acceptodds
Under review as a conference paper at ICLR 2027

The Model’s Tell: Measuring Context-Leakage Attack Signals with Behavior Gauges

Abstract

LLMs increasingly rely on external contexts, such as pre-defined system prompts or retrieved documents, to improve generation quality. However, processing these contexts alongside user queries creates an attack surface: adversarial inputs can induce models to disclose them. Prior probing studies suggest that leakage-related signals emerge in hidden states, yet the need to extract these states poses additional deployment challenges. In this paper, we explore whether this internal signal leaves a more accessible “tell” before decoding. We propose LeakGauge, which probes this response by appending a suffix that gauges leakage behavior and mapping its prefill token probabilities to an attack-risk score. While a direct gauge uses the initial tokens of confidential content, we find that a content-agnostic one that verbalizes leakage behavior yields more robust signals. Across 11 LLMs, including GLM-5.2 (753B) and Kimi-K3 (2.8T), LeakGauge reaches an AUROC range of 0.944–0.996 on unseen attacks. The signal remains stable when the content changes language or the attack shifts from verbatim to semantic disclosure. By activation-steering interventions, we further show that the risk score is sensitive to an internal leakage-related direction, relating the observable signal to the model's internal representation. In addition, \sysname enables an input detector with fewer than 0.5K extra parameters and 10.34 ms of added latency, 77% lower than activation-based baselines. Code: https://anonymous.4open.science/r/leakgauge-1190.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Secutiry, privacy, leakage

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_H7hkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces LeakGauge, a method for detecting context-leakage attacks before an LLM generates a response. It appends a short gauge sequence to the input and feeds the sequence’s prefill token log-probabilities into a lightweight classifier. The authors compare content-specific Exact gauges with content-agnostic Behavior gauges and find that Behavior generally provides stronger generalization. Experiments cover eleven models, system-prompt and RAG leakage, and held-out attacks and protected content. Cross-lingual tests, cross-objective transfer, and activation-steering experiments further investigate the signal, while a deployment experiment reports 10.34 ms of additional latency and fewer than 500 classifier parameters. Additional stress tests demonstrate both sensitivity to hard negatives and vulnerability to detector-aware attacks, with alternative and combined gauges providing partial mitigation.

Critical strengths
  1. Simple and potentially useful detection mechanism. LeakGauge extracts a useful signal through prefill probabilities without requiring access to intermediate activations or autoregressive generation, offering a practical alternative to existing model-derived detectors.

  2. Broad empirical investigation. The evaluation spans eleven models, two leakage settings, multiple distribution shifts, and several controls, while activation steering provides additional evidence connecting the observable score to leakage-related model behavior.

Critical weaknesses
  1. The deployment reliability of the default gauge remains limited. Table 20 reports a 22.18% hard-negative false-positive rate on GLM-5.2, and Table 21 shows only 8.0% detection recall against successful leaks under single-probe-aware optimization. Alternative gauges and a two-probe detector improve these results, but their combined false-positive rates and deployment costs are not established; although the white-box test exceeds the main threat model, it reveals an important limitation of a fixed exposed detector.
  2. The reported operating-point performance requires clarification. Table 8 reports combinations such as F1 of 0.993 and TPR@FPR5 of 0.963 for Gemma-4-E4B on the RAG in-distribution test, which appear difficult to reconcile under standard binary metrics on the same predictions and the stated class proportions. Furthermore, best F1 over thresholds does not establish performance with a threshold selected independently of the test set, leaving uncertainty about practical deployment performance.
Questions for authors
  1. How are F1 and TPR@FPR5 in Table 8 calculated and aggregated across seeds? Please clarify the example identified above and provide the corresponding per-seed operating points or confusion matrices.

  2. How are thresholds selected for the deployment and stress-test evaluations? What performance is obtained when a validation-selected threshold is frozen across ordinary benign inputs, hard negatives, and held-out attacks?

  3. What are the hard-negative false-positive rate and additional latency of the two-probe configuration in Table 21?

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer H7hk (1/2)

Dear Reviewer H7hk, thank you very much for your careful reading and constructive feedback. We appreciate your recognition of our simple and practical detection mechanism and broad empirical investigation. We hope the following responses address your concerns and clarify the evaluation protocol and deployment results.


