Tamper-Evident Measurement Invariance Audits Across Data Silos

Before anyone concludes that a survey instrument, a stress model, or a risk score behaves differently for women and men, or for old and young people, they first have to show that the measuring instrument means the same thing in each group. Psychometricians call this measurement invariance and test it with multi-group structural equation modeling. This method fits one latent-variable model per demographic group and asks whether constraining the groups to share parameters makes the fit noticeably worse. In practice the data for such audits sits in separate hospitals, insurers, or agencies that cannot pool item-level responses. Some of those institutions also have a stake in the answer. They may want a disparity to disappear, or to appear.

This note argues that the cryptographic core of the problem is smaller than it looks, and that the genuinely hard parts lie elsewhere. The invariance decision depends on the data only through per-group counts, means, and covariance matrices. These are integer sums when the items are Likert-scored. So the computation that must be protected is a secure sum, not iterative model fitting under encryption. Three problems remain open. First, the approximate encryption scheme most HE statistics work uses must add noise at decryption, and that noise can swamp the tiny fit-index changes psychometric decisions rest on. Second, a dishonest site can quietly flip an invariance verdict by submitting malformed or strategically chosen contributions. Third, sites that will not agree on a shared trusted setup need a one-shot protocol, and the only known route to one with malicious security runs through superpolynomial-time simulation. We propose to build an exact-arithmetic audit protocol, measure how much a cheating site can move the verdict, and show that bounding that influence is exactly what a simulation-based security guarantee buys.

Core Claim

Our claim has three parts, each testable. First, distributed configural, metric, scalar, and strict invariance testing can be made numerically identical to centralized lavaan output by securely aggregating per-group moment sums with an exact lattice scheme. BFV is such a scheme: it computes on encrypted integers modulo a large number, with no rounding. Because it computes exactly, it sidesteps the decryption-leakage attack on approximate encryption and the protective noise that attack forces.

Second, when a participant is malicious, the right question is not "is the computation correct?" but "how large a verdict change can a participant force while its input still looks like a valid dataset of its declared size?" We call this quantity the manipulation margin and will measure it empirically.

Third, in a two-message protocol with no trusted setup, the only way to guarantee that a cheater's influence is capped by that margin is a simulator that can extract the cheater's effective input. For two rounds in the plain model, that requires a superpolynomial-time simulator.

Where This Sits in Prior Work

Three lines of work motivate the proposal. Short version of the smartphone addiction scale: Measurement invariance across gender runs the full four-level invariance hierarchy on 1,112 students and accepts invariance on changes of ΔCFI < 0.004 and ΔRMSEA < 0.013. These margins are what any privacy mechanism must preserve. Gender-specific perception of job stressors and resources shows why pooling fails: on 4,118 employees, workload predicts stress at β = 0.33 for men and 0.23 for women, and working hours matter only for women. Both studies need centralized microdata, and our protocol removes that need.

Two lines supply the method and constrain it. Collaborative Privacy-Preserving Analysis of Oncological Data using Multiparty Homomorphic Encryption runs survival analysis and logistic regression across hospitals with multiparty approximate encryption. It assumes honest-but-curious parties and never tests invariance. We reuse its multiparty key and decryption structure but replace approximate arithmetic with exact arithmetic and add a malicious model. Securing Approximate Homomorphic Encryption Using Differential Privacy proves that approximate (CKKS) decryption leaks the secret key unless Gaussian noise is added to outputs. It also proves that noise calibrated on the fly still leaks. This is the negative result that pushes us to BFV rather than to a cleverer CKKS noise schedule. FHAIM injects differential-privacy noise inside CKKS to train synthetic-data generators, including on COMPAS. We share its concern about outputs but target a hypothesis test, not a generator, and rely on its trusted decryption service in no way.

The superpolynomial-simulation line sets the security bar. How To Play Almost Any Mental Game Over The Net introduced simulators allowed more time than the adversary as a way to get composable security without setup. Succinct Non-Interactive Secure Computation combines fully homomorphic encryption, delegation proofs, and superpolynomial extraction into a maliciously secure two-round protocol with no common reference string, under subexponential LWE. It is a feasibility result for arbitrary circuits. We do not claim to make it practical. Instead we specify the concrete functionality an invariance audit needs, which is moment aggregation with an input-validity predicate. We show that this functionality's circuit is shallow enough that the distance between feasibility and deployment becomes a measurable question. The Concrete Security of Two-Party Computation offers input indistinguishability as a simpler definition, but proves it equivalent to simulation only for invertible functionalities. Aggregated moments are not invertible, so that shortcut does not apply here.

Research Program

The first component is a reduction plus an exactness proof. Maximum-likelihood multi-group factor analysis with mean structure depends on group $g$'s data only through its count $n_g$, its sum vector, and its cross-product matrix. For $k$ Likert items coded $0$ to $L$, all of these are integers bounded by $n_g L^2$. We encode each site's per-group contributions as BFV plaintexts, sum the ciphertexts across sites, and threshold-decrypt only the pooled per-group moments. These pooled moments are exactly what published MG-SEM papers routinely report. The four nested models are then fit in plaintext. For intersectional groups too small to release safely, we add a variant that computes only the group-count check under encryption and suppresses release below a threshold. Leakage in this variant is limited to what the fitted models themselves reveal.

