acceptodds
Under review as a conference paper at ICLR 2027

Hidden in Plain Sight: Advancing Harmful Content Understanding via Perception-Aware Adversarial Self-Play

Abstract

Despite significant advances in language understanding, Large Language Models (LLMs) can still miss meaning hidden in plain sight, failing to identify harmful intent in expressions readily understood by humans. Malicious users exploit this gap by strategically disguising harmful content to evade automated moderation, posing a serious threat to online platform safety. The vulnerability reflects a fundamental cognitive asymmetry between human readers, who integrate contextual, phonetic, and visual cues to decipher the intent behind deliberate obfuscation, and text-based LLMs, which rely primarily on token sequences and have limited exposure to evolving evasions. To bridge this asymmetry, we propose PASP (Perception-Aware Adversarial Self-Play), which combines multi-perceptual views with adversarial self-play to improve understanding of harmful intent as evasions evolve. Specifically, PASP constructs textual, phonetic, and visual views of each input, preserving its original context while exposing pronunciation and glyph cues obscured in token space. During adversarial self-play, an attacker uses defender feedback to generate intent-preserving evasions, while the defender is trained using an intent-invariance reward to understand harmful intent across changing surface forms. Extensive experiments show that PASP reduces adaptive attack success rates by up to 73.71% and 74.49% on ChineseHarm-Bench and EvoHarmBench, respectively, preserves model capabilities with minimal degradation, and effectively interprets real-world evasive expressions in e-commerce.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Content Safety, Adversarial Expressions, Adversarial Self-Play

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_uM6mOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines harmful content within evasive expressions and proposes PASP to improve LLM robustness. PASP combines (i) multi-perceptual inputs that expose textual, phonetic, and visual cues, and (ii) adversarial self-play, where an attacker iteratively generates intent-preserving evasions based on defender feedback while the defender is optimized with an intent-invariance objective. Experiments on ChineseHarm-Bench, EvoHarmBench, and in-the-wild e-commerce data show substantial improvements in robustness against adaptive attacks while largely preserving general model capabilities

Critical strengths
  1. The motivation and method are clear, and the visualization is helpful.

  2. The authors provide sufficient evidence to support the effectiveness of their proposed method.


  3. This paper is well-written, making it easy to follow.

Critical weaknesses
  1. Incremental methodological novelty. The two major components of PASP—perception-aware cues (e.g., phonetic/glyph information) and adversarial self-play—are closely related to existing work. The paper should more clearly articulate the technical novelty beyond combining these ingredients with an intent-invariance reward.

  2. Limited baselines and lack of frontier-model comparison. Strong contemporary safety models such as WildGuard, LlamaGuard, ShieldGemma, Qwen3Guard, or Shieldstral are not evaluated [1-5]. Moreover, zero-shot evaluation with strong frontier general-purpose LLMs (e.g., GPT-6 Astra) would provide an important reference for assessing whether PASP offers benefits beyond task-specific safety fine-tuning.

  3. Limited evidence of in-domain generalization. The experiments mainly rely on within-benchmark train/test splits, while the gains on the in-the-wild data are considerably more modest. It therefore remains unclear whether PASP generalizes beyond the benchmark distributions. Stronger cross-dataset evaluation (e.g., training on ChineseHarm-Bench and testing on EvoHarmBench), as well as evaluation on English harmful-content detection ( represents a fundamental setting for general-purpose content moderation) would better establish its practical generalizability.

  4. Lack of cross-lingual OOD generalization. Beyond in-domain generalization, it remains unclear whether PASP transfers to substantially different linguistic systems. Since the current implementation relies on Chinese-specific cues such as Pinyin and character-level obfuscation, evaluation on OOD languages with different writing and phonological systems, such as Korean or Arabic, would help determine whether PASP learns a language-agnostic perception-aware mechanism or primarily Chinese-specific robustness.

[1] Wildguard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of llms
[2] Llama guard: Llm-based input-output safeguard for human-ai conversations
[3] Shieldgemma: Generative ai content moderation based on gemma
[4] Qwen3guard technical report
[5] Shieldstral

Questions for authors

Refer to weaknesses.

Reviewer_ne42OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PASP for recognizing harmful intent concealed by human-readable obfuscations. It combines textual, Pinyin, and rendered-image inputs with feedback-guided adversarial search and intent-supervised GRPO training, including transformed benign examples to discourage over-moderation. Experiments on two Chinese benchmarks and two multimodal backbones report substantial reductions in adaptive attack success, supplemented by held-out attack-family, unseen-attacker, capability-retention, and e-commerce evaluations.

