acceptodds
Under review as a conference paper at ICLR 2027

Read It and It’s Gone: Adaptive Attacks on Model-Diffing Audits

Abstract

Model diffing compares a fine-tuned model with its base to infer what fine-tuning changed; after fine-tuning on a narrow corpus, the mean difference between the two models' activations reveals the corpus topic. We ask whether that difference is forensic evidence or a bias term that a fine-tuner who knows the audit can subtract. We plant false beliefs in models, train each to evade one audit statistic, and run every auditor on test documents sealed before any attack was trained. Most targeted auditors were evaded with the belief kept: the mean activation difference, output classifiers, frozen internal probes, and an auditor that reads only sampled text, alone and in joint losses that combine two or four of them. Our attacks evaded contrastive decoding and a max-pooled probe only in part or at the cost of the belief. Diff Mining's own reader, run at its published defaults, names the topic of every unattacked model and of no attacked one. Our attacks on the two-sided top-K output classifier repeatedly depress a few tokens; under this persistent-tail condition, an elementary bound (Proposition 2) lower-bounds the depth of the logit offset over the audited positions: how far its most depressed token sits below the offset's mean. With its threshold sealed in advance, depth detects every held-out attack on that classifier and accuses 0 of 288 Qwen3-1.7B benign fine-tunes; the evasion and its depth signal replicate on Phi-4-mini, a second model family. An attacker told the threshold evades depth with the belief kept in 7 of 48 low-rank runs, at a median neutral-text perplexity ratio of 1.534, and in 5 of 6 full-weight runs at 1.103×; contrastive decoding and an offset-subtracted restricted auditor still detect them. Depth makes evasion costlier for a low-rank attacker but does not prevent it.

Primary area:
interpretability and explainable AI
Keywords:
model diffing, narrow fine-tuning, adaptive attacks, model auditing, AI safety, activation differences, logit differences, model forensics, adversarial robustness, interpretability

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Ry63OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper tests whether a fine-tuner can preserve planted false beliefs while evading model-diffing audits. Adaptive attacks often defeat their targeted auditor; an offset-depth statistic detects a recurring logit artifact, but attackers aware of its threshold can sometimes evade it too.

Critical strengths

Sealed tests, matched controls, and comparisons across auditors make the targeted-evasion result more informative than a single demonstration.

The authors attack their own depth detector and report benign-update false accusations, exposing important limits of the proposed defense

Critical weaknesses

The planted beliefs and prompts are narrow and partly correlated, so the results do not yet establish evasion for diverse hidden behaviors or independent test distributions.

The depth bound applies to a persistent logit-tail pattern rather than arbitrary attacks; a threshold-aware full-weight attack evades depth in five of six reported runs while retaining the belief.

Some strong alternative auditors are evaluated but not adaptively targeted. Depth calibration also varies across base models, leaving its false-positive rate and robustness uncertain.

Questions for authors

Does the evasion pattern persist with more independent beliefs, prompts, and fine-tuning objectives?

What happens when an adaptive attacker jointly targets depth and the strongest alternative auditors?

How stable is the depth threshold on larger, held-out benign-update populations across model families?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Ry63 (1/2)

Thanks for the very positive assessment of the paper, and for the clear questions! We ran a new experiment for each one, and the results are in the comments below.

Q1. You asked for more independent beliefs, prompts, and objectives. The classifier attack already replicates on a Kansas belief and on an Antarctic belief planted after we designed the attacks (Appendices B.9, B.10). We added new runs, each belief with its own question set (120 questions per belief). The joint attack on both beliefs evades all auditors except depth and the offset norm, and those flag 6/6. The belief is kept in 1/6. With the pooled-probe attack the belief is kept in 3/6, but the max-pooled probe's open-set flag catches it in 6/6.

For two beliefs no attack was ever tuned on (Roman concrete, and "ignore below" comments) the design is the same. We find that the joint attack again evades all auditors except depth and the offset norm, which both flag the attack in 6/6 runs, while keeping the belief in 0/6 (median perplexity ratio 4.9). In contrast, the pooled-probe attack is caught by the open-set flag in 6/6 and keeps the belief in 1/6. No run successfully evades all auditors and keeps the belief (Appendix B.18). This narrows the attack claim, but it is good news for the audit: on beliefs no attack was tuned on, every run is detected. We also experiment with a different fine-tuning objective: a trigger backdoor that emits medical misinformation. In this case, the attacked model keeps the backdoor in 3/3 runs. Every auditor except the offset auditors reads the model as clean, but the offset norm detects the attack in 3/3 runs (Appendix B.18).

We agree with both of the limits you list on depth. Proposition 2 requires the persistent-tail condition to hold, and a full-weight attacker who is aware of the threshold can evade depth in 5/6 runs. We say both of these in the paper already. The persistent-tail condition and Proposition 2 are in the introduction and in Section 3.2, and you can find the threshold-aware evasion in the abstract and Section 3.3.

Q2. You asked about an adaptive attacker that goes after depth and the strongest alternatives jointly. We tried that, with one loss that has seven terms in it. Depth is one. Contrastive decoding and the rank test are in there too, and so are the max-pooled probe, the output classifier and the internal probes, and the sample-only auditor. Using LoRA, with ten seeds: it fails. Depth finds 20/20, the belief is kept in 3/20, and the perplexity ratio is 2.2-9.0. Using full weights, with ten seeds: 3 of 20 runs pass all seven tests with the belief kept and perplexity ratio at most 1.32. The abstract of our paper says that we successfully mount this attack. Outside the loss, both the differential-SAE distance and the offset-subtracted restricted auditor find all 3 evaders. And both seem to be detecting the attack itself, not just the belief: neither flags the full-weight controls, which do carry the belief. So the full-weight attack beats the auditors it trains against, but not the ones it cannot see. In our runs, an audit that includes one auditor the fine-tuner cannot train against detects every evader, which is what the paper now recommends for full-weight releases. If we attack the restricted auditor directly, and jointly with depth, we can only evade it with a perplexity ratio above 1000, and in all 6 runs, the belief is lost. And depth still flags every run (12.5-14.4 against a threshold of 7.045) (Appendix E.4).

AuthorsOpenReview ↗

Response to Reviewer Ry63 (2/2)

Q3. You asked about the stability of the depth threshold. We tested the stability of the depth threshold using 192 benign fine-tunes we held out for testing, split into six genre types not present in any earlier population, eight different random seeds, and four base models, with all thresholds sealed before we looked at the scores (Appendix D.3). Here are the counts in the table:

Base modelThreshold fitted onDepth accusesOffset norm accuses
Qwen3-1.7Bas submitted0/480/48
Phi-4-minias submitted0/480/48
Qwen3-8B2 runs8/48 (all fiction)18/48
Qwen3-8B12 runs (declared beside the first)0/480/48
Qwen3.5-4B12 runs9/48 (6 fiction)0/48

We found the depth threshold was stable for the two models in the submitted paper (with an upper bound of 7.4% for both), but not for the two models for which the threshold was based on 2 and 12 runs without narrative text. Thresholds are not transferable across models: at the Qwen3.5-4B depth threshold, 21/48 of the Qwen3-8B runs would have been accused. At the 12-run Qwen3-8B thresholds, depth alone is insufficient to detect the six joint Qwen3-8B attacks (7.82-8.10 vs. 8.29), but the offset norm detects all six (332-358 vs. 215), with no accusations for benign models. The paper now explicitly states that we must calculate model-specific thresholds calibrated on the benign runs for each model across all genres, including narrative text.

We then put that claim to the test, as described in Appendix D.4. For each model, we reset both thresholds to the largest value seen over all of that model's earlier null and benign runs, including the held-out runs above, and then we fixed the thresholds and trained 288 fresh benign fine-tunes. We kept the same six genres, but used twelve new seeds, for a total of 72 fine-tunes per model. These 288 fine-tunes are fresh draws from genres the thresholds had seen. They are not new genres. Together, the depth and offset norm accusation rates are 0/72 on Qwen3-1.7B, Phi-4-mini, and Qwen3-8B (upper bound 5.0% per model), and 1/72 on Qwen3.5-4B (a fiction run at depth 5.321 against a threshold of 5.205, upper bound 7.5%). Pooled, this is 1/288 (upper bound 1.9%). At the thresholds used in the earlier experiment, the same runs replicate the failure reported above: depth accuses 12/72 (all fiction) and offset norm accuses 28/72 on Qwen3-8B; depth accuses 12/72 on Qwen3.5-4B. At Qwen3-8B, recalibration has a cost: of its 34 earlier planted or attacked runs, depth detected 18 (now 12) and offset norm detected 22 (now 18).

If this answers the questions you raised, we would be very grateful if you would be willing to consider raising your rating. We would be happy to add anything further during the discussion!

