acceptodds
Under review as a conference paper at ICLR 2027

PHASE: Packet-level Hierarchical Adversarial Semantic Exploitation for Targeted Steering of Retrieval-Augmented Diffusion Models

Abstract

Retrieval-augmented diