acceptodds
Withdrawn from ICLR 2027

FG-GUARD: FINE-GRAINED PERTURBATION DEFENSE WITH FULL PATH AUDIO ATTACK ANALYSIS

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Audio-Language Models (ALMs) exhibit a safety inconsistency: a harmful request may be rejected when spoken neutrally yet elicit an unsafe response when the same words are delivered differently. Existing attacks establish this vulnerability empirically, but do not provide a unified account of how an audio manipulation affects the safety of the generated answer. We develop a local, first-order analysis of the complete pathway from acoustic transformation to safety-relevant output. Signal-level transformations first induce structured changes in the log-Mel representation. The ALM and output Jacobians then propagate these changes through the model to a refusal-related opening-token probability. The resulting first-order expression quantifies the signed local change in this continuous proxy. The acoustic transformation determines the time–frequency pattern of the input change, while its effect on the proxy depends on the specific model and input. Motivated by two consequences of the analysis, we propose Fine-Grained Perturbation Defense (FG-Guard). The additive log-Mel formulation motivates a perturbation-based defense in the same representation space, while the joint time–frequency structure of the derived changes motivates a two-dimensional mask over the perturbation. The mask support is estimated empirically from jailbreak and transcription gradients. On Qwen2-Audio-7B-Instruct, FG-Guard reduces AdvWave attack success from 84.61% to 41.03% and AdvBench-Audio from 1.92% to 0.00%.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Audio-Language Model safety, attack and defense

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HsTGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies safety inconsistencies in audio-language models, where the same harmful textual content may receive different responses under different acoustic deliveries. It derives a local first-order analysis that connects acoustic transformations such as pitch, tempo, and noise to changes in the log-Mel representation, propagates these changes through the audio-language model using Jacobians, and maps them to a continuous refusal-opening-token proxy. Based on this analysis, the paper proposes FG-Guard, which learns a reusable perturbation in log-Mel space using a sparse two-dimensional time-frequency mask selected by the ratio between jailbreak and transcription gradients. The perturbation is trained with both refusal and transcription-preservation objectives. Experiments on Qwen2-Audio and Kimi-Audio show meaningful reductions in attack success for several benchmarks, especially on Qwen2-Audio, with relatively small changes in the reported utility metrics. However, the theoretical analysis is only locally validated, and the defense evaluation is limited to non-adaptive, within-family attacks with family-specific perturbations.

Critical strengths
  1. The paper addresses an important and emerging safety problem in audio-language models. FG-Guard provides a concrete input-space defense that is computationally simple at inference time and explicitly attempts to balance safety and transcription preservation.

  2. The experimental section includes several useful baselines and controlled ablations. On Qwen2-Audio, FG-Guard reduces AdvWave SRoA from 84.61% to 41.03% and AdvBench-Audio SRoA from 1.92% to 0.00%, while maintaining a similar AIR-Bench score. The no-mask and no-transcription-loss ablations provide some evidence that both components contribute to the observed behavior.

Critical weaknesses
  1. The central theoretical contribution is not sufficiently connected to the proposed algorithm or empirically validated. Equations (2)–(4) mainly provide a conditional local first-order expansion and chain-rule composition. However, the actual mask is selected using empirical jailbreak/transcription gradient ratios, and the derived acoustic tangents, model Jacobians, and refusal-opening proxy are not used directly to determine the mask or perturbation. The main validation is a qualitative temporal RMS comparison on one utterance; there is no population-level signed test showing that the predicted change in the refusal proxy correlates with the actual model output or final refusal rate. This weakens the paper’s main mechanistic claim.

  2. The robustness evaluation is narrow and potentially optimistic. The method trains one mask and perturbation per attack family and evaluates each family using its corresponding checkpoint. The attacks are not re-optimized against the deployed perturbation, and there is no cross-family or unseen-delivery evaluation. Therefore, the results demonstrate suppression of non-adaptive, within-family attacks rather than robust defense against an adaptive attacker. This limitation is especially important because AdvWave has only 39 examples, while AdvWave and AdvBench-Audio already have 0% attack success on the undefended Kimi-Audio model.

Questions for authors

Clarification requests:

  1. Were every family-specific mask and perturbation trained strictly using the training partitions, with validation data used for hyperparameter selection and no test examples involved in mask construction or perturbation training? Please clarify the exact split used for each defense.

  2. The paper reports utility metrics averaged over four family-specific checkpoints. What would the per-checkpoint safety and utility results be, and how would the method be deployed when the attack family is unknown at inference time?

Material evaluation question:

  1. Can the authors provide a quantitative population-level validation of Eq. (4), such as the correlation between the predicted first-order change in (m_R) and the actual finite-difference change across many inputs and multiple transformations? In particular, does this predicted quantity explain the observed refusal gains of FG-Guard?
Reviewer_kkqfOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper raises question on analyzing and defending against attacks against ALMs, specifically through utilizing different stages and features of the audio input, and how to modify the log-Mel to the input of the ALMs to prevent from the model to answer to harmful request.

Critical strengths
  1. The paper aims to analyze the full pipeline and all of the stages of the audio attack, and utilize this analysis to develop a guardrail against white-box attacks.
  2. The proposed FG-Guard defend successfully against some attacks.
Critical weaknesses
  1. Looking at the main results table, FG-Guard is very model-dependent and also attack-dependent. Only in JALB on Qwen2-Audio was FG-Guard the only defense that scored the lowest SRoA. It is hard to conclude that FG-Guard is a promising method to defend against these existing audio attacks.

  2. The paper only tested with 2 models with a single seeded run, which makes it hard to claim the robustness of the results across different seeds or multiple runs. Also, especially for Kimi-Audio, it already shows that the SRoA score is 0.00 with no defense, which raises doubts about the purpose of the defense in the first place.

  3. The defense works in a white-box setting, which is a very strong assumption in an audio attack scenario, especially in a real-world setting.

Questions for authors
  1. Are the authors aiming to claim that FG-Guard is able to replace existing audio defenses? If so, how do the illustrated results prove the paper's claim?

  2. Did the authors consider any black-box defense methods?

Reviewer_ZaH3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper delves into jailbreak defenses of Audio-Language Models (ALMs). The authors propose FG-Guard, which utilizes a fine-grained two-dimension time-frequency mask to identify regions that are highly sensitive to jailbreak-related objectives while having minimal impact on speech transcription.

Critical strengths
  1. The paper addresses a practically problem by investigating how acoustic variations affect the safety of ALMs while considering the trade-off between safety and utility.

  2. FG-Guard utilizes a fine-grained two-dimension time-frequency perturbation and jointly optimizes refuse and accept objectives, which is simple and concise to understand.

  3. The theoretical analysis, experiments across various models, and ablation studies all demonstrate the effectiveness of FG-Guard.

Critical weaknesses

See questions

Questions for authors

1 .I am very curious about the data in Page 7, Table 1. On Qwen2-Audio, SRoA of FG-Guard is 41.03% while that of ALMGuard is 15.38%. Could the authors provide further insights or analysis into why FG-Guard underperforms ALMGuard in this setting?

  1. Could the authors compare 1D and 2D masks under identical settings to better demonstrate the effectiveness of the proposed 2D mask?

  2. Could the authors consider evaluating FG-Guard against adaptive attacks to further assess its robustness under more challenging scenarios?

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

Related papers

Discussion (0)

Sign in to comment.