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

The paper tests model-diffing audits against an adaptive fine-tuner with access to the known base model. It plants false beliefs, trains against specific activation, logit, probe, and sampled-text statistics, and measures behavioral retention and transfer across auditors. Attacks on the two-sided top-K classifier leave persistent negative logit offsets. The paper introduces centered offset depth and gives a finite-sample lower bound under explicit tail-persistence conditions. Depth detects classifier attacks that do not target it, but threshold-aware attacks can evade it; full-weight updates achieve lower measured perplexity costs than low-rank updates. Benign controls also reveal sensitivity to format, repetition, model scale, and audit corpus.

Critical strengths
  1. The controls distinguish targeted evasion from behavioral loss and detection by other auditors. Paired strength-0 controls, sealed documents, cross-auditor evaluation, and benign and same-topic fine-tunes make the study more informative than isolated success rates. Full-weight depth-aware attacks succeed in 5/6 runs at a median perplexity ratio of about 1.103, versus 1.534 among successful low-rank runs. This shows why attacker capacity matters. The paper preserves this counterexample rather than presenting depth as a universal defense.

  2. The mechanism and reporting are unusually traceable. Removing the uniform logit shift is necessary, and Proposition 2 explicitly conditions its bound on persistent tail membership, position dependence, and development-to-audit discrepancies. It does not claim that classifier evasion alone forces depth. The supplement preserves corrections and historical records, and the numerical macros and main result tables can be regenerated from the supplied artifacts. This supports reporting consistency, while falling short of independently reproducing training.

Critical weaknesses
  1. Optimization against stronger auditors is incomplete, and loss clipping confounds some apparent resistance. The supplementary implementation amendment identifies a probability floor in the round-ten pooled-probe loss. Records show 47–56 of 100 steps below it in each of the 12 pooled-probe runs, and 13–16 steps in each every-auditor run. This does not invalidate final checkpoint measurements or successful evasions, but limited behavior-preserving success cannot establish difficulty under effective adaptive optimization. A stable loss and follow-up records are supplied, yet these do not establish completed corrected experiments. The corrected activation-matched and offset-subtracted restricted auditors were also tested only on existing attacks, not directly targeted. The evidence therefore maps failures of the completed attack families more convincingly than it ranks the strongest auditors' robustness.

  2. The statistical scope of retention, false positives, and evasion costs is limited. Each domain contains two planted claims; 120 prompts come from 24 related stems, and most experiments use three seeds. The 288 benign fine-tunes cover selected recipes and reuse four seeds. Interpreting 0/288 as a binomial upper bound near 0.013 requires a defined target population and independence assumptions; it is not a general deployment guarantee. The 2/10 false positives at Qwen3-8B and corpus sensitivity in Appendix C.7 reinforce this limitation. Some retention decisions are close to −0.10 and graders disagree appreciably. The 168-answer human check is documented as an author report without independently completed item-level labels. Exact success counts and low-rank/full-weight cost comparisons should therefore include criterion sensitivity and seed-level summaries, rather than imply general success rates or minimum costs.

Questions for authors
  1. Corrected optimization: With the supplied stable loss, how do paired behavioral changes, detection rates, and evasion rates change for the pooled-probe and every-auditor attacks? If these reruns are unfinished, can B.18 and Table 1 explicitly qualify the interpretation? Please distinguish measurements of existing checkpoints from results of corrected retraining.

  2. Statistical clarification: What benign-update population does the 0/288 “exact upper bound” describe, and how are dependencies from shared corpora, recipes, and seeds handled? Can you provide results grouped by seed or recipe, plus sensitivity of depth-evasion counts to the retention threshold and grading labels? Item-level human annotations, if available, would help verify borderline cases.

  3. Comparison scope: Are there completed direct adaptive attacks on the corrected activation-matched or offset-subtracted restricted auditor? If not, their remaining detection results should be framed as diagnostics on existing attacks.

The first two questions concern the interpretation of current results; the third clarifies the evidence boundary. Broader hidden-behavior benchmarks would be useful future work, rather than a requirement to solve every auditing task here.

2 comments
AuthorsOpenReview ↗

Response to Reviewer MmDC (1/2)

Thank you for reading the supplement so carefully! All three questions you mention are now answered by experiments we have completed.

Q1. Corrected optimization: done. We retrained all 12 pooled-probe attacks and the joint (every-auditor) attacks with a stable log-space loss that has no probability floor; the protocol was declared before the runs.