W1 & Q2 & Q3. Request for settings and results of the stress tests in Appendix G.

W1. The deployment reliability of the default gauge remains limited. Table 20 reports a 22.18% hard-negative false-positive rate on GLM-5.2, and Table 21 shows only 8.0% detection recall against successful leaks under single-probe-aware optimization. Alternative gauges and a two-probe detector improve these results, but their combined false-positive rates and deployment costs are not established; although the white-box test exceeds the main threat model, it reveals an important limitation of a fixed exposed detector.

Q2. How are thresholds selected for the deployment and stress-test evaluations? What performance is obtained when a validation-selected threshold is frozen across ordinary benign inputs, hard negatives, and held-out attacks?

Q3. What are the hard-negative false-positive rate and additional latency of the two-probe configuration in Table 21?

Reply to W1 & Q2 & Q3: Thank you for highlighting this concern. As you noted, the white-box GCG test goes beyond our main threat model; it characterizes the failure boundary of an exposed detector. For this additional comparison, to directly calibrate the FPRs, we select a single threshold for max⁡(p1,p2)\max(p_1, p_2) on the validation set and freeze it before evaluation. Using the same GCG attack suffixes as in Table 21, this configuration achieves 81.5% detection recall and 4.45% post-filtered ASR.

Table R4. Llama-3.1-8B-Instruct, system-prompt leakage. One- and two-probe detectors are evaluated under single- and two-probe-aware GCG attacks, respectively. (Now included as Table 22 in Appendix G.2 of the revision.)

DetectorThresholdBenign FPRHard-neg. FPRGCG recallPost-filtered ASR
One probe0.50317.61%5.64%8.03%27.52%
Two probes0.81423.46%6.73%81.5%4.45%

Table R4 reports benign and hard-negative FPRs of 3.46% and 6.73% for the two-probe configuration, together with 81.5% recall against successful two-probe-aware GCG leaks. Scoring both gauges sequentially takes 22.6 ms, compared with 10.34 ms for one probe, and remains below Attention-Tracker (46.03 ms) and I'vDtL (70.91 ms) in our deployment setup. For held-out attack evaluation, Table 8 separately reports frozen-threshold F1 scores of 0.953 on Held-Out Attacks and 0.944 on Held-Out Strict for Llama-3.1-8B-Instruct in the system-prompt setting.

Table 20 shows that gauge design can substantially improve hard-negative robustness. On GLM-5.2, a pre-specified user-intent gauge reduces FPR from 22.18% to 9.64% while increasing attack TPR from 84.93% to 95.76%, without tuning on the hard-negative set. This improvement in both metrics supports gauge selection as a practical mitigation.

AuthorsOpenReview ↗

Author Response to Reviewer H7hk (2/2)

W2 & Q1. Clarification on metric computation and frozen-threshold performance.

W2. The reported operating-point performance requires clarification. Table 8 reports combinations such as F1 of 0.993 and TPR@FPR5 of 0.963 for Gemma-4-E4B on the RAG in-distribution test, which appear difficult to reconcile under standard binary metrics on the same predictions and the stated class proportions. Furthermore, best F1 over thresholds does not establish performance with a threshold selected independently of the test set, leaving uncertainty about practical deployment performance.

Q1. How are F1 and TPR@FPR5 in Table 8 calculated and aggregated across seeds? Please clarify the example identified above and provide the corresponding per-seed operating points or confusion matrices.

Reply to W2 & Q1: Thank you for checking this carefully, and we are sorry for causing ambiguities in Appendix Table 8.

  • Resolving the underspecification of "best F1". As we aim to measure a deployment-oriented detection performance, we rigorously curate the validation set, which shares no samples with the test set. Concretely, the "best-F1" threshold is tuned on the validation set, while the data in column F1 is computed on the test set at this frozen "best-F1" threshold. Thus, F1 in Table 8 already reflects performance at a threshold fixed before testing. As the test set is never used for threshold selection, the best F1 threshold on val set does not necessarily equal the best on test set.
  • Double-checking the results. In contrast, TPR@FPR5 is the highest test TPR among thresholds whose test FPR is at most 5%. The two metrics are therefore reported at different thresholds: the frozen threshold can have a test FPR above or below 5%. As described in Appendix B.1, both metrics are computed per run and then averaged over five runs. Table R3 reports per-run confusion counts and F1 at the frozen threshold, showing that the gap in the Gemma-4-E4B-it comes from runs 4 and 5, whose frozen thresholds operate at a test FPR above 5% (5.6% and 5.4%).

