A new class of models maps between biological modalities: Omni-DNA, an autoregressive genomic foundation model, generates images and functional annotations directly from DNA sequences, reaching a 0.987 Macro F1 score when mapping DNA motifs to image tokens on its Needle-in-DNA benchmark Omni-DNA; ST-Net models predict the expression of hundreds of genes from histology image patches PPML-Hybrid. These models sit exactly where privacy risk is worst, because genomes are permanent, uniquely identifying, and implicate relatives Fellay et al.. Meanwhile, the membership inference literature — attacks that determine whether a specific record was in a model's training set — has developed powerful cross-modal attacks: measuring how tightly an overfitted model binds a training image to its training caption (the feature-based attack of M4I, up to 94.83% success), or re-feeding a model's own generated text back to it and reading the token-level entropy (the MaxRényi-K% attack of VL-MIA, AUC 0.815 on GPT-4). Nobody has ever run these attacks on a biological cross-modal model. The utility-focused genomic modeling literature contains zero privacy evaluation, and the attack literature has never left natural image–text pairs.
This proposal does two things about that gap. First, it builds the first cross-modal membership inference benchmark for genomic and spatial-transcriptomic models, adapting the strongest published attacks to Omni-DNA and ST-Net settings. Second, it designs and tests a defense targeted at where these attacks actually read the signal: not the final classification softmax, but the cross-modal alignment — the projection and decoding machinery where DNA features and image/text features are pulled into correspondence. The defense places only that alignment "bridge" inside a hardware enclave (a trusted execution environment — a physically isolated region of the processor that decrypts and computes on data in plaintext while shielding it from the host operating system), leaving the large unimodal encoders outside, and combines this with homomorphic-encryption-based aggregation for federated training, following the pattern that made SAFETY run 1.93 to 4.82 times faster than pure enclave execution SAFETY.
Core Claim
The claim has a measurement half and a defense half. Measurement: cross-modal membership inference transfers to biological models, and leaks donor or tissue-sample membership through predicted gene expression and generated annotations. Defense: because the strongest cross-modal attacks observe latent alignment and intermediate token logits rather than final softmax outputs, enclave placement must shield the cross-modal bridge — the alignment head and decoder — rather than the network tail that DarkneTZ isolated for unimodal classifiers. Concretely, write the model as a composition
$$ f = d \circ \pi \circ (e_{\mathrm{img}}, e_{\mathrm{dna}}), $$
where $e_{\mathrm{img}}$ and $e_{\mathrm{dna}}$ are unimodal encoders, $\pi$ is the cross-modal alignment/projection, and $d$ is the generative decoder. The proposal runs $e_{\mathrm{img}}$ and $e_{\mathrm{dna}}$ in the untrusted rich environment and $\pi$ plus the sensitive tail of $d$ inside the enclave, so that cross-modal latents and intermediate logits never leave hardware isolation. The hypothesis — falsifiable, and the placement-sweep experiment below is designed to settle it — is that the bridge, not the tail, is the leaky channel. This is exactly the open question left by DarkneTZ, which showed last-layer isolation drops white-box attack precision from roughly 85% to 50% (random) at only 3% CPU overhead, but only for unimodal classifiers whose attacks read the softmax DarkneTZ.
An honest scoping note up front: enclave shielding removes the attacker's observation channel (latents and logits across the bridge); it does not by itself make the released outputs private. The haplotype-based attacks of Haplotype-based membership inference from summary genomic data prove that even carefully aggregated genomic outputs leak — 475 of 1,350 target haplotypes received posterior membership probability 1.0 — so this proposal's defense claim is scoped to neutralizing latent- and logit-based cross-modal attacks at near-native speed, with output-level leakage measured and reported rather than assumed away.
Research Program
The first workstream is the attack benchmark. The feature-based attack of M4I — train a shadow encoder projecting both modalities into a shared latent space, then measure how abnormally close training pairs sit — is the natural instrument, since it needs no ground-truth captions and already achieved 82% accuracy on clinical chest X-ray report generation, the closest biological setting published M4I. We adapt it twice: on the 10x Genomics Visium human breast cancer spatial transcriptomics dataset (30,612 spots across 23 patients, predicting the top 250 genes from histology patches), measuring whether predicted expression vectors reveal that a tissue spot or an entire patient was in training; and on Omni-DNA's Needle-in-DNA and DNA2Function tasks, measuring whether generated image tokens or RNA-type annotations reveal sequence membership. The re-feeding protocol with MaxRényi-K% entropy slicing VL-MIA provides a second, logit-based instrument; we expect it to transfer well because autoregressive genomic decoders output token distributions just like vision-language models. One prediction worth stating: leakage per patient may be higher than in natural-image settings, because 23-patient cohorts invite heavy overfitting — but the coarse patient-level granularity (23 members vs. thousands) may cut the other way; the benchmark resolves this empirically.
