acceptodds
Under review as a conference paper at ICLR 2027

Vec2DNA: Feature-Consistent Genomic Embedding Inversion Attacks and Task-Aligned Defense

Abstract

Genomic foundation models are increasingly produced through mean-pooled embeddings, yet it remains comparatively underexplored when it comes to what such vectors leak about the underlying DNA from a privacy perspective. We introduce Vec2DNA, a black-box embedding inversion framework that constructs single-nucleotide-perturbed anchor–variant pairs and exploits their known correspondence through a feature-level consistency objective. Rather than treating these variants as conventional data augmentation, Vec2DNA uses their pairwise structure to preserve internal representations at unchanged target positions while localizing the effect of sequence perturbations. Across four genomic embedding models on hg38 20-mer dataset, Vec2DNA consistently outperforms alternative inversion methods in reconstruction quality. In particular, it achieves up to a 14.5% relative improvement in Levenshtein similarity over standard inversion baselines. Our analysis shows that the structured correspondence between anchor–variant pairs provides the key additional signal through feature-level consistency. Beyond sequence reconstruction, we further show that genomic embeddings retain sufficient fine-grained information to reliably distinguish single-nucleotide variants and recover allele identity well above chance. Furthermore, we propose Task-Aligned Genomic Quantization (TAG-Q), a utility-aware defense that selectively preserves task-relevant embedding directions while suppressing information useful for inversion. TAG-Q substantially reduces inversion performance while retaining near-clean downstream utility on DNABERT-2, and achieves a similarly favorable privacy–utility trade-off on GENA-LM. Together, our results highlight the privacy risks of genomic embeddings and demonstrate complementary attack and defense mechanisms for controlling fine-grained genomic information leakage.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Genomic foundation models, Embedding inversion attacks

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_kBVGOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how much of the underlying DNA sequence is leaked by mean-pooled embeddings from genomic foundation models. The authors propose Vec2DNA, a black-box embedding inversion attack that substitutes 1 to 3 nucleotides in each training sequence to form anchor–variant pairs, queries the frozen target model for both embeddings, and trains a non-autoregressive encoder-only inverter. The novel component is a feature-level consistency loss that constrains the L2 distance between pre-head features of anchor and variant at slots where the ground-truth token IDs match, on the grounds that a single substitution can change BPE segmentation and break direct positional correspondence. The authors analyze allele-level and locus-level disclosure, and propose TAG-Q, a post-hoc defense that fits linear probes for the intended downstream tasks, builds a basis that puts those probe directions first, and then spends more quantization bits on directions that are both task-relevant and high-variance.

Critical strengths
  1. The paper correctly diagnoses why existing post-hoc defenses fail, and this diagnosis outlives TAG-Q itself. Section A.2 shows the two standard defenses failing in opposite directions rather than simply failing: Gaussian noise couples the two objectives, so on DNABERT-2 the attacker reaches the random baseline only once probing accuracy has collapsed from 0.81 to 0.63, while scalar quantization preserves utility even at one bit (0.77 on DNABERT-2, 0.93 on GENA-LM) yet floors inversion at 0.50 and 0.80.

  2. The cross-model spread changes what the field should be measuring. Encoder-Only reaches 0.8476 on GENA-LM and 0.8300 on GROVER but only 0.5771 on DNABERT-2, all under BPE, refuting the tokenization-based explanation of inversion resistance. The spread across encoders exceeds any gain this paper’s attack achieves, so no single-encoder evaluation can establish that a model is safe. Vec2DNA holds its advantage at both ends, separating attack strength from encoder resistance as independent axes.

Critical weaknesses
  1. The main comparison seems unfair: Vec2DNA was allowed four times as many queries as the baseline, and its own new idea explains only about a tenth of the gain. Every time the attack asks the target model for an embedding, that counts as one query. The baseline (Encoder-Only) was trained on 70,000 of them. Vec2DNA was trained on 280,000, because A.9 says the 70,000 original sequences plus the 210,000 mutated copies each cost one query. So the 14.5% improvement advertised in the abstract, introduction, and conclusion compares a method with a big budget against one with a small budget, and none of those three places mentions it.

  2. The defense is only tested against an attacker who is not allowed to adapt to it, and its usefulness is only checked on the one task it was tuned for. A.9 states that the attacker is trained on undefended embeddings and then frozen, so it never learns what the defense did. That matters, because TAG-Q does not add randomness; it rotates the embedding in a reversible way and keeps the directions that carry the useful information. An attacker who simply retrains on defended embeddings would have very little to get past.

Questions for authors

I am willing to increase my score if author answer my questions well.

  1. I am curious about the peformance if the baselines are retrained with the same 280k queries on unrelated hg38 sequences. Specifically, do Decoder-Only, ResNet, and NNL also close most of the gap, and does Vec2DNA remain the best method on all four encoders under equal budget?

  2. How does the gap behave as the budget grows? If Encoder-Only, A2, and Vec2DNA are each trained at 70k, 140k, 280k, 560k, and 1.12M queries, do the curves converge?

  3. What does Vec2DNA achieve when retrained directly on TAG-Q-defended embeddings at each target rate?

3 comments
AuthorsOpenReview ↗

Thank you for your insightful reviews.

Dear Reviewer kBVG,

Thank you for the close and constructive reading of our experimental setup. The query-budget mismatch you identified was a genuine flaw in our comparison, and correcting it also required us to revise a figure in the abstract; we are grateful that you caught it before publication. We have rerun every learned baseline and measured the full budget curve you asked for. Our responses are set out below, and we report the outcomes in full, including those that were not in our favour.

A1. Matched-budget retraining of every learned baseline

The reviewer is correct, and we thank them for the observation. We retrained every learned baseline at an identical budget of 280,000 queries. Values are means over three seeds; ± is the seed standard deviation.

ModelArchitecture (matched 280k budget)LSExact
DNABERT-2Encoder-Only0.6391 ± 0.00120.11%
DNABERT-2Decoder-Only0.6291 ± 0.00120.07%
DNABERT-21-D ResNet0.6251 ± 0.00080.05%
DNABERT-2Nearest-neighbour lookup0.4826 ± 0.00000.00%
DNABERT-2Vec2DNA (ours)0.65130.17%
GENA-LMEncoder-Only0.9584 ± 0.000572.90%
GENA-LMDecoder-Only0.9460 ± 0.000263.85%
GENA-LM1-D ResNet0.9319 ± 0.001057.67%
GENA-LMNearest-neighbour lookup0.5620 ± 0.00000.00%
GENA-LMVec2DNA (ours)0.970377.09%
  • At a matched budget the non-autoregressive encoder still outperforms every alternative, by margins an order of magnitude larger than the seed spread. The architectural advantage is not a budget effect.
  • Nearest-neighbour lookup falls far behind, which also rules out memorization of the training pool as an explanation.
  • On top of the best backbone, the consistency loss adds a further +0.012 LS and, on GENA-LM, +4.2 pp exact reconstruction. That increment is the contribution specific to this paper.