Table R3. Per-run confusion counts and F1 at the frozen threshold, together with TPR@FPR5. Gemma-4-E4B-it, RAG, In-Dist (2,074 attack / 1,040 benign).

RunTPFPTNFNF1TPR@FPR5
120682103860.9980.997
220650104090.9980.996
3205801040160.9960.994
420735898210.9860.914
520745698400.9870.913
Mean0.9930.963

We have corrected the caption of Table 8 and Appendix B.1 in the revision.

Thanks for your reply. My concerns have addressed. I have raised the rating to 3.

AuthorsOpenReview ↗

Thank You for Your Response

Thank you for your reply and for raising your rating. We are glad that our response addressed your concerns.

Reviewer_eo2qOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LeakGauge detects context-leakage attacks by appending a short behavioral probe and classifying features derived from its token probabilities. The paper evaluates system-prompt and RAG leakage across 11 open models, with held-out contexts, attack templates, languages, and objectives. Its main contribution is evidence that leakage-related behavior can be detected through the probability interface, together with an efficient implementation that reuses the context prefix.

Critical strengths

The probability-based interface is a useful engineering contribution. It avoids extracting hidden states and allows prefix-cache reuse. The held-out-context and held-out-template evaluations test meaningful forms of distribution shift.

The appendices add substantive controls: unrelated suffixes and randomized labels, hard benign queries, and detector-aware attacks. These experiments help distinguish a leakage-related signal from a classifier that simply recognizes the original attack templates.

Critical weaknesses
  1. The reported operating-point metrics are inconsistent under their stated definitions. Table 8 defines F1 as the best F1 over thresholds. For GLM-5.2 on the system-prompt Held-Out Strict split, it reports F1 = 0.921 and TPR = 0.966 at 5% FPR. With the 1,600 attack and 1,472 benign examples in Table 1, that operating point already gives binary F1 of approximately 0.960. The best F1 cannot be lower. The same issue appears for Llama-3.1-8B on RAG In-Dist: the 2,074 attack and 1,040 benign examples in Table 7 and TPR = 0.998 imply F1 of approximately 0.987 at 5% FPR, above the reported best F1 of 0.973. Different definitions or aggregation could change this calculation, but the paper must specify them and reconcile these entries. This directly affects the interpretation of its headline detection results.

  2. The proposed defense lacks a matched assessment of false alarms at its deployed operating point. Best-threshold F1 and TPR read from a test ROC do not establish performance at a threshold calibrated once and then frozen. Appendix G.2 combines two probes with an OR rule and reports improved resistance to adaptive attacks, without reporting the corresponding joint benign false-positive rate. This omission matters because Appendix G.1 reports a 22.18% false-positive rate for the default GLM probe on hard negatives. The single-probe and two-probe defenses need comparison at the same total false-positive budget, including hard benign queries. Otherwise the benefit of adding a probe cannot be separated from a more aggressive blocking policy.

I recommend weak rejection because the headline metrics cannot be reconciled as reported and the defense lacks a comparison at a matched false-positive rate.

Questions for authors
  1. What exact F1 definition, threshold-selection procedure, and aggregation produce the two Table 8 examples above? Please provide the corresponding split sizes and confusion counts for each run, and correct the affected results.

  2. With calibration separated from testing and thresholds frozen, how do one-probe and two-probe defenses compare at the same total benign false-positive rate? Results on ordinary benign queries, hard negatives, and attacks aware of both probes would determine whether the deployment claim holds.

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer eo2q (1/2)

Dear Reviewer eo2q, thank you very much for your careful review and constructive feedback. We appreciate your recognition of our probability-based interface as a useful contribution, generalization evaluations, and substantive experimental controls. We hope the following responses address your concerns and clarify the metric definitions and evaluation protocol.