The second component is the adversarial analysis. A malicious site can do two things. It can submit ciphertexts that decrypt to impossible values, such as negative counts or cross-products inconsistent with the sums, and nothing in honest-but-curious protocols stops this. Or it can submit a well-formed but fabricated dataset. We define the ideal functionality to accept only inputs satisfying a validity predicate: item values lie in range, the declared $n$ is consistent, and the covariance is positive semidefinite. Against the first attack, validity checking is decisive. Against the second, no protocol can do better than the manipulation margin, the smallest fraction of fabricated records that flips the scalar-invariance decision. We will compute this margin by adversarial optimization over valid synthetic records.

The theoretical deliverable is a proof sketch that the one-shot, no-setup version of this functionality can be realized by instantiating the Morgan–Pass–Polychroniadou construction. The key link is that superpolynomial extraction is what lets a security proof say "the cheater's effect equals that of some valid input." That statement turns the cryptographic guarantee into the empirical margin.

The third component is the noise analysis that justifies exact arithmetic. For the CKKS alternative, we compute how much the IND-CPAD-mandated output noise perturbs ΔCFI and ΔRMSEA. We then measure how often invariance decisions flip relative to the exact reference, across group sizes from roughly one hundred to several thousand.

Baselines

The naive reference is centralized plaintext lavaan/semTools on pooled microdata. It defines the correct verdicts and fit indices, and the protocol must match it on every decision and on fit indices to printing precision.

The strongest existing encrypted competitor is the multiparty CKKS pipeline of Geva et al., run with output noise sized per Li et al. It will be applied to the same moment sums, on the same hardware and ring dimension, with its noise tuned as favorably as the security proof allows. Geva et al. report results that "matched cleartext results up to 7 decimal digits" and logistic regression taking "about 500 seconds" for 100 iterations. Those figures are without IND-CPAD noise. We beat this baseline if BFV matches centralized verdicts in every configuration where CKKS-with-required-noise flips at least some, at comparable or lower wall-clock time.

The local-privacy baseline has each site add differential-privacy noise to its moments before pooling, at ε = 1 to match FHAIM's setting. Our protocol beats it if the noisy version's decision error rate exceeds ours at realistic group sizes.

The pooled-covariate baseline is a single regression with gender dummies, run on the same data. It shows what is lost when invariance and moderation are never tested.

Two ablations isolate the key ingredients. Validity checking on versus off isolates the malicious-security ingredient. BFV versus CKKS at matched parameters isolates the exact-arithmetic ingredient.

Why not just…?

Why not just have each site send its per-group moments over TLS to a trusted analyst? This is a real threat, because our protocol is at heart secure aggregation of those same moments. The difference appears when sites hold overlapping demographic groups: per-site moments of small groups then leak more than pooled ones. It also appears when no analyst is trusted by all sites. If the experiments show that per-site moments at realistic sizes leak nothing beyond pooled moments, we will say so and narrow the claim to settings where a single honest analyst does not exist.

Why not just add differential privacy locally and run lavaan centrally? Invariance decisions hinge on ΔCFI shifts below 0.004. The local-DP baseline directly measures how often ε = 1 noise pushes verdicts across that line. We expect it to fail at the few-hundred-per-group sizes typical of the cited studies, but this is an empirical question the experiment settles.

Why not just use CKKS, as in existing HE statistics work? Li et al. show that CKKS outputs need Gaussian noise for security and that on-the-fly noise estimation still leaks. The data here are integers, so CKKS's floating-point convenience buys nothing, and the BFV-versus-CKKS ablation quantifies the cost of the noise.

Why not just assume a trusted setup or common reference string and use standard simulation? For institutions that will accept a setup ceremony, this is the right engineering choice. We will benchmark such a variant as the deployable path. Superpolynomial simulation matters only for the one-message, no-setup regime, where Goldreich–Oren rules out standard simulation. We treat it as the theoretical guarantee that justifies interpreting the manipulation margin as a hard bound, not as the implementation.

Why not use input indistinguishability for simpler proofs? Bellare et al. establish its equivalence to simulation only for invertible functionalities. Pooled moments discard information by design.

Proposed Experiments

Real item-level data from the cited cohorts may not be public, and we do not assume access. Our primary testbed is therefore synthetic populations calibrated to the reported structure. For the scale study, this means one factor, 529 and 583 per group, and CFI near 0.985. For the workplace study, it means 2,059 per gender with differing structural paths. For each, we generate matched invariant and non-invariant ground truths. Where item-level data can be obtained, we will rerun on it. COMPAS (7,214 records) serves as a public secondary case, auditing group-conditional error rates, which are also functions of aggregated counts.

Data are split across three to ten simulated sites with both clean and skewed demographic mixes. We report four metrics: decision agreement with the centralized verdict at each invariance level, absolute error in CFI and RMSEA, runtime and communication, and manipulation margin with and without validity checking.

The decisive checks are three. BFV must reproduce every centralized decision exactly. CKKS with the security-mandated noise must flip a nonzero fraction of decisions at small group sizes; if it does not, the exactness argument is unnecessary and we drop it. Without validity checks, a single site must be able to flip scalar invariance with one malformed contribution, while with checks the required fabricated fraction should approach the statistical margin. If this margin turns out to be tiny even for honest-format inputs, that is itself the headline finding. It would mean invariance verdicts across silos are fragile to any participant regardless of cryptography, and audit design must require larger groups rather than stronger encryption.