating site's actual input out of its proof. Classical proofs do this by rewinding: they rerun the cheater from a midpoint. That argument breaks when many training rounds and studies run at once. Our proposal avoids the problem with one observation. If each site first compresses its update to a vector of signs (+1 or −1), as an existing linear-cost encrypted aggregation scheme already does, then "well-formed" reduces to "every coordinate is ±1." That property is cheap to prove, and the proof can be built so the simulator reads the input off the proof in a single forward pass, with no rewinding.
In two sentences: we take a linearly scaling, threshold-homomorphic aggregation protocol that only resists honest-but-curious participants. We add to each site's upload a proof that its compressed update contains only signs, and the proof is extractable without rewinding. Together this bounds any site's influence to one vote per coordinate, removes the need for a trusted clean dataset on the server, and keeps security when sessions run concurrently. We test whether this costs little enough to train a multimodal model of MCI-to-AD conversion across simulated ADNI sites.
Core Claim
The closest existing system is RSA from Secure Aggregation in Federated Learning using Multiparty Homomorphic Encryption. Homomorphic encryption means encryption that allows sums to be computed on ciphertexts. RSA uses threshold homomorphic encryption: the decryption key is split among clients so that no single party can decrypt. It also uses a Sign-quantization Random Linear Compressor (SRLC), which projects each gradient to a lower dimension $d$ and keeps only the signs. After a one-time quadratic setup, each round costs $O(N)$ in the number of clients $N$. RSA reports 95.20% CIFAR-10 accuracy against 95.48% for uncompressed non-secure training. Its security, however, holds only against semi-honest parties, meaning parties that follow the protocol but try to learn from what they see.
We claim that SRLC's output format is what makes malicious-client security affordable. Each coordinate of a valid update $x \in \mathbb{Z}^d$ satisfies $x_i^2 = 1$. The relation each client must prove is therefore
$$ R = \{ (c, \mathsf{cm}) ; (x, r) \;:\; c = \mathsf{Enc}_{pk}(x; r),\ \mathsf{cm} = \mathsf{Commit}(x),\ x_i^2 = 1 \ \forall i \le d \}, $$
where $c$ is the client's homomorphic ciphertext, $\mathsf{cm}$ a polynomial commitment to $x$, and $r$ the encryption randomness. A valid proof caps any client's per-coordinate influence on the aggregate at $1/N$. This is the same bounded-vote structure as sign-majority training. No root dataset and no server-side similarity scoring are needed.
We make the proof straight-line extractable with the Commit-then-Extract compiler of Witness-Succinct Universally-Composable SNARKs. That compiler encodes the witness as polynomial coefficients, commits with KZG, and applies Fischlin's proof-of-work transform. The transform forces the prover to query a global random oracle, which is a shared idealized hash function. The simulator recovers the witness by reading those query logs. The claim to test is that, for this specific binary relation, the added overhead is linear in $d$ with a small constant. This would give universally composable (UC) security against static malicious clients. UC is a standard notion under which security survives arbitrary concurrent composition with other protocols. Per-round cost would remain $O(N + d)$.
Why straight-line, and what constrains us
The motivating evidence is On the Necessity of Rewinding in Secure Multiparty Computation. It builds explicit functionalities for which stand-alone security proven by a rewinding black-box simulator fails even under two concurrent sessions. Adaptive Concurrent Non-Malleability with Bare Public-Keys shows that rewinding extraction also fails when the adversary changes its statement after being restarted. A federated consortium invites exactly that behavior, because a cheating site can choose its next update in response to the global model. Endemic oblivious transfer in Endemic Oblivious Transfer via Random Oracles, Revisited supplies the same query-logging technique in a different primitive, and we reuse the idea.
If a global random oracle with observable queries is considered too strong, Non Observability in the Random Oracle Model offers an extractable commitment built only from oracle programming. We treat it as a fallback, not as the main path.
Two results constrain the design. First, On the Connection between Leakage Tolerance and Adaptive Security shows that simulation-based tolerance of even polylogarithmic leakage is equivalent to adaptive security. It also forces secret-key spaces almost as large as the message space. Threshold lattice keys reused across hundreds of rounds cannot meet that bound, so we claim only static corruption and make no leakage-resilience claim. Second, the KZG commitment inherits a trusted setup ceremony, the "Powers of Tau." We accept that setup as a consortium-run ceremony and flag it as a limitation.