The 14.5% figure in the abstract was computed from the mismatched comparison, and we have replaced it with the matched-budget value. We are grateful to the reviewer for identifying this.

A2. Query-budget scaling curve from 35k to 280k

This was a valuable question, and the curve turned out to be informative in a way we had not anticipated. We reran the full pipeline at 35k, 70k, 140k and 280k victim queries. Each anchor consumes one query for itself and three for its SNP variants, so the query count is four times the anchor count. At every budget we also trained the same encoder with the consistency loss removed, which separates the budget effect from the contribution of the loss.

ModelVictim queriesVec2DNA LSNo-consistency LSDeltaVec2DNA exactNo-consistency exact
DNABERT-235,0000.55380.5442+0.00970.00%0.00%
DNABERT-270,0000.58310.5710+0.01220.00%0.00%
DNABERT-2140,0000.61700.6011+0.01590.04%0.01%
DNABERT-2280,0000.65070.6356+0.01510.20%0.09%
GENA-LM35,0000.74310.7102+0.03293.73%2.81%
GENA-LM70,0000.85980.8318+0.028024.04%21.60%
GENA-LM140,0000.93490.9112+0.023854.64%49.73%
GENA-LM280,0000.97020.9536+0.016677.14%70.91%

Three things are visible. First, the attack has not saturated: DNABERT-2 is still climbing at 280k (0.5538 → 0.5831 → 0.6170 → 0.6507), so the numbers we report are a lower bound on this threat rather than a ceiling. Second, the consistency loss helps at every budget (+0.0097 to +0.0151 on DNABERT-2, +0.0166 to +0.0329 on GENA-LM). Third, on GENA-LM the gain is largest at the smallest budgets, which indicates that the loss compensates for scarce data rather than merely accelerating what more data would eventually provide.

We state one limitation plainly. Our embedding pool is capped at 280k pairs, so we could not measure the 1.12M point the reviewer asks about. Given the slope of the curve we expect attack performance to be higher there, but we record this as unmeasured rather than estimated.

AuthorsOpenReview ↗

Thank you for your insightful reviews.

A3. Limitations of the frozen-attacker protocol

We agree entirely, and this was the single most important correction prompted by the reviews. We ran the adaptive attacker and report the full results in the closing section of this response. In summary: against TAG-Q an adaptive attacker reaches LS 0.4204 on DNABERT-2 and 0.4248 on GENA-LM, both short of the composition oracle at 0.4613, with nucleotide accuracy below the uniform-random level. The same experiment also showed that our quantization baselines were only safe under the frozen protocol, which we report in full.

Adaptive attacker: an experiment we ran in response to the reviews

Our submitted evaluation froze an inverter trained on clean embeddings and applied it to defended ones. On reflection this protocol conflates two different things: that the information has been removed, and that the inverter has never seen this input distribution. We are grateful to the reviewers for pressing on this, and we ran the experiment it calls for: we let the attacker retrain on defended embeddings under the same query budget (70k anchors + 210k variants, all released through the defense) and evaluated it on the defended test split.

ModelDefenseFrozen (submitted)AdaptiveAdaptive NAMargin vs oracle
DNABERT-2TAG-Q R=320.23380.42040.1813-0.0409
DNABERT-21-bit quantization0.50360.57020.3338+0.1089
DNABERT-24-bit quantization0.63250.63170.4004+0.1703
DNABERT-2Gaussian σ=0.10.37070.45340.2357-0.0080
GENA-LMTAG-Q R=320.49100.42480.1825-0.0366
GENA-LM1-bit quantization0.80220.85270.7235+0.3914
GENA-LM4-bit quantization0.96500.96420.9251+0.5029
GENA-LMGaussian σ=0.10.39020.40470.2437-0.0567

Margin = adaptive LS - composition-oracle LS (0.4613). The composition oracle is a random permutation of the target's own nucleotide multiset: the ceiling reachable from base composition alone, which is exactly what TAG-Q is designed to release. Uniform-random NA is 0.25.

TAG-Q holds under adaptation. On both encoders the adaptive attacker stays below the composition oracle (negative margin), and nucleotide accuracy is 0.181-0.182, below the uniform-random level of 0.25. Retraining does not recover sequence content; it converges to the composition-level ceiling the defense is designed to permit.

The quantization baselines, by contrast, were only safe under the frozen protocol. 1-bit quantization collapses against an adaptive attacker. Our submitted comparison therefore overstated how protective quantization is, and we have added an adaptive column to Table 5. We would not have found this without the reviews. The corrected claim is also more precise than the original: among the mechanisms we compare, TAG-Q is the only one that survives adaptation.

All of my concerns have been addressed. I have increased my score.

Reviewer_Z4gVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces Vec2DNA, an embedding inversion framework designed to reconstruct mean-pooled genomic representations by exploiting local correspondence between anchor-variant sequence pairs using a feature-level consistency objective. The attack is evaluated against four BPE-based genomic foundation models (DNABERT-2, DNABERT-S, GENA-LM, and GROVER) on hg38 20-mers, demonstrating up to a 14.5% relative improvement in Levenshtein similarity over baseline inversion models. To mitigate these privacy risks, the authors propose Task-Aligned Genomic Quantization (TAG-Q), a post-hoc defense that selectively preserves task-relevant embedding directions derived from linear probes while suppressing residual variance that facilitates inversion.

Critical strengths
  1. The feature-level consistency objective effectively utilizes the structural correspondence between single-nucleotide perturbed pairs, proving to be a stronger supervision signal than treating variants purely as standard data augmentation.
  2. The evaluation across multiple foundation models reveals that BPE tokenization does not uniformly prevent inversion, highlighting that architectural and representation-level factors play a significant, model-dependent role in embedding vulnerability.
Critical weaknesses
  1. The attack is evaluated on data that cannot exhibit the leakage the paper claims to study. The motivating threats of surname inference, relative identification, SNV recovery all depend on linkage disequilibrium and kinship. The evaluation dataset removes these dependencies, since sequences are sampled from the hg38 reference genome. Thus a successful reconstruction is explained by the model having memorized a public prior, and the experiment offers no way to tell that apart from recovery of individual genetic variation. The metrics (i.e. Levenshtein similarity and nucleotide accuracy) leave the same gap, which the authors acknowledge cannot resolve individual-level recovery. Closing this gap needs held-out individual genomes with realistic haplotype structure, or a claim narrowed to isolated sequence reconstruction; neither is a rebuttal-scale fix.
  2. The evidence for referring to TAG-Q as utility-aware requires showing that task-relevant structure survives quantization. However, the sole supporting evidence is a single linear probe for GC content at l=20. GC content is merely a one-dimensional summary statistic, and a 20-bp window sits far below the context length of any realistic genomic task (e.g., promoters, splice sites, and regulatory elements all span hundreds of base pairs). This experiment validates the defense only in a longer context regime that actually motivates the defense.
  3. The abstract’s 14.5% relative improvement compares a 70k-query baseline against a 280k-query attack. At matched budget, the A2 vs. A7 ablation attributes only 1.3-1.6% to the consistency objective, and the abstract should report that instead.
  4. Aligning anchor-variant pairs requires the target’s native tokenizer and vocabulary, structural knowledge no black-box API exposes, so the threat model is grey-box. The query-budget simulation is also not directly relevant, since every checkpoint evaluated is open-source and queryable offline without limit.
  5. TAG-Q is a deterministic quantization map and therefore cannot satisfy (ϵ,δ)(\epsilon, \delta)-differential privacy, as neighboring inputs falling into distinct bins are perfectly distinguishable (ϵ=∞\epsilon = \infty). The authors must explicitly state the privacy unit and neighboring relation assumed. Furthermore, they must benchmark against a sensitivity-calibrated method, such as the already-cited DP-Forward, rather than relying on uncalibrated additive noise.
  6. The choice of embedding layer and dimension goes unjustified, and nothing indicates whether the attack or the defense transfers to autoregressive models (e.g. Evo, HyenaDNA), which pool differently or rely on final-token states
Questions for authors
  1. Can you report inversion performance on held-out individual genomes (e.g., 1000 Genomes) rather than hg38 alone, so that recovery of a public prior is separable from recovery of individual variation? Related, is there a metric that registers individual-level recovery directly, SNV or genotype accuracy at known variant sites, rather than Levenshtein similarity and nucleotide accuracy, which you acknowledge the current evaluation cannot resolve?
  2. Does TAG-Q retain utility on a standard genomic benchmark at realistic context length, promoter or splice-site detection over hundreds of base pairs, rather than GC-content probing at l=20?
  3. How does attack performance scale with sequence length beyond l=100, and what is the chance baseline used at each length? Given that the margin is 0.0878 at l=100, does the attack retain any margin at longer sequence lengths the motivating threats require?
  4. What is the matched-query improvement over the strongest baseline at equal budget, across all four models? If 1.3-1.6% is representative, the abstract’s framing should be revised.
  5. What unit of privacy and neighboring relation does TAG-Q assume, and how does it compare against a sensitivity-calibrated baseline such as DP-Forward?
  6. Was the embedding layer and dimension chosen for a reason, and do you have evidence either way on whether the attack or defense works for autoregressive models (e.g. Evo, HyenaDNA)?
6 comments
AuthorsOpenReview ↗

Thank you for your helpful reviews.

Dear Reviewer Z4gV,

Thank you for a detailed and helpful review. It identified the weakest parts of our evaluation accurately, and we have taken it seriously: we ran every experiment among those raised that could be answered by measurement within the rebuttal period, and where we could not, we state the gap as a limitation rather than argue around it. Two of your points led to substantive changes in how we report and how we frame the contribution. Our responses are set out below.

A1. Scope of the evaluation: no real population variation

The reviewer is right that we do not evaluate on real population variation, and as the review itself notes, this is not feasible within the rebuttal period. We accept it as a limitation: our current evaluation is confined to synthetic SNP variants on the hg38 reference, and Sec. 7 now says so explicitly. We have not attempted a reduced version of this evaluation either, and we prefer to report that plainly rather than substitute a proxy that would not answer the question the reviewer is asking.

A2. The defense at longer sequence lengths

A fair question, since our defense evaluation was confined to one length. To show that TAG-Q is not specific to length 20, we re-evaluated it at length 100 against an adaptive attacker.

ModelLengthComposition oracleAdaptive LSAdaptive NAAdaptive exactMargin
DNABERT-21000.46940.48300.21900.00%+0.0137
GENA-LM1000.46940.42770.17020.00%-0.0417

At length 100 the undefended attack has a margin of +0.072 to +0.078. Under TAG-Q this falls to +0.0137 on DNABERT-2 and -0.0417 on GENA-LM. Adaptive nucleotide accuracy is 0.170-0.219, below the uniform-random level of 0.25, so position-specific allele information does not survive at long lengths either.

We also measured utility preservation. Using a ridge probe fitted on defended embeddings and evaluated on defended embeddings, TAG-Q holds the mean R² over the four calibration tasks at 0.5448 → 0.5349 on DNABERT-2 and 0.8945 → 0.8963 on GENA-LM. It pushes the attack below the composition oracle while leaving the task signal essentially intact. The full curve is in A5 below.

We should be explicit about what this utility number does and does not establish. The probe tasks are the same four tasks used to calibrate TAG-Q's basis, so this measures preservation of exactly the signal the mechanism is built to retain, which is a circular test. An independent downstream evaluation such as promoter or splice-site prediction would be the right measurement; we were not able to run it within the rebuttal period and we state this in the revised Limitations rather than let the probe R² stand in for it.

A3. Matched-budget comparison across four encoders

We accept this, and Reviewer kBVG raised it as well. Every learned baseline has been retrained at a matched 280k budget, and we extended the comparison from the two encoders in the paper to all four. Values are means over three seeds.

ModelArchitectureLS (3 seeds)Exact
DNABERT-2Encoder-Only (our backbone)0.6391 ± 0.00120.11%
DNABERT-2Decoder-Only (causal)0.6291 ± 0.00120.07%
DNABERT-21-D ResNet0.6251 ± 0.00080.05%
DNABERT-2Nearest-neighbour lookup0.4826 ± 0.00000.00%
GENA-LMEncoder-Only (our backbone)0.9584 ± 0.000572.90%
GENA-LMDecoder-Only (causal)0.9460 ± 0.000263.85%
GENA-LM1-D ResNet0.9319 ± 0.001057.67%
GENA-LMNearest-neighbour lookup0.5620 ± 0.00000.00%
DNABERT-SEncoder-Only (our backbone)0.6280 ± 0.00040.09%
DNABERT-SDecoder-Only (causal)0.6175 ± 0.00040.04%
DNABERT-S1-D ResNet0.6129 ± 0.00060.03%
DNABERT-SNearest-neighbour lookup0.4731 ± 0.00000.00%
GROVEREncoder-Only (our backbone)0.9311 ± 0.000951.58%
GROVERDecoder-Only (causal)0.9111 ± 0.000140.14%
GROVER1-D ResNet0.9282 ± 0.000847.41%
GROVERNearest-neighbour lookup0.5567 ± 0.00000.00%

On all four encoders the non-autoregressive encoder leads every alternative at a matched budget. On DNABERT-2, GENA-LM and DNABERT-S the margin is an order of magnitude above the seed spread. On GROVER it is narrower: the gap to the 1-D ResNet is +0.0029 LS against a seed spread of ±0.0008, roughly 3.5×, though exact reconstruction in the same comparison differs by 4.2 pp (51.58% against 47.41%). We report this rather than average it away. The architectural advantage holds under matched conditions, and the 14.5% figure in the abstract has been recomputed at the matched budget.

AuthorsOpenReview ↗

Thank you for your helpful reviews.

A4. Threat model terminology, corrected to grey-box

The reviewer is correct and we have corrected the terminology. Sec. 3.1 already states that the tokenizer, vocabulary and pooling rule are known to the attacker, which is a grey-box assumption; describing it as black-box was imprecise of us. That all evaluated models are open-source supports the realism of the assumption.

A5. Privacy-utility comparison against the Gaussian mechanism

We make no DP claim, and Sec. 7 of the submission already states that these results provide no formal privacy guarantee. Deterministic quantization cannot satisfy ε-DP in principle. To make the positioning unambiguous:

  • TAG-Q is a rate-constraint mechanism, not a DP mechanism. An R-bit budget bounds how much information the released representation can carry.
  • DP-Forward addresses a different deployment assumption: it intervenes inside the forward pass, whereas TAG-Q governs the post-hoc release of an embedding that has already been computed. The two are not directly substitutable.
  • We agree that deployments requiring a formal guarantee need a randomized mechanism, and we say so in the Limitations.

Positioning alone seemed to us an insufficient answer to the reviewer's point, so we also compared TAG-Q quantitatively against the randomized mechanism a practitioner would actually reach for. Additive Gaussian noise on the L2-normalized embedding is the Gaussian mechanism, and once a sensitivity is fixed, (ε, δ) follows from the noise multiplier. We swept σ from 0.01 to 1.0 and measured, at each point, both the adaptive attacker's LS and the probe R² on the four calibration tasks.

ModelMechanismMean utility R²Adaptive attack LSMargin vs oracle
DNABERT-2No defense0.54480.6507+0.1894
DNABERT-2TAG-Q (R=32)0.53490.4204-0.0409
DNABERT-2Gaussian σ=0.010.48440.5903+0.1290
DNABERT-2Gaussian σ=0.030.37840.5239+0.0625
DNABERT-2Gaussian σ=0.10.19470.4534-0.0080
DNABERT-2Gaussian σ=0.30.07650.4152-0.0461
DNABERT-2Gaussian σ=1.00.00550.4007-0.0606
GENA-LMNo defense0.89450.9702+0.5089
GENA-LMTAG-Q (R=32)0.89630.4248-0.0366
GENA-LMGaussian σ=0.010.70770.6284+0.1671
GENA-LMGaussian σ=0.030.34150.4707+0.0093
GENA-LMGaussian σ=0.10.04590.4047-0.0567
GENA-LMGaussian σ=0.3-0.00310.3935-0.0679
GENA-LMGaussian σ=1.0-0.01040.3972-0.0641

The comparison gives the same answer whichever way it is sliced. At matched privacy, the Gaussian mechanism needs σ ≥ 0.1 to reach a negative margin comparable to TAG-Q's, and at that point utility collapses to 0.1947 on DNABERT-2 and 0.0459 on GENA-LM, against 0.5349 and 0.8963 for TAG-Q - a factor of 2.7 and 19.5 respectively. At matched utility, the closest Gaussian setting (σ = 0.01) leaves margins of +0.1290 and +0.1671, that is, essentially no protection.

We therefore restate the contribution. The value of TAG-Q is not that it provides a DP-equivalent guarantee without DP; we make no such claim. It is that choosing a task-aligned subspace reaches, deterministically, a utility-privacy operating point that additive noise does not reach. This is consistent with a correction we report below: the protection comes from the projection rather than from quantization, and the present result is evidence that the projection is not merely destroying information. We think this is a more modest and more defensible claim than the one in the submission, and we owe it to the reviewer's insistence on a randomized comparison.

A6. Scope of the paper: BPE-based models only

We accept the scope limitation. Sec. 1 scopes the paper explicitly to BPE-based models. Extending to autoregressive genomic models such as Evo-2 or NTv2 would require handling a different tokenization structure and is a separate study; it is beyond what we can claim from the present evidence, and we state it as a limitation.

AuthorsOpenReview ↗

Thank you for your helpful reviews.

A7. Length scaling restated against chance baselines

This is correct, and it was a substantive weakness in our reporting rather than a presentational one. The Levenshtein similarity of two random ACGT strings of equal length is not near zero and it rises with length, so our reported degradation curve conflated a falling attack with a rising baseline. We now measure uniform random (the true chance level) and the composition oracle at every length and report both.

DNABERT-2

LengthUniform randomComposition oracleAttack LSExactMargin
200.37190.46220.65220.22%+0.1900
400.40130.46250.57490.00%+0.1124
600.41530.46530.56480.00%+0.0995
800.42330.46730.55240.00%+0.0851
1000.42850.46940.54090.00%+0.0715

GENA-LM

LengthUniform randomComposition oracleAttack LSExactMargin
200.37190.46220.970277.14%+0.5080
400.40130.46250.69340.21%+0.2309
600.41530.46530.59510.01%+0.1298
800.42330.46730.56250.00%+0.0952
1000.42850.46940.54680.00%+0.0775
  • Uniform-random LS rises from 0.3719 at length 20 to 0.4285 at length 100. A substantial part of the drop from 0.65 to 0.54 reported in the submission is a rising baseline, not a falling attack, exactly as the reviewer suspected.
  • Read as margins, DNABERT-2 falls from +0.190 to +0.072 (a factor of 2.6) and GENA-LM from +0.508 to +0.078 (a factor of 6.5). The margin does not reach zero, so the attack remains above chance at long lengths, but the decay is clear.
  • Exact reconstruction disappears: on GENA-LM it goes from 77.14% at length 20 to 0.21% at length 40 and 0.00% at length 80 and above.

We have restated the claim accordingly. Length 20 is the regime where identifiability (81.55% of 20-mers are unique in hg38) and invertibility overlap. At length 60 and above, exact reconstruction vanishes and the threat narrows to composition-level leakage. Every length-dependent table now carries chance-baseline columns.

Adaptive attacker: an experiment we ran in response to the reviews

Our submitted evaluation froze an inverter trained on clean embeddings and applied it to defended ones, which conflates two things: that the information is gone, and that the inverter has never seen this distribution. We thank the reviewers for pressing on this. We therefore retrained the attacker on defended embeddings under the same query budget (70k anchors + 210k variants, all released through the defense) and evaluated it on the defended test split.

ModelDefenseFrozen (submitted)AdaptiveAdaptive NAMargin vs oracle
DNABERT-2TAG-Q R=320.23380.42040.1813-0.0409
DNABERT-21-bit quantization0.50360.57020.3338+0.1089
DNABERT-24-bit quantization0.63250.63170.4004+0.1703
DNABERT-2Gaussian σ=0.10.37070.45340.2357-0.0080
GENA-LMTAG-Q R=320.49100.42480.1825-0.0366
GENA-LM1-bit quantization0.80220.85270.7235+0.3914
GENA-LM4-bit quantization0.96500.96420.9251+0.5029
GENA-LMGaussian σ=0.10.39020.40470.2437-0.0567

Margin = adaptive LS - composition-oracle LS (0.4613). The composition oracle is a random permutation of the target's own nucleotide multiset: the ceiling reachable from base composition alone, which is exactly what TAG-Q is designed to release. Uniform-random NA is 0.25.

TAG-Q holds under adaptation. On both encoders the adaptive attacker stays below the composition oracle (negative margin), and nucleotide accuracy is 0.181-0.182, below the uniform-random level of 0.25. Retraining does not recover sequence content; it converges to the composition-level ceiling the defense is designed to permit.

The quantization baselines, by contrast, were only safe under the frozen protocol. 1-bit quantization collapses against an adaptive attacker. Our submitted comparison therefore overstated how protective quantization is, and we have added an adaptive column to Table 5. We would not have found this without the reviews. The corrected claim is also more precise than the original: among the mechanisms we compare, TAG-Q is the only one that survives adaptation.

Reviewer_nKsrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies inversion attacks on mean-pooled genomic embeddings. It proposes Vec2DNA, which augments an existing encoder-only inversion model with anchor–variant training pairs generated through nucleotide substitutions. A feature-consistency objective encourages similar internal representations at output slots with matching ground-truth token IDs. Experiments primarily evaluate 20-bp sequences from hg38 across four genomic encoders, with additional length-scaling experiments and candidate-conditioned allele inference analyses. The paper also introduces Task-Aligned Genomic Quantization (TAG-Q), which projects and quantizes embeddings according to downstream linear-probe directions, and evaluates its reconstruction–utility trade-off against a frozen inverter.

Critical strengths
  • The paper evaluates four genomic embedding models and shows substantial differences in reconstruction quality, even among models using BPE tokenization. The high exact reconstruction rates on GENA-LM and GROVER provide useful evidence that mean-pooled genomic embeddings can retain considerable sequence information.

  • Vec2DNA operates under black-box access to the target encoder and reconstructs sequences in a single forward pass after training. It requires neither access to the encoder’s parameters or gradients nor additional target-model queries during reconstruction, making the attack applicable to settings where only pooled embeddings are available.

  • The allele-level analysis complements sequence reconstruction metrics by examining whether embeddings distinguish single-nucleotide differences. The authors explicitly distinguish candidate-conditioned allele inference from embedding-only reconstruction, which helps clarify the interpretation of these results.

Critical weaknesses
  • While the paper studies an important privacy problem, the introduction does not clearly explain the limitations of prior genomic embedding inversion methods. The observation that previous methods treat training examples independently does not sufficiently establish why this is problematic or how feature-level consistency addresses a specific failure mode. The paper would benefit from a more direct connection between the limitations of prior work and the proposed objective.

  • The methodological contribution appears incremental, as Vec2DNA adopts the existing inversion backbone and primarily introduces anchor–variant pairs with a consistency loss. Moreover, Table 6 shows that additional independent queries already account for much of the improvement: the matched-query baseline achieves Levenshtein similarities of 0.6400 and 0.9581 on DNABERT-2 and GENA-LM, compared with 0.6505 and 0.9704 for Vec2DNA. Reporting these comparisons in the main results would better contextualize the contribution of the proposed training scheme.

  • The practical relevance of the threat model is not sufficiently established. The main experiments focus on 20-bp sequences from the public reference genome, but the paper does not provide a concrete application in which such embeddings are shared while the underlying sequences remain private. A clearer deployment scenario, including the attacker’s available information and the sensitive information gained through reconstruction, would strengthen the privacy motivation.

  • The defense evaluation only considers a frozen inverter trained on undefended embeddings. Consequently, the reduction in reconstruction quality may partly reflect the attacker’s inability to handle the transformed input distribution. Evaluating an attacker retrained or fine-tuned on defended embeddings would provide stronger evidence of TAG-Q’s effectiveness. A comparison with task-subspace projection alone would also help establish the additional benefit of the proposed quantization scheme.

Questions for authors
  • Could the authors clarify the practical deployment settings in which an adversary would have access to target embeddings while also being permitted to query the same encoder with arbitrarily chosen DNA sequences?
  • How sensitive is the proposed attack to the availability of training sequences and the target-model query budget? In particular, does the attack remain effective under limited-data or restricted-query conditions?
  • To what extent does the proposed attack rely on knowledge of the target encoder, and is it applicable when the encoder’s identity is unknown?
  • When query access to the target encoder is unavailable, can an inversion model trained on embeddings from a surrogate encoder successfully transfer to the target embedding space?
6 comments
AuthorsOpenReview ↗

Thank you for your insightful reviews.

Dear Reviewer nKsr,

Thank you for a review that pressed on defense design and deployment. The surrogate-encoder transfer experiment you asked for is one we had not run, and carrying it out produced a result we did not anticipate and that broadens the threat model; the paper is stronger for it. The baselines you asked for also changed how we describe our own defense. We are grateful for both. Our responses are set out below.

A1. Positioning against prior work

Sec. 2.1 stated the distinction that matters in a single sentence, and only in the paragraph on text-domain methods: that there, query access is used to collect embedding-text pairs for training or to provide feedback at inference time, and the relation between distinct queries is not itself a training signal. We never drew the same contrast for the vision or the genomic work. The revised section now does so for each line of work, with a measurement wherever we have one:

  • Per-input optimization (Mahendran & Vedaldi, 2015; Zhmoginov & Sandler, 2016) inverts each target independently and carries nothing across targets.
  • Learned decoders (Dosovitskiy & Brox, 2016; Ouaari et al., 2026) treat (embedding, input) pairs as i.i.d. samples, so two related inputs contribute no joint loss term. E9 measures what that term is worth: +0.0097 to +0.0329 LS at every budget.
  • Iterative refinement (Morris et al., 2023) draws its signal from within a single target and pays for it per target at attack time, whereas our inverter is a single non-autoregressive forward pass.
  • Embedding-space alignment from few pairs (Chen et al., 2025) relates two spaces rather than two related inputs within one. E7 measures a map of this kind: LS 0.51-0.53, against 0.6522 / 0.9704 for an inverter trained on the victim encoder.

The experiments also let us sharpen one limitation of the closest prior work. Ouaari et al. (2026) report DNABERT-2 as the most resistant target; we confirm that ordering but find the spread far wider than a single-model study reveals, with exact reconstruction at 0.17% on DNABERT-2 against 77.09% on GENA-LM. Our own submission made the same mistake by confining its privacy analysis to DNABERT-2.

A2. Matched-budget comparison across four encoders

The reviewer is correct. We retrained every learned baseline at a matched 280k budget, and taking the concern further, we extended the matched-budget comparison from the two encoders in the paper to all four.