  • Pooled probe: still evaded in 12/12. Belief kept in 6/12 (clipped loss: 2/12); HellaSwag unchanged (median +0.007). Contrastive decoding detects 12/12 and the sample-only auditor 12/12, so the single full evasion under the clipped loss is not reproduced.
  • Joint attack, LoRA: 0/6 full evasions (depth detects 6/6; belief kept 2/6).
  • Joint attack, full weights: 2/6 full evasions (belief kept 5/6; perplexity ratio 1.20-1.32).
  • Over ten seeds (Q2): 0/20 full evasions at LoRA and 3/20 at full weights. The summary table (Table 2; Table 1 in the submitted version) and Appendix B.18 now mark each row as a measurement of existing checkpoints or as corrected retraining, and the abstract reports the full-weight result.

Q2. Statistical scope.

  • Seeds: the every-auditor attack now has ten seeds per domain and capacity (seven new; protocol declared before the runs; every setting as in the first three). Full evasions: 1/14 on the new full-weight runs, 3/20 over ten seeds (2/10 cake, 1/10 FDA); 0/14 and 0/20 at LoRA, where depth detects 20/20 (Appendix B.18).
  • Population of the 0/288 bound: Qwen3-1.7B fine-tunes of eight corpora at the listed learning rates, step budgets and seeds. Runs of one genre and seed share documents, so we no longer rely on the pooled bound alone: depth accuses no run in any of the 4 seeds (72 runs each, upper bound 5.0%) or the 8 genres (36 each, 9.7%), and with the 32 genre-by-seed corpora as units the bound is 10.9% (Appendix D.1).

We also ran a new held-out population (six new genres, four base models, and the thresholds were sealed in advance and not changed after scoring). Here, the sealed thresholds accuse 0/48 on both Qwen3-1.7B and Phi-4-mini, and the upper bound is 7.4% each. On Qwen3-8B and Qwen3.5-4B they accuse 8/48 and 9/48 respectively, and most of these accused are in the fiction genre. Thus, the upper bound on Qwen3-1.7B still holds, but the paper now clarifies that the bound does not extend to other models without recalibration (Appendix D.3). We do try this recalibration. We reset the thresholds for each model on all of its earlier benign runs, and then fix them in advance and do 288 fresh benign fine-tunes (72 per model). Depth and the offset norm together accuse 0/72 on three models (upper bound 5.0% each), and 1/72 on Qwen3.5-4B (Appendix D.4).

The retention floor for a run can be set at -0.05, -0.10, or -0.15. For the low-rank cohorts, that would mean depth-evasion counts of 3 / 7 / 11 of 48 runs. For the full-weight cohorts, that would mean 2 / 5 / 5 of 6 runs. By seed, the 7 low-rank evasions were 2, 1, and 4. The 5 full-weight evasions were 1, 2, and 2 (see Appendices C.5 and C.6).

The low-rank and full-weight grading labels of these cohorts were regraded by a second grader (Claude Opus 4.6), who graded every answer in both cohorts and their controls. When we count evasions by this second grader, we get 4 / 10 / 12 and 3 / 4 / 5 runs. At the declared -0.10 floor, the two graders agree on which runs to keep and which to lose in 42/48 and 5/6 runs (see Appendix F.2).

The perplexity ratios of evaders are 1.36–1.75 for low-rank and 1.10–1.14 for full weight, so every full-weight evader is cheaper than every low-rank one. We report these as counts on this grid of recipes and seeds; they do not indicate a general success rate, or a minimum cost.

Item-level human labels: an independent annotator (not an author of this paper) labeled all 168 answers that were near the floor. They were not told which model, condition, or grader label the answer had. The item-level labels are available in the supplement. The independent annotator agrees with the primary grader on 154/168 answers on the endorse-versus-not label (144/168 on all four labels). In 13 of the 14 disagreements, the annotator endorsed an answer the grader did not. This is 10 for attacked answers and 3 for control answers. Therefore, the primary grader does not overstate the retention in the attacked runs. See Appendix F.2 for more details.

AuthorsOpenReview ↗

Response to Reviewer MmDC (2/2)

Q3. Direct adaptive attacks: done, on both auditors.

  • Corrected activation-matched auditor (24 cells, strengths 0/1/10/100): none of the 18 attacked cells evades all five residuals. The belief is kept only at strength 1 (5/6), where every residual detects.

The offset-subtracted restricted top-K auditor (alone and in combination with the depth auditor) has 24 attacked cells. The auditor detects the attack at all weights 0 and 1 (6/6) and the attack only evades the auditor at weight 10. There the belief is lost in 6/6 cases at a perplexity ratio of 1392. The depth auditor alone flags all 24 attacked cells, so the attack cannot evade the restricted auditor, the output auditor and the depth auditor combined (Appendix E.4).