Research Program
The build has three parts. The first is RSA's threshold BFV aggregation (BFV is an exact-integer homomorphic encryption scheme) with SRLC compression, re-implemented as the semi-honest base. The second is a proof gadget for $R$. Inside a single polynomial identity, it checks that the committed polynomial's coefficients are roots of $X^2 - 1$ and that the committed vector matches the packed BFV plaintext. The ciphertext-consistency check is the expensive part. We will measure it separately and consider proving consistency against a commitment to the ciphertext's linear decryption share instead. The third part is the Fischlin wrapper and a UC proof sketch that exhibits the straight-line simulator.
The analytical deliverable is the proof that the simulator extracts every corrupted client's $x$ from oracle logs within a single session. A corollary bounds poisoning: under the proof, any coalition of $t$ malicious clients shifts each aggregated coordinate by at most $t/N$.
The application regime is Alzheimer's progression modeling. Predicting Alzheimer's disease progression using multi-modal deep learning approach trains a recurrent model on longitudinal cognitive scores, cerebrospinal fluid amyloid and tau, MRI, and demographics, and reports 81% accuracy (AUC = 0.86) for MCI conversion. Cognitive-only input reaches about 75% (AUC = 0.83). Multimodal models matter here. Early role of vascular dysregulation on late-onset Alzheimer's disease based on multifactorial data-driven analysis reports that vascular signals from arterial spin labeling MRI are the earliest and most pronounced abnormality. Those signals are invisible to cognitive tests and require pooled imaging the sites cannot share.
We partition the ADNI cohort by acquisition site into simulated hospitals with naturally skewed demographics and protocols. That split is non-identically distributed, which matches the failure mode described in Assessing the Generalizability of ADRDs Subtypes Across Diverse Cohorts HABS‐HD and ADNI. A secondary workload is the brain-age convolutional network on UK Biobank from Secure Neuroimaging Analysis using Federated Learning with Homomorphic Encryption, which trains with CKKS-encrypted aggregation (CKKS is an approximate-arithmetic homomorphic scheme). That paper reports no degradation but gives no numeric error; no reported MAE was found on Lacuna for it.
Baselines
The naive reference points are centralized training on pooled ADNI data, using the multimodal recurrent model, and plaintext federated averaging over the same site split. The centralized figure the cited paper reports is 81% accuracy, AUC 0.86. We will rerun it ourselves on our preprocessing, because cohort filtering differs. Both references use the same architecture, epochs, and hyperparameter search budget as every secure variant. They set the accuracy ceiling and the cost floor.
The strongest competitor on cost and accuracy is RSA (Secure Aggregation in Federated Learning using Multiparty Homomorphic Encryption), run with identical BFV parameters, compression dimension $d$, and client counts. It is the most direct comparison because our system is RSA plus the proof. Any gap in accuracy or per-round time is therefore attributable to the proof alone. Beating RSA here means matching its accuracy within one point AUC and staying within a 2× per-round wall-clock factor. RSA stays strictly better when all clients are honest. The point is to pay little for malicious security.
The strongest competitor on robustness is RFLPA (RFLPA: A Robust Federated Learning Framework against Poisoning Attacks with Secure Aggregation). It weights packed-secret-shared updates by cosine similarity to a 200-sample clean server dataset. Under a 30% gradient-manipulation attack on MNIST it reports 95% accuracy, against 32% for plain averaging. We give RFLPA its best case: a root set drawn from held-out ADNI subjects at the largest site, with the same attack fractions. Beating it means equal or better AUC under attack without any root data, plus a margin when the root set comes from a demographically different site.
We include BatchCrypt (BatchCrypt: Efficient Homomorphic Encryption for Cross-Silo Federated Learning) and MicroSecAgg (MicroSecAgg: Streamlined Single-Server Secure Aggregation) as cost references. BatchCrypt reports a 23–93× speedup over stock FATE; MicroSecAgg reports being about 100× faster than Bonawitz et al. at 500 users. Neither defends against malicious clients, so they bound how much of our overhead is intrinsic to homomorphic aggregation.