Critical strengths
  1. PASP targets a consequential problem distinct from conventional jailbreak defense: understanding disguised harmful user content rather than preventing harmful model outputs. Combining complementary perceptual cues with newly generated attacks is well motivated, while similarly transformed benign examples explicitly address the risk of treating obfuscation itself as evidence of harm.
  2. The empirical improvements are substantial and reasonably broad. Table 2 shows consistently lower adaptive ASR across both benchmarks and backbones; the held-out-family and unseen-attacker evaluations provide meaningful evidence beyond the training opponent. The marketplace evaluation also offers useful real-world support, although the brand-interpretation scores remain modest in absolute terms.
Critical weaknesses
  1. The individual building blocks have substantial precedent: TextShield exploits phonetic and glyph information, the cited Pinyin-based work addresses phonetic camouflage, and adversarial self-play, GRPO, and conditional reference injection are established techniques. Their integration for disguised-content moderation may still be a useful contribution, but its distinctive benefit is incompletely isolated. Table 5 already compares perception-aware SFT, perception-aware offline training, and text-only self-play, yet does not separate phonetic and visual inputs, the intent reward, or reference injection; there is also no perception-matched PICL comparison. Thus, the incremental contribution of the intent-invariance objective remains unclear.
  2. Attack validity and robustness comparability need stronger verification. In the stratified validation samples of Appendix B.3, 66/250 (26.4%) and 37/250 (14.8%) of judge-accepted rewrites are human-invalid, while many judge-rejected rewrites are human-valid; these figures do not establish the error rates specifically among successful attacks against each defender. DeepSeek also constructs reference intents and scores intent alignment, raising a risk of correlated errors rather than proving evaluation bias. Conditioning ASR on initially correct examples is legitimate, but creates potentially different source subsets; defender-specific strategies likewise yield different benign FPR test distributions. Without per-defender validation and uncertainty estimates, the direction and size of these effects remain unresolved.
Questions for authors
  1. Table 5's online AT is described as text-only self-play, not a separate online SFT control. Would PASP still outperform perception-matched PICL and perception-aware online adversarial SFT under matched data and training budgets? Intent-reward and injection ablations would be particularly informative for attributing the improvement.
  2. Do the ASR rankings persist when adaptive attacks are rerun on the intersection of initially correct source examples, with human validation of successful rewrites for each defender? Please report effective denominators, and test benign FPR on a shared validated set. A shared rewrite pool could additionally assess transfer robustness, but should not replace defender-adaptive evaluation.
  3. Section 3.3.2 explicitly adopts S-GRPO-style reference injection; this is not inherently an invalid training procedure. How are old-policy log probabilities and token-wise advantages computed for the injected response in Eq. (5), and what fraction of groups require injection? An injection-free comparison would clarify how much of the gain comes from this supervised anchor rather than the intent reward.
Reviewer_8NEQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses an important challenge in harmful content understanding: the underlying harmful intent may remain unchanged even when the surface form of a text is deliberately altered. The authors propose PASP, a framework designed to improve the robustness of LLMs in understanding harmful content under adversarial text obfuscation. The proposed method combines multi-perceptual input representations with feedback-guided adversarial self-play. Specifically, PASP augments the original input with three complementary views: textual, phonetic, and visual. The phonetic view is constructed through Pinyin conversion, while the visual view is obtained by rendering the text as an image to preserve character shapes and spatial layouts. Meanwhile, the attacker iteratively generates adversarial expressions that preserve the original intent based on feedback from the defender. The defender is optimized using a reward function that incorporates intent alignment, moderation correctness, and output format validity.

Critical strengths
  1. Unlike existing methods such as PICL, PASP uses the underlying intent of the original text as a supervision anchor and directly rewards the model for accurately recovering the latent intent. This design encourages the defender to achieve semantic correctness rather than merely producing consistent predictions across different surface forms.

  2. The attacker in PASP receives not only the defender's risk classification results but also its explicit interpretation of the input's underlying intent, which can be used to refine subsequent attack strategies. Compared with attack search methods that rely solely on binary success signals, this mechanism provides richer feedback, enabling the attacker to target specific errors in the defender's semantic understanding and generate more informative adversarial training examples.

Critical weaknesses
  1. PASP directly inherits the original text's intent and risk label as supervision targets for adversarial rewrites. However, complex semantic obfuscation, metaphors, and contextual restructuring may alter the actual meaning of the original expression. Since the validity of these rewrites is primarily assessed by an LLM judge, whose precision on human-validated accepted samples is only 79.4%, incorrect judgments about intent preservation may introduce noisy supervision signals and undermine the reliability of intent-invariant learning.

  2. The authors provide the defender with textual, phonetic, and visual representations simultaneously. However, the reliability of these views may vary substantially across different obfuscation strategies, potentially introducing conflicting or ambiguous information. The proposed framework does not explicitly model the reliability of individual views or how inconsistencies across them should be resolved, which may limit its ability to effectively integrate complementary perceptual cues.

Questions for authors
  1. Given that the LLM judge achieves only 79.4% precision on accepted adversarial rewrites, how does PASP mitigate the impact of incorrectly preserved intents on the reliability of its training signals?

  2. How does PASP handle potential ambiguities or conflicts among textual, phonetic, and visual views, particularly when their reliability varies across different obfuscation strategies?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.