The second workstream is the defense. We implement the partition above on Arm TrustZone with OP-TEE, reusing DarkneTZ's world-switching machinery but placing the boundary at the bridge DarkneTZ. Feasibility follows from size asymmetry: unimodal encoders dominate parameter counts (Omni-DNA spans 20M to 1B parameters Omni-DNA), while alignment heads and decoder tails are small enough for the roughly 14MB of usable secure memory on edge TrustZone configurations, or the 128MB enclave page cache of SGX class hardware. Two known TEE weaknesses are addressed directly: memory access patterns are masked with the fixed 4KB page-chunking strategy PRESAGE used for oblivious access PRESAGE, and the API never returns intermediate logits — only final generated outputs — which is precisely what denies the re-feeding attack its input.
The third workstream covers federated training across institutions, where the enclave alone scales badly. SAFETY showed that pure-SGX federation degrades as data owners accumulate, because each contributor forces another secure channel and in-enclave decryption, while adding encrypted counts homomorphically before enclave entry restores flat scaling SAFETY; PPML-Hybrid showed the same stabilizing effect of noise-free CKKS-encrypted updates on image-to-transcriptomic regression, with an HE-client ratio of 0.5 to 0.8 approaching the all-HE baseline's error PPML-Hybrid. We combine both: clients send CKKS-encrypted gradients, the untrusted aggregator homomorphically sums them, and the enclave performs a single decryption of the aggregate before aplaintext update — never exposing any individual site's update. This deliberately avoids the alternative of running training under pure FHE, whose costs are documented below.
Proposed Experiments
The primary testbed is the Visium breast cancer dataset with ST-Net, because it is the one setting in the file where a real biological cross-modal model, real data, and a published privacy-preserving training pipeline all coexist PPML-Hybrid. The secondary testbed is Omni-DNA's Needle-in-DNA and DNA2Function benchmarks with released checkpoints at the 20M-to-1B scale Omni-DNA. Attacks are run at matched query budgets against every defense; metrics are attack AUC and success rate (and TPR at low false-positive rates, since MrM showed low-FPR true-positive rate is where defenses matter most MrM), against utility (MSE on the 250-gene regression; Macro F1 on Needle-in-DNA) and wall-clock overhead relative to fully untrusted execution. For the federation experiments we measure runtime as a function of the number of contributing sites, directly against SAFETY's reported pattern of pure-SGX degradation SAFETY. The decisive check for the defense is a placement sweep — nothing in the enclave, DarkneTZ-style tail only, bridge only, bridge plus decoder tail — run under identical attacks: it isolates whether the bridge is the load-bearing ingredient, and it will also tell us if the simpler tail-only placement suffices, in which case we report that instead. The decisive check for the attack benchmark is calibration: attacks must beat their random baseline (AUC 0.5 / 50% success) with confidence intervals that survive correction for the small patient count.
Baselines
The "do nothing" reference point is the undefended model served through a standard API that returns full outputs and, in the permissive setting, token logits and latents. The strongest published numbers for this regime come from M4I: the feature-based attack reached 94.83% success in the unrestricted threat model on natural image-captioning data and 82% accuracy on the clinical R2GenCMN chest X-ray generator M4I; VL-MIA's MaxRényi-K% reached an AUC of 0.815 on GPT-4 VL-MIA. No reported result found on Lacuna for cross-modal membership inference on any genomic or spatial-transcriptomic model, so on our testbeds the undefended number must be established by this project — the M4I clinical figure of 82% is the closest anchor and the level we expect to approach or exceed.
The strongest existing defense at feasible cost is DarkneTZ-style tail partitioning, which dropped white-box membership-inference precision from roughly 85% to about 50% while adding only 3% CPU overhead (about 10% when the enclave's full memory budget is used) on CIFAR-100 and ImageNet classifiers DarkneTZ. It is the strongest competitor because it is the only published hardware defense with a demonstrated near-random attack result at single-digit overhead, and it is the natural "just port it" answer to this proposal. We run it on identical hardware, identical enclave software, identical datasets, and an identical attack suite and tuning budget; our defense counts as beating it only if the bridge placement suppresses the cross-modal attacks (feature-based and re-feeding) measurably below what tail placement achieves at comparable or lower overhead. Heuristic defenses — L2 weight regularization and aggressive augmentation, which M4I found inconsistent, and the prompt-level filters that MrM bypassed M4I, MrM — are included as weak reference points under the same protocol. Differential privacy is included not as a straw man but as the strongest theoretical-privacy competitor: DP-SGD at several privacy budgets, with the utility-privacy curve reported in full, since M4I found DP acts as a hard defense at severe ROUGE-L cost and PPML-Hybrid found DP-only omics regression markedly worse than noise-free hybrid training M4I, PPML-Hybrid. The key ablation isolating this proposal's ingredient is the placement sweep described above: same model, same enclave, same attacks, only the partition boundary moved.
Why not just…?