Two ablations isolate the key ingredients. The first keeps the identical proof but uses plain Fiat–Shamir, which is extractable only by rewinding, instead of Fischlin. Its cost difference is the exact price of straight-line extraction. The second removes the proof entirely and lets an attacker submit scaled non-sign vectors, which shows what the validity predicate buys.
Why not just…?
Why not just use a Fiat–Shamir proof and argue sequential security? This is the most serious alternative, because the proof is identical and cheaper, and a consortium may in practice run rounds sequentially. Our answer has two parts. Concurrency is not hypothetical here. Sites run several studies at once, and a cheating site can adapt across restarts, which is the failure Yao et al. describe. More importantly, the Fiat–Shamir ablation measures the price directly. If Fischlin's overhead on the binary relation is small (we target under 1.5×), composable security is nearly free. If it is large, we will report that honestly and recommend the rewinding variant for sequential deployments.
Why not just use RFLPA with a public Alzheimer's dataset as the root set? This is a real threat, since ADNI subsets are available and would remove RFLPA's privacy objection. The question is whether cosine similarity to a root drawn from one population still separates honest skewed sites from attackers. Our cross-site root experiment settles it. If RFLPA with a mismatched root matches our robustness, the advantage of our design narrows to removing a data dependency and gaining a formal guarantee.
Why not just clip or detect outliers after decryption? The server only ever decrypts the sum. A single scaled update is invisible as an individual and dominant in the aggregate. The no-proof ablation demonstrates this failure on the ADNI split.
Why not just pairwise-masking secure aggregation with range checks? MicroSecAgg's own authors note its cost grows with vector length through group exponentiations, and it is proven only against semi-honest parties. Adding range proofs there would need the same extraction machinery we build. Homomorphic aggregation also lets us reuse the BFV pipeline for the survival-style statistics that Collaborative Privacy-Preserving Analysis of Oncological Data using Multiparty Homomorphic Encryption computes in oncology.
Proposed Experiments
We run 5, 10, 20, and 50 simulated sites on the ADNI site partition and measure four quantities. The first is client proving time and proof size as functions of $d$, with the prediction of linear growth in $d$. The second is server verification time per client. The third is end-to-end per-round wall clock against RSA. The fourth is held-out MCI-conversion AUC against the centralized and plaintext references.
The decisive check is a curve of AUC versus attacker fraction (0–30%) under scaling and sign-flip attacks for our protocol, RSA, RFLPA with matched and mismatched roots, and the no-proof ablation. Success requires two things. Our protocol must stay within two AUC points of its clean value at 30% sign-flip attackers, where RSA and the no-proof ablation collapse. And its honest-case cost must stay within 2× of RSA.
A theory check accompanies the measurements: the written straight-line simulator, validated by an implemented extractor that recovers every corrupted client's update from oracle logs in a single forward pass of the protocol. The UK Biobank brain-age run tests whether the overhead profile carries over to a convolutional model with about one million parameters.
References
- Secure Aggregation in Federated Learning using Multiparty Homomorphic Encryption
- Witness-Succinct Universally-Composable SNARKs
- On the Necessity of Rewinding in Secure Multiparty Computation
- Adaptive Concurrent Non-Malleability with Bare Public-Keys
- Endemic Oblivious Transfer via Random Oracles, Revisited
- Non Observability in the Random Oracle Model
- On the Connection between Leakage Tolerance and Adaptive Security
- Predicting Alzheimer’s disease progression using multi-modal deep learning approach
- Early role of vascular dysregulation on late-onset Alzheimer’s disease based on multifactorial data-driven analysis
- Assessing the Generalizability of ADRDs Subtypes Across Diverse Cohorts HABS‐HD and ADNI
- Secure Neuroimaging Analysis using Federated Learning with Homomorphic Encryption
- RFLPA: A Robust Federated Learning Framework against Poisoning Attacks with Secure Aggregation
- BatchCrypt: Efficient Homomorphic Encryption for Cross-Silo Federated Learning
- MicroSecAgg: Streamlined Single-Server Secure Aggregation
- Collaborative Privacy-Preserving Analysis of Oncological Data using Multiparty Homomorphic Encryption