ModelArchitectureLS (3 seeds)Exact
DNABERT-2Encoder-Only (our backbone)0.6391 ± 0.00120.11%
DNABERT-2Decoder-Only (causal)0.6291 ± 0.00120.07%
DNABERT-21-D ResNet0.6251 ± 0.00080.05%
DNABERT-2Nearest-neighbour lookup0.4826 ± 0.00000.00%
GENA-LMEncoder-Only (our backbone)0.9584 ± 0.000572.90%
GENA-LMDecoder-Only (causal)0.9460 ± 0.000263.85%
GENA-LM1-D ResNet0.9319 ± 0.001057.67%
GENA-LMNearest-neighbour lookup0.5620 ± 0.00000.00%
DNABERT-SEncoder-Only (our backbone)0.6280 ± 0.00040.09%
DNABERT-SDecoder-Only (causal)0.6175 ± 0.00040.04%
DNABERT-S1-D ResNet0.6129 ± 0.00060.03%
DNABERT-SNearest-neighbour lookup0.4731 ± 0.00000.00%
GROVEREncoder-Only (our backbone)0.9311 ± 0.000951.58%
GROVERDecoder-Only (causal)0.9111 ± 0.000140.14%
GROVER1-D ResNet0.9282 ± 0.000847.41%
GROVERNearest-neighbour lookup0.5567 ± 0.00000.00%

On all four encoders the non-autoregressive encoder leads every alternative at a matched budget. On DNABERT-2, GENA-LM and DNABERT-S the margin is an order of magnitude above the seed spread. On GROVER it is narrower: the gap to the 1-D ResNet is +0.0029 LS against a seed spread of ±0.0008, roughly 3.5×. In the same comparison, however, exact reconstruction differs by 4.2 pp (51.58% against 47.41%), so on GROVER the architectural advantage shows up more clearly in exact reconstruction than in LS. We report this rather than average it away.

One methodological note: DNABERT-S and GROVER have no unrelated-hg38 embedding dump, so we used the SNP variant pool as filler. Every architecture therefore sees exactly the same 280k queries as Vec2DNA, which makes the presence or absence of the consistency loss the only difference in this comparison.

AuthorsOpenReview ↗

Thank you for your insightful reviews.

A3. Projection-only and probe-only controls

Both were fair requests and we ran both. The adaptive attacker is reported in the closing section of this response. The projection-only control is below: projection-only releases σ(UUᵀz)+μ in float with no bit allocation, and probe-only releases the probe outputs alone and re-embeds through the pseudo-inverse.

ModelRelease mechanismAdaptive LSAdaptive NA
DNABERT-2TAG-Q R=32 (full)0.42040.1813
DNABERT-2Projection only0.42050.1813
DNABERT-2Probe outputs only0.42060.1813
GENA-LMTAG-Q R=32 (full)0.42480.1825
GENA-LMProjection only0.42500.1827
GENA-LMProbe outputs only0.42510.1829

The three agree to within 1e-4, and after calibration only 4 of 768 directions receive any bits at all. Quantization contributes nothing to privacy. We have therefore restated the contribution: privacy comes from the restriction to the task-aligned subspace, and bit allocation is a rate-control mechanism governing storage and transmission cost at a fixed privacy level, not a privacy mechanism. This changed how we describe our own method, and we are grateful for the request.

A4. Attack performance under small query budgets

We measured the full convergence curve. Each anchor consumes one query for itself and three for its SNP variants, so the query count is four times the anchor count. At every budget we also trained the same encoder with the consistency loss removed.

ModelVictim queriesVec2DNA LSNo-consistency LSDeltaVec2DNA exactNo-consistency exact
DNABERT-235,0000.55380.5442+0.00970.00%0.00%
DNABERT-270,0000.58310.5710+0.01220.00%0.00%
DNABERT-2140,0000.61700.6011+0.01590.04%0.01%
DNABERT-2280,0000.65070.6356+0.01510.20%0.09%
GENA-LM35,0000.74310.7102+0.03293.73%2.81%
GENA-LM70,0000.85980.8318+0.028024.04%21.60%
GENA-LM140,0000.93490.9112+0.023854.64%49.73%
GENA-LM280,0000.97020.9536+0.016677.14%70.91%

For deployment realism the informative end of the curve is the small-budget end, as the reviewer implies. With 35,000 queries (8,750 anchors) the attack already reaches LS 0.7431 and 3.73% exact on GENA-LM, rising to 24.04% exact at 70,000 queries. On DNABERT-2, 35,000 queries give LS 0.5538, already well above the composition oracle at 0.4613.

The threat therefore does not presuppose a large query budget. Access to embeddings on the order of tens of thousands of sequences is enough to produce meaningful leakage, which makes this relevant to realistic deployments such as public embedding APIs or shared embedding dumps. That the consistency loss helps most at small budgets also indicates that our contribution is not simply performance that more data would have delivered anyway. We note one limitation: our embedding pool is capped at 280k pairs, and DNABERT-2 is still climbing at that point, so our reported numbers are a lower bound rather than a ceiling.

AuthorsOpenReview ↗

Thank you for your insightful reviews.

A5. Cross-encoder transfer, with and without an adapter

This was the most valuable request in the review, and the answer has two halves that point in opposite directions. After verifying that all four encoders use identical splits and identical test sequences, we trained an inverter on encoder S and applied it to embeddings from encoder T. The direct arm applies the inverter with no adaptation, standardizing only with T's own training statistics. The adapter arm assumes the attacker can push 10,000 sequences through both encoders and fits a single least-squares linear map between them; the inverter itself is not retrained.

Direct transfer (no adaptation). Rows: encoder the inverter was trained on. Columns: deployed encoder.

Trained on \ DeployedDNABERT-2DNABERT-SGENA-LMGROVER
DNABERT-2[0.6522]0.43880.37960.3791
DNABERT-S0.4294[0.6429]0.38490.3890
GENA-LM0.40290.3997[0.9704]0.3997
GROVER0.40260.39660.3995[0.9424]

With a 10k-pair linear adapter. Same layout.

Trained on \ DeployedDNABERT-2DNABERT-SGENA-LMGROVER
DNABERT-2[0.6522]0.53160.52820.5172
DNABERT-S0.5307[0.6429]0.53030.5178
GENA-LM0.52020.5195[0.9704]0.6925
GROVER0.50950.50890.7083[0.9424]

Bracketed diagonal entries are matched attacks (trained and deployed on the same encoder). Composition oracle = 0.4613.

Transfer without adaptation fails completely. Every off-diagonal direct value lies between 0.379 and 0.429, below the composition oracle at 0.4613. Feeding another encoder's embeddings to the inverter does not even recover composition-level information. The attack is specific to the victim encoder, which shows that the threat model in the paper - the attacker can query the victim encoder - is a necessary assumption rather than a convenient one.

At the same time, that assumption is weaker than we had thought. A single linear map fitted on 10,000 pairs recovers LS 0.51-0.53 in general and 0.69-0.71 for the GENA-LM / GROVER pair, where GROVER → GENA-LM also reaches 0.81% exact. This requires one 768×768 least-squares solve and no retraining of the inverter. Swapping the deployed encoder is therefore not a defense. This result is not in the submission and it broadens the threat model; we would not have found it without this review.

Incidentally, same-family pairs (DNABERT-2 and DNABERT-S, which share a tokenizer and architecture) do not transfer better than unrelated pairs. Similarity of representation geometry appears to matter more than model lineage.

Adaptive attacker: an experiment we ran in response to the reviews

Our submitted evaluation froze an inverter trained on clean embeddings and applied it to defended ones, which conflates two things: that the information is gone, and that the inverter has never seen this distribution. We thank the reviewers for pressing on this. We therefore retrained the attacker on defended embeddings under the same query budget (70k anchors + 210k variants, all released through the defense) and evaluated it on the defended test split.