Why not just use differential privacy and skip the hardware entirely? This is the most serious alternative, and we treat it as such rather than arguing around it. The published evidence says DP pays a steep utility price in exactly our settings: M4I reports severe ROUGE-L collapse when DP defends generative captioning M4I; PPML-Hybrid reports significantly higher MSE for DP-only spatial-transcriptomic regression, with errors recovering only when noise-free encrypted updates anchor the model PPML-Hybrid; and the haplotype attack shows that noise calibrated for independent variants fails against linkage-disequilibrium structure Haplotype-based membership inference from summary genomic data. But this evidence is about generation and linked variants — a reviewer could reasonably hypothesize that moderate noise suffices for a 250-gene regression. Our DP-SGD sweep at multiple budgets on the same testbed settles this head to head, and if a modest budget matches our hardware defense's privacy at acceptable MSE, the hardware case weakens and we will say so.
Why not just put the whole model inside the enclave? enclave memory makes this infeasible: PRESAGE hit the 128MB SGX page-cache ceiling and the paging penalties that come with exceeding it PRESAGE, and DarkneTZ's TrustZone offered roughly 14MB of usable secure memory DarkneTZ — nowhere near a 1B-parameter model. SAFETY further showed pure-enclave federation degrades linearly in the number of contributing sites SAFETY. Partitioning is not an ornament here; it is what makes the defense fit in hardware at all.
Why not just run everything under homomorphic encryption? encryption lets a server compute on data without seeing it, but the cost is documented at every turn in this literature: PRESAGE measured homomorphic methods roughly 120 times slower than enclave execution for genomic queries PRESAGE; non-linear activations force polynomial approximation and bootstrapping cycles that dominate runtime SAFENet, SecureNets] — and our models add attention softmax on top; and ciphertext expansion can push a single deep-inference request to tens of gigabytes of transfer Haralampieva. We use encryption only where it is cheap and strong: additive aggregation of gradients before a single enclave decryption, the exact role SAFETY and PPML-Hybrid validated SAFETY, PPML-Hybrid.
Why not just port DarkneTZ's last-layer trick as-is? This is the threat we take most seriously, because if tail-only placement neutralizes cross-modal attacks at 3% overhead, the bridge-placement hypothesis is unnecessary. The counter-argument is mechanistic, not rhetorical: the feature-based attack reads latent alignment between modalities, which is computed in the projection head upstream of any tail layer M4I, and the re-feeding attack reads token logits of intermediate description segments, not final softmax scores VL-MIA — both channels survive a tail-only boundary. But this is precisely the open empirical question, which is why the placement sweep is the decisive experiment rather than an afterthought, and why its result will be reported whichever way it falls.
Why not just avoid individual-level models and publish aggregate statistics? the haplotype attack closed this exit: frequency tables and Beacon responses yielded posterior membership probabilities of 1.0 on 475 haplotypes and enabled reconstruction of 406 novel haplotypes Haplotype-based membership inference from summary genomic data, and cross-modal methods like S-PrediXcan invert summaries back to individual-level quantities S-PrediXcan. Aggregation reduces, but does not eliminate, the attack surface — which is why this proposal measures output-level leakage explicitly rather than claiming the enclave makes it moot.
Decisive Checks and Honest Failure Modes
Three outcomes would falsify or reframe the proposal, and each has a named experiment. If the undefended cross-modal attacks land near chance on both biological testbeds — plausible if genomic pretraining over 30 billion nucleotides dilutes fine-tuning memorization, given VL-MIA's pre-training detection AUC of only about 0.688 VL-MIA — then the defense is unmotivated, and the benchmark's contribution becomes a documented negative result with calibration analysis. If the placement sweep shows tail-only placement matches bridge placement, the DarkneTZ port wins and we report it. If unimodal encoder activations alone carry membership signal across the world-switch boundary, bridge shielding is insufficient by construction, and the fix — shrinking what crosses the boundary or encrypting the crossing — becomes the follow-up. Success, by contrast, is concrete: feature-based and re-feeding attacks suppressed toward the random baseline on both testbeds at overhead within the DarkneTZ envelope of roughly 3 to 10 percent, with utility within a few percent of undefended training, and federated runtimes that stay flat as institutions join — a demonstration that the leakiest component of a cross-modal genomic model can be made private at the cost the TEE literature has already shown to be affordable.
References
- Omni-DNA: A Unified Genomic Foundation Model for Cross-Modal and Multi-Task Learning
- A Privacy-Preserving Federated Learning Method with Homomorphic Encryption in Omics Data
- Practical Solutions for Protecting Individual Genomic Privacy
- M$^4$I: Multi-modal Models Membership Inference
- Membership Inference Attacks against Large Vision-Language Models
- SAFETY: Secure gwAs in Federated Environment through a hYbrid Solution
- DarkneTZ: towards model privacy at the edge using trusted execution environments
- Haplotype-based membership inference from summary genomic data
- PRESAGE: PRivacy-preserving gEnetic testing via SoftwAre Guard Extension
- MrM: Black-Box Membership Inference Attacks Against Multimodal RAG Systems
- SAFENet: A Secure, Accurate and Fast Neural Network Inference
- SecureNets: Secure Inference of Deep Neural Networks on an Untrusted Cloud
- A Systematic Comparison of Encrypted Machine Learning Solutions for Image Classification
- Exploring the phenotypic consequences of tissue specific gene expression variation inferred from GWAS summary statistics