W1 & Q1. Clarification on the F1 threshold in Appendix Table 8.

W1. The reported operating-point metrics are inconsistent under their stated definitions. Table 8 defines F1 as the best F1 over thresholds. For GLM-5.2 on the system-prompt Held-Out Strict split, it reports F1 = 0.921 and TPR = 0.966 at 5% FPR. With the 1,600 attack and 1,472 benign examples in Table 1, that operating point already gives binary F1 of approximately 0.960. The best F1 cannot be lower. The same issue appears for Llama-3.1-8B on RAG In-Dist: the 2,074 attack and 1,040 benign examples in Table 7 and TPR = 0.998 imply F1 of approximately 0.987 at 5% FPR, above the reported best F1 of 0.973. Different definitions or aggregation could change this calculation, but the paper must specify them and reconcile these entries. This directly affects the interpretation of its headline detection results.

Q1. What exact F1 definition, threshold-selection procedure, and aggregation produce the two Table 8 examples above? Please provide the corresponding split sizes and confusion counts for each run, and correct the affected results.

Reply to W1 & Q1: Thank you for checking this carefully, and we are sorry for causing ambiguities in Appendix Table 8.

  • Resolving the underspecification of "best F1". As we aim to measure a deployment-oriented detection performance, we rigorously curate the validation set, which shares no samples with the test set. Concretely, the "best-F1" threshold is tuned on the validation set, while the data in column F1 is computed on the test set at this frozen "best-F1" threshold. As the test set is never used for threshold selection, the best F1 threshold on val set does not necessarily equal the best on test set, which thus leads to the reviewer's observation.
  • Double-checking the results. It therefore does not necessarily equal the highest-scoring F1 on the test set and thus can be different from the F1 score corresponding to the TPR@5%FPR's threshold. Both metrics are averaged over five runs. Table R1 reports per-run confusion counts and F1 at the frozen threshold, together with TPR@FPR5.

Table R1. Per-run confusion counts and F1 at the frozen threshold, together with TPR@FPR5. Left: GLM-5.2, System Prompt, Held-Out Strict (1,600 attack / 1,472 benign). Right: Llama-3.1-8B, RAG, In-Dist (2,074 attack / 1,040 benign).

RunTPFPTNFNF1TPR@FPR5TPFPTNFNF1TPR@FPR5
113985114212020.91700.9656207211192920.97350.9990
214236014121770.92310.9663202251989520.97520.9986
314035514171970.91760.9675200950990650.97220.9957
414065314191940.91930.9669207313190910.96910.9995
514416614061590.92760.9631206810993160.97290.9971
Mean0.92090.96590.97260.9980

We have corrected the caption of Table 8 and Appendix B.1 in the revision.

AuthorsOpenReview ↗

Author Response to Reviewer eo2q (2/2)

W2 & Q2. Request for a comparison of one- and two-probe settings in Appendix G.

W2. The proposed defense lacks a matched assessment of false alarms at its deployed operating point. Best-threshold F1 and TPR read from a test ROC do not establish performance at a threshold calibrated once and then frozen. Appendix G.2 combines two probes with an OR rule and reports improved resistance to adaptive attacks, without reporting the corresponding joint benign false-positive rate. This omission matters because Appendix G.1 reports a 22.18% false-positive rate for the default GLM probe on hard negatives. The single-probe and two-probe defenses need comparison at the same total false-positive budget, including hard benign queries. Otherwise the benefit of adding a probe cannot be separated from a more aggressive blocking policy.

Q2. With calibration separated from testing and thresholds frozen, how do one-probe and two-probe defenses compare at the same total benign false-positive rate? Results on ordinary benign queries, hard negatives, and attacks aware of both probes would determine whether the deployment claim holds.

Reply to W2 & Q2: Thank you for the practical suggestions! We note that this white-box GCG test goes beyond our main threat model and characterizes the failure boundary of exposed detectors. Table R2 reports the joint benign and hard-negative FPRs of one-probe and two-probe defenses on Llama-3.1-8B-Instruct. For this additional comparison, we apply a single threshold to max⁡(p1,p2)\max(p_1, p_2) to directly calibrate the FPRs.