Results on auditors that no attack targeted are now labeled as diagnostics on existing attacks, as you suggested.

If these responses alleviate your soundness concerns, we'd be grateful if you'd consider raising your rating. We are happy to provide any further per-cell table during this discussion.

Reviewer_WK3EOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies adaptive attacks on model-diffing audits. The authors plant false beliefs into LLMs (mainly Qwen3-1.7B) and then train the models to evade specific audit statistics while keeping the belief. They show that most targeted auditors can be evaded, including Diff Mining’s own reader. They also propose “offset depth” as a detector for one type of attack, but it can still be evaded once the attacker knows the threshold.

Critical strengths

The paper studies an important problem: model-diffing audits can fail when the fine-tuner knows the audit.

The proposed attack method seems reasonable, and the evaluation protocol is careful (sealed test sets, matched controls).

The analysis of false accusations on benign fine-tunes is useful.

Critical weaknesses

The presentation can be improved; it’s very hard to follow. There are many new terms (mean trace, calibrated release, disclosed offset, depth, stem, round), and the main text is mostly dense numbers that point to the appendix.

Examples of why this is a critical safety issue might help. Who is the attacker and who is the auditor in practice? The planted beliefs are low-stakes (e.g., cake baking), and no harmful hidden behavior (e.g., backdoors) is tested.

The results section is hard to understand, e.g., how the subsections relate to each other and what the implications are. It is not clear what a practitioner should do based on these results.

The auditors used are fairly weak (nearest-centroid with two references per class), so it is unclear whether the attacks would work against stronger auditors.

Questions for authors

Can you give a concrete real-world scenario where this attack causes harm?

Would combining several auditors be robust? What does it cost an attacker to evade all of them?

Do the results hold for stronger auditors and larger models?

2 comments
AuthorsOpenReview ↗

Response to Reviewer WK3E

Thanks! You're right that the paper was a bit hard to follow, and we did a major rewrite around this feedback. Presentation-wise, here are the major changes:

  • The abstract is a bit shorter and describes each finding in 1-2 sentences.

  • The introduction now starts by grounding the paper in a concrete setting (e.g. a platform deciding whether to host a fine-tune), and each of the three findings is clearly stated in one bolded sentence.

  • We use plain language names instead of our internal names (mean activation difference, offset norm, question) and added a Table 1 glossary of terms.

  • The Results section starts with a roadmap of four questions. The attack-by-attack details moved to Appendix B, and the main text contains one summary table (Table 2).

  • The Discussion now ends with a paragraph specifically for practitioners.

Q1. You asked: who is the attacker and who is the auditor? What is the harm? The auditor is an entity such as a model hub or regulator. It receives a fine-tuned model from a known open base model, and then audits it using model diffing before listing or certifying it. The attacker is the fine-tuner. For example, it might be a vendor who has fine-tuned a customer-support model to assert a false drug-safety claim, or to misbehave when given a trigger. We show that this attacker can pass the audit by adding only one term to its loss function. The auditor then reads the model as a normal fine-tuned model, and the behavior is still present in the model (see Sections 1 and 2). Beyond low-stakes beliefs, we also added a backdoor whose trigger causes the model to provide medical misinformation. After the attack, the backdoor still fires on every prompt with the trigger, and doesn't fire without the trigger, 3/3 runs. Every auditor except the offset auditors reads the model as clean. The offset norm detects all 3 (see Appendix B.18).

Q2. Is a combination of auditors robust? What is the cost for an attacker to evade all auditors? To address this question, we first created a loss that uses all seven auditors and placed the auditors all in the same loss. We then ran ten different seeds for each attacker capacity. A LoRA attacker was not able to evade the audit and all 20 of its runs were detected. A full-weight attacker was able to evade the audit in 3 of its 20 runs. In those three runs, all seven auditors were able to pass and the implanted belief was kept, at a neutral-text perplexity ratio of at most 1.32. Thus, combining published auditors raises the attacker's cost but does not stop a full-weight attacker. An additional finding in our runs is that the auditors that are not in the loss are able to detect all 3 evaders. We also find that those auditors do not flag the full-weight controls, which carry the belief but do not contain the attack. The practitioner paragraph says this.

Q3. You said we should use a stronger auditor, or a larger model. We experimented with stronger auditors, by using a stronger output classifier (nine references per class instead of two; a 3-nearest-neighbour vote; logistic regression; all 288 benign fine-tunes added as references). None of them name any of the 180 attacked checkpoints that the original auditor doesn't name (Appendix E.2). We tried Diff Mining's own reader, using the default settings. It names every unattacked model and no attacked one in the original cohort (Appendix E.4). We tried directly attacking two stronger auditors: an activation-matched auditor, and a restricted offset-subtracted auditor. Neither of them are evaded when we keep the belief (Appendix E.4).

You said we should use a larger model. At Qwen3-8B, the joint attack evades every auditor except the offset auditors, which detect 6/6; we keep the belief in 4/6 runs (Appendix B.18). Using the same attack at Qwen3-32B (4-bit weights, with three changes to the training recipe driven by memory constraints, which are described in Appendix B.18 and in the amendments section of the supplement), the attack evades the output auditor, the internal-probe auditor, the sample-only auditor, and the contrastive-decoding auditor on 6/6 runs; the offset auditors detect 6/6 runs, and the rank test detects 5/6; the belief is kept in 2/6 runs, and no run evades every auditor.

We have added a table in the general comment section, which maps each request to the section where we address it. We hope that this will resolve the issues behind your presentation and soundness scores. We would be grateful if you could reconsider your rating. We would also be happy to clarify anything further during the discussion!

Thank you for the response and the revised manuscript. The revision is clearly easier to follow. Given the revision and the new results, I will raise my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions and new experiments

We thank the three reviewers for their careful reviews. We have run every experiment the reviews requested, and the results are in the revised PDF and its supplement. Below is a table matching each review's request to where we have addressed it.

RequestWhat we ranResultWhere
Corrected loss (MmDC Q1)All 12 pooled-probe attacks and the joint attack retrained with a stable lossPooled probe still evaded 12/12; belief now kept in 6/12 (was 2/12)App. B.18
Combine auditors; attacker's cost (WK3E Q2, Ry63 Q2)One loss carrying seven auditors, depth included, at LoRA and at full weightsTen seeds per capacity. LoRA: detected 20/20. Full weights: 3/20 runs pass all seven with the belief kept, at a perplexity ratio of at most 1.32; an auditor outside the loss (differential-SAE distance) flags all 3App. B.18
Direct attacks on the strongest alternative auditors (MmDC Q3, Ry63 Q2, WK3E Q3)42 attacked runs against the corrected activation-matched auditor and the offset-subtracted restricted auditor, alone and with depthNeither is evaded with the belief keptApp. E.4
Stronger output auditors (WK3E)Nine references per class; 3-nearest-neighbour, logistic regression, 288 benign referencesNone names any of the 180 evading checkpointsApp. E.2
More beliefs, harmful behavior, other objectives (Ry63 Q1, WK3E Q1)Joint attack on four more beliefs (two never used to tune any attack) and on a medical-misinformation backdoorEvery auditor except the offset auditors is evaded; the offset norm detects 6/6, 6/6 and 3/3; no run on the new beliefs evades every auditor with the belief keptApp. B.18
Larger models (WK3E Q3)Joint attack at Qwen3-8B and at Qwen3-32B (4-bit weights)Qwen3-8B: same pattern; the offset auditors detect 6/6. Qwen3-32B: the offset auditors detect 6/6 and the rank test 5/6; no run evades every auditorApp. B.18
Threshold stability (Ry63 Q3, MmDC Q2)192 held-out benign fine-tunes: six new genres, four base models, thresholds sealed first; then 288 fresh benign fine-tunes read at thresholds recalibrated on every earlier benign run0/48 accused on two models; 8/48 and 9/48 on the two whose thresholds rested on 2 and 12 runs. Recalibrated: 1/288 fresh runs accused (0/72 on three models, 1/72 on Qwen3.5-4B)App. D.3, D.4
Statistical scope (MmDC Q2)Bounds by seed, genre and corpus; three retention floors; a second graderTwo graders make the same keep/lose call on 42/48 and 5/6 runsApp. C.5, C.6, D.1, F.2
Presentation, threat model (WK3E)Shorter abstract; introduction opens with the concrete scenario and states each finding in one bold sentence; glossary, results roadmap, practitioner paragraph, plain names in the main textSec. 1-4

One claim has changed from the original paper. Using our corrected loss function, an attacker who uses full weights can evade all seven targeted auditors at once with a belief that is maintained in 3 of 20 runs, over ten different random seeds (2 of 6 on the first three seeds, and 1 of 14 on seven additional seeds). We have now updated the abstract to reflect that this is a successful attack. All other conclusions are unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.