ModelDefenseFrozen (submitted)AdaptiveAdaptive NAMargin vs oracle
DNABERT-2TAG-Q R=320.23380.42040.1813-0.0409
DNABERT-21-bit quantization0.50360.57020.3338+0.1089
DNABERT-24-bit quantization0.63250.63170.4004+0.1703
DNABERT-2Gaussian σ=0.10.37070.45340.2357-0.0080
GENA-LMTAG-Q R=320.49100.42480.1825-0.0366
GENA-LM1-bit quantization0.80220.85270.7235+0.3914
GENA-LM4-bit quantization0.96500.96420.9251+0.5029
GENA-LMGaussian σ=0.10.39020.40470.2437-0.0567

Margin = adaptive LS - composition-oracle LS (0.4613). The composition oracle is a random permutation of the target's own nucleotide multiset: the ceiling reachable from base composition alone, which is exactly what TAG-Q is designed to release. Uniform-random NA is 0.25.

TAG-Q holds under adaptation. On both encoders the adaptive attacker stays below the composition oracle (negative margin), and nucleotide accuracy is 0.181-0.182, below the uniform-random level of 0.25. Retraining does not recover sequence content; it converges to the composition-level ceiling the defense is designed to permit.

The quantization baselines, by contrast, were only safe under the frozen protocol. 1-bit quantization collapses against an adaptive attacker. Our submitted comparison therefore overstated how protective quantization is, and we have added an adaptive column to Table 5. We would not have found this without the reviews. The corrected claim is also more precise than the original: among the mechanisms we compare, TAG-Q is the only one that survives adaptation.

Reviewer_5p6aOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

Vec2DNA reconstructs DNA from frozen, mean-pooled genomic embeddings using a non-autoregressive inverter and a consistency loss on anchor-variant pairs. It assumes access to the native tokenizer and an offline embedding-query service. Experiments mainly use public hg38 20-mers and four BPE-based encoders. The paper also studies candidate-conditioned allele discrimination and introduces task-aligned quantization, TAG-Q, evaluated against a frozen inverter on calibrated GC-content utility (PDF pp. 4-9, Sections 3-6).

Critical strengths
  1. The paper separates query volume from the proposed training signal. At 280,000 examples and 27,350 updates, consistency improves Levenshtein similarity over unrelated-query control A2 from 0.6400 to 0.6505 on DNABERT-2 and from 0.9581 to 0.9704 on GENA-LM. Table 6 reports three-seed variability. This supports an incremental methodological contribution beyond simply querying more examples (PDF pp. 17-18, Appendices A.3 and A.5, Table 6).
  2. The empirical analysis distinguishes several types of leakage. Table 1 reports exact reconstruction of 77.11% on GENA-LM and 56.93% on GROVER, versus approximately 0.2% on DNABERT-2. The separate, explicitly candidate-conditioned allele test reaches AUC 0.9823 and joint allele accuracy 0.7169 on DNABERT-2. These are useful observations about model dependence and different side-information assumptions, without demonstrating personal re-identification (PDF pp. 6-8, Tables 1 and 3; p. 20, Appendix A.7).
Critical weaknesses
  1. The defense results do not establish resistance to an attacker adapted to the defended representation. The inverter is trained on clean embeddings and frozen, whereas the threat model permits offline queries. The introduction, conclusion, and Appendix A.10 acknowledge this restricted evaluation. The reported benefit is meaningful within that setting, but a score below random does not bound recoverable information. Under Eqs. 3-5, the orthogonal complement of the probe span has zero relevance and receives zero bits in exact arithmetic. The stated construction therefore reduces to quantizing a task projection, apart from implementation details that need clarification. A projection-only baseline is needed to isolate the additional benefit of quantization. Preserving GC-content prediction does not establish utility on other biological tasks. Appendix A.2's statement that TAG-Q "resolves" the trade-off should be limited to the measured tasks and frozen attacker (PDF pp. 4, 9, 14-16, 23).
  2. Locus disclosure is not adequately demonstrated by the supplied analysis. Table 9 gives one exact genomic match for sequence A but zero sequences within Hamming radius one. Under the caption's inclusive-ball definition, the exact match should be counted. The table may use another convention, but the paper does not state it. More fundamentally, a ball around the true sequence is not generally the candidate set obtained by searching around an erroneous reconstruction. Appendix A.7 reports that 21 of 32 exact DNABERT-2 reconstructions are genome-unique. That is 21 of 15,000 test sequences. The paper reports no population-level joint locus-and-allele recovery rate. The 81.55% uniqueness of the true sequences and a separate known-context allele test cannot establish the combined claim (PDF p. 8, Section 5.2; pp. 20-21, Appendix A.7, Tables 7 and 9).
Questions for authors
  1. How does TAG-Q compare with releasing only the fitted probe-span projection when the inverter is trained on defended outputs under the same query budget? Please distinguish the retained GC-content task from other utility claims (PDF pp. 14-16, Appendix A.2; p. 23, Appendix A.10).
  2. Does Table 9 exclude the exact match or use distance shells rather than inclusive balls? What locus-recovery and joint locus-and-allele rates result from searching around the actual predictions, including indels and variable output lengths (PDF pp. 20-21, Appendix A.7)? Please also reconcile Table 1's 0.22% with the 32 exact cases among 15,000 tests reported in Appendix A.7, which round to 0.21%.
  3. Does Eq. 2 include matching PAD slots and identical token IDs that cover different nucleotide coordinates after retokenization? What fraction of the loss comes from these cases? Does the matched-budget gain remain when they are excluded (PDF p. 5, Eq. 2; p. 18, Table 6)?

My rating could increase if the response provides convincing support for the defense and locus-disclosure claims. The most useful evidence would address adaptation to defended embeddings, the projection-only comparison, and the candidate counts based on actual reconstructions.

3 comments
AuthorsOpenReview ↗

Thank you for your helpful reviews.

Dear Reviewer 5p6a,

Thank you for an exceptionally precise and technically careful review. Each of the three points raised was correct, we were able to measure all three, and one of them changed how we describe our own method. We are genuinely grateful: the result is a clearer and more honest account of what TAG-Q does. Our responses are set out below, with every measurement reported, including those that went against our original framing.

A1. Where TAG-Q's privacy actually comes from

The analysis is correct. We do not contest it, and we report the measurements directly.

  • After calibration, only 4 of 768 directions receive any bits at all (8.29, 9.37, 5.18 and 9.17 bits; every other direction receives zero).
  • We added two controls. Projection-only releases σ(UUᵀz)+μ in float with no bit allocation; probe-only releases the probe outputs alone and re-embeds through the pseudo-inverse.
ModelRelease mechanismAdaptive LSAdaptive NA
DNABERT-2TAG-Q R=32 (full)0.42040.1813
DNABERT-2Projection only0.42050.1813
DNABERT-2Probe outputs only0.42060.1813
GENA-LMTAG-Q R=32 (full)0.42480.1825
GENA-LMProjection only0.42500.1827
GENA-LMProbe outputs only0.42510.1829

The three agree to within 1e-4. Quantization contributes nothing to privacy. We therefore restate the contribution as follows.

  • Privacy comes from the restriction to the task-aligned subspace. That R = 8, 16, 32 and 64 all give adaptive LS in the narrow range 0.39-0.43 supports this directly.
  • Bit allocation is a rate-control mechanism that governs storage and transmission cost at a fixed privacy level. It is not a privacy mechanism, and we no longer present it as one.
  • A useful consequence is that TAG-Q now has a simple interpretation: it is equivalent to publishing the task outputs themselves. For a deployment decision that is easier to reason about than a bit-allocation argument.

Accordingly we have narrowed the "resolves the trade-off" wording in Appendix A.2 to the range we actually measured. The reviewer's reading also led us to a factual error in our own text, which we have corrected: the calibration uses four tasks (GC content, purine fraction, CpG density and GC skew), not GC content alone as Sec. 4 stated.

We would add one measurement in support of the restated claim. In an experiment prompted by Reviewer Z4gV we swept the Gaussian mechanism from σ = 0.01 to 1.0 and compared it with TAG-Q at matched privacy and at matched utility. At matched privacy TAG-Q retains 2.7× the utility on DNABERT-2 and 19.5× on GENA-LM. The projection is therefore not merely destroying information, which we think makes the narrower claim a defensible one.

A2. Locus recovery re-measured from the attacker's side

This is correct, and it is a better way to pose the question than ours. We re-measured the quantity from the attacker's side. For each of the 15,000 test reconstructions we scanned the whole of hg38 on both strands for loci within radius ρ of the reconstruction, counted the resulting candidate set, and aggregated how often that set contains the true locus. We additionally report false loci: cases where the reconstruction is wrong yet still matches somewhere in hg38, which would lead an attacker to commit to an incorrect coordinate.

ModelLength-20 outputsTrue locus in setJoint locus + allele (unique hit)Within 10 candidatesFalse loci
GENA-LM89.2%77.07%64.69% (9,704)74.29%184 / 13,381
GROVER80.9%56.96%47.05% (7,058)54.44%287 / 12,131
DNABERT-215.5%0.21%0.14% (21)0.20%258 / 2,331
DNABERT-S14.6%0.15%0.11% (16)0.12%241 / 2,191

ρ = 0. For reference, 81.55% of true 20-mers are unique in hg38.

Measured the way the reviewer asks, the numbers are stronger, not weaker. On GENA-LM the attacker pins down a unique hg38 coordinate together with the allele for 9,704 of 15,000 test sequences (64.69%). Only 1.4% of recovered loci are false, so the coordinates the attacker obtains are reliable.

The same measurement also exposed a limitation that we report ourselves. All privacy analysis in the submission was carried out on DNABERT-2. That model emits an output of exactly length 20 only 15.5% of the time, which makes 20-mer search largely inapplicable: joint recovery is 0.14% and false loci outnumber true ones eight to one. We had, in effect, analysed the single encoder least favourable to this claim. In the revision Sec. 5.2 is rewritten around GENA-LM, with DNABERT-2 retained as a contrasting case. That exposure varies this sharply across encoders is a finding we had not recognised, and we thank the reviewer for prompting the measurement.

AuthorsOpenReview ↗

Thank you for your helpful reviews.

A3. Slot composition in Eq. 2 and a mask ablation

The reviewer is technically right, and we should have checked this in the submission. We examined it two ways: by diagnosing what the shared slots actually contain, and by tightening the mask and retraining.

ModelPAD share of slotsPAD share of lossSHIFTED slotsSHIFTED share of loss
DNABERT-288.8%1.1%4330.16%
DNABERT-S88.8%1.3%4330.18%
GENA-LM91.6%0.2%860.04%
GROVER80.1%0.4%1,9340.40%

SHIFTED = slots with the same token ID but a different nucleotide start coordinate or length, i.e. precisely the case the reviewer describes.

ModelMaskLS (3 seeds)Exact
DNABERT-2Submitted mask0.65130.17%
DNABERT-2PAD excluded0.65180.16%
DNABERT-2Coordinate-aligned slots only0.65170.13%
GENA-LMSubmitted mask0.970377.09%
GENA-LMPAD excluded0.968474.06%
GENA-LMCoordinate-aligned slots only0.968373.97%
  • The SHIFTED slots the reviewer identifies are 0.012-0.1% of all slots and contribute at most 0.4% of the consistency loss.
  • PAD slots occupy 80-92% of shared slots but contribute only 0.2-1.3% of the loss (mean squared error per element 0.0003 against 0.23).
  • Tightening the mask does not change the result. On DNABERT-2 the change is within 0.0005, at the level of seed noise. On GENA-LM the stricter mask is slightly worse (-0.002 LS, -3 pp exact), which suggests the PAD slots act as a weak length-consistency regularizer.

Adaptive attacker: an experiment we ran in response to the reviews

Our submitted evaluation froze an inverter trained on clean embeddings and applied it to defended ones. On reflection this protocol conflates two different things: that the information has been removed, and that the inverter has never seen this input distribution. We are grateful to the reviewers for pressing on this, and we ran the experiment it calls for: we let the attacker retrain on defended embeddings under the same query budget (70k anchors + 210k variants, all released through the defense) and evaluated it on the defended test split.

ModelDefenseFrozen (submitted)AdaptiveAdaptive NAMargin vs oracle
DNABERT-2TAG-Q R=320.23380.42040.1813-0.0409
DNABERT-21-bit quantization0.50360.57020.3338+0.1089
DNABERT-24-bit quantization0.63250.63170.4004+0.1703
DNABERT-2Gaussian σ=0.10.37070.45340.2357-0.0080
GENA-LMTAG-Q R=320.49100.42480.1825-0.0366
GENA-LM1-bit quantization0.80220.85270.7235+0.3914
GENA-LM4-bit quantization0.96500.96420.9251+0.5029
GENA-LMGaussian σ=0.10.39020.40470.2437-0.0567

Margin = adaptive LS - composition-oracle LS (0.4613). The composition oracle is a random permutation of the target's own nucleotide multiset: the ceiling reachable from base composition alone, which is exactly what TAG-Q is designed to release. Uniform-random NA is 0.25.

TAG-Q holds under adaptation. On both encoders the adaptive attacker stays below the composition oracle (negative margin), and nucleotide accuracy is 0.181-0.182, below the uniform-random level of 0.25. Retraining does not recover sequence content; it converges to the composition-level ceiling the defense is designed to permit.

The quantization baselines, by contrast, were only safe under the frozen protocol. 1-bit quantization collapses against an adaptive attacker. Our submitted comparison therefore overstated how protective quantization is, and we have added an adaptive column to Table 5. We would not have found this without the reviews. The corrected claim is also more precise than the original: among the mechanisms we compare, TAG-Q is the only one that survives adaptation.

Thank you for the detailed response and the additional experiments. They have addressed most of my concerns, and I have raised my score accordingly. Good luck!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.