Following your requests, Table R2 answers two questions with two threshold settings:

  • (1) Frozen thresholds: what FPRs do the detectors incur when calibration is separated from testing? Thresholds are selected on the validation set and frozen before evaluation, reflecting deployment performance.
  • (2) A common FPR budget: how do the two detectors compare under a common FPR budget? Thresholds are selected post hoc on the evaluation negatives so that each detector's FPR is at most 5% on both ordinary benign queries and hard negatives, providing a controlled comparison.

Table R2. Llama-3.1-8B-Instruct, system-prompt leakage. (Now included as Table 22 in Appendix G.2 of the revision.)

Threshold selectionDetectorThresholdBenign FPRHard-neg. FPRGCG recallPost-filtered ASR
Validation set, frozenOne probe0.50317.61%5.64%8.03%27.52%
Validation set, frozenTwo probes0.81423.46%6.73%81.5%4.45%
Common FPR budget (≤ 5%)One probe0.57102.79%4.73%4.82%28.49%
Common FPR budget (≤ 5%)Two probes0.85043.40%4.91%71.0%6.97%

With frozen thresholds, the two-probe detector has a lower benign FPR and a hard-negative FPR only 1.09 percentage points higher, while raising GCG recall from 8.03% to 81.5%. When both detectors operate within the 5% budget on ordinary benign and hard-negative queries, it still raises GCG recall from 4.82% to 71.0% and reduces post-filtered ASR from 28.49% to 6.97%. These results suggest that the gain comes from improved detection rather than more aggressive blocking.

Thank you for the response and the revision. The clarification of the F1 definition, together with the per-run confusion counts, resolves the metric inconsistency I raised, and Table 22 provides the joint false-positive comparison I asked for. I am raising my rating to 3. For the final version, please align the "best F1" wording in Section 6 and Table 5 with the corrected definition, and summarize the hard-negative and exposed-probe limitations from Appendix G in the main text.

AuthorsOpenReview ↗

Thank You for Your Response

Thank you for your reply and for raising your rating. We are glad that our response addressed your concerns. We will revise the final version as requested.

Reviewer_oKAUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes LeakGauge, which detects system-prompt and RAG-context leakage attacks using the target model's token probabilities during prefill. Its core contribution is to turn the target model's token probabilities for descriptions of specific behaviors into low-cost attack-detection signals available before generation. The main experiments measure attack-input recognition rather than directly predicting whether each request will actually cause leakage. Oveall, this manuscript is well-motivated and organized.

Critical strengths
  1. Nice idea. LeakGauge uses the per-token log-probabilities of a short gauge to identify leakage attacks without generating responses or extracting intermediate-layer activations. Its additional classifier has fewer than 0.5K parameters. This turns the target model's own signal into a readily deployable detector.

2.Good experiements. Table 2 covers more than ten models and both system-prompt and RAG leakage, with separate evaluations on unseen content, unseen attacks, and both simultaneously. The performance achieves an AUROC of at least 0.94 across these settings. Extensive evaluation further supports its effectiveness.

Critical weaknesses

Unclear experimental settings:

  1. How are the detection thresholds determined? Are the thresholds used in Tables 20 and 21 selected on independent validation data and then frozen?
  2. What information and optimization settings are used in the adaptive attack evaluation?
  3. In the cost comparison, which advantages arise from the algorithms themselves, and which arise from serving-interface limitations?
Questions for authors

See Weaknesses.

3 comments
AuthorsOpenReview ↗

Author Response to Reviewer oKAU

Dear Reviewer oKAU, thank you very much for your careful review of our paper. We are encouraged by your recognition of our core idea of using token log-probabilities and our thorough experimental evaluation. We hope the following responses address your concerns and clarify the experimental settings.


W1. Question about the threshold for Appendix Table 20 and 21.

How are the detection thresholds determined? Are the thresholds used in Tables 20 and 21 selected on independent validation data and then frozen?

Reply to W1: Thank you for pointing this out! Our experiments in Appendix Table 20 and 21 mainly intend to reflect a deployment-ready performance. Concretely, we select the F1-maximizing threshold on validation set and then transfer it for evaluation without additional adjustment. The validation set shares no samples with any test or stress-test data.

We have clarified this setting in the revision's Appendix G.


W2. Question about the adaptive attack settings in Appendix G.2.

What information and optimization settings are used in the adaptive attack evaluation?

Reply to W2: Thank you for the thoughtful question. We do this white-box stress test beyond our primary threat model: the attacker has full access to the target model, the protected system prompt, and the deployed probe(s). GCG optimizes a 20-token suffix for 300 iterations to minimize the NLL of the protected content plus the probe-predicted attack probability. At each iteration, it samples 64 candidates from the top-128 gradient-based token replacements per position.

We have clarified the settings in the revision's Appendix G.2.


W3. Question about the sources of the cost advantages in Table 5.

In the cost comparison, which advantages arise from the algorithms themselves, and which arise from serving-interface limitations?

Reply to W3: Thank you for raising this point. We appreciate the opportunity to better compare LeakGauge with baselines regarding cost efficiency.

  • Efficiency advantages of LeakGauge. LeakGauge has two inherent advantages: (1) generation-free scoring of a short gauge, and (2) lower deployment costs, as it uses the log-probabilities, which are existingly supported by mainstream serving frameworks like vLLM and SGLang, and does not require internal model states.
  • What about baselines. As detailed in Appendix E.2, the additional 8B replicas required by baselines, and part of their latency, stem from serving-interface limitations. Eliminating these replicas would require additional engineering to expose internal signals in the serving stack.
AuthorsOpenReview ↗

Kind Reminder to Reviewer oKAU

Dear Reviewer oKAU,

Thank you again for your positive assessment of our work and for the time you have spent reviewing it. We understand that this is a busy period. If you have a chance before the discussion period ends, we would be grateful if you could take a look at our response.

For your convenience, we have summarized our replies to your three questions below.

  • W1: Threshold selection. The thresholds used in Tables 20 and 21 are selected by maximizing F1 on an independent validation set and then frozen for evaluation. The validation set shares no samples with the test or stress-test data. We have clarified this protocol in Appendix G.
  • W2: Adaptive attack settings. We specify the attacker's access and the GCG optimization settings in Appendix G.2. The attacker has access to the target model, the protected system prompt, and the deployed probes. GCG optimizes a 20-token suffix for 300 iterations, jointly targeting content leakage and detector evasion.
  • W3: Sources of the cost advantages. We distinguish the benefits of generation-free scoring and the existing log-probability interface from costs caused by serving-interface limitations. As detailed in Appendix E.2, the additional model replicas required by the baselines, and part of their latency, arise from limitations in exposing internal signals through the serving stack.

We would appreciate hearing whether our response addresses your questions.

Best regards,

The authors of Submission 1190

Reviewer_oKAUOpenReview ↗

Thanks for you response.

Appreciate the authors detailed responses. I am satisfied with their clarifications.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We sincerely appreciate your valuable time and effort spent reviewing our manuscript.

We are encouraged by the reviewers’ positive assessments, including their remarks on the paper’s nice idea (oKAU), useful probability-based interface (eo2q), deployable detection mechanism (oKAU & H7hk), and meaningful distribution-shift evaluations and substantive controls (eo2q).

Following the discussion, all three reviewers confirmed that our responses addressed their questions and concerns. We thank Reviewers eo2q and H7hk for raising their ratings to 3, and all three reviewers for their feedback, which helped clarify the evaluation protocols and better communicate the scope of our approach.

In response to the reviewers’ comments, we provided the following clarifications and additional evaluations during the rebuttal:

  • Threshold protocol and metrics. We clarified that the F1 threshold is selected on an independent validation set and frozen for testing, and provided per-run confusion counts to clarify the reported results in Table 8 (eo2q, H7hk).
  • One- and two-probe evaluation. We added comparisons under a matched false-positive budget, including hard negatives, and reported joint false-positive rates and deployment costs in Table 22 (eo2q, H7hk).
  • Stress-test settings and efficiency analysis. We specified the adaptive-attack settings in Appendix G.2 and distinguished the method’s efficiency gains from serving-interface limitations in Appendix E.2 (oKAU). We also clarified the relationship between the white-box stress test and our primary threat model (H7hk).

Thank you for your time and consideration.

The authors of Submission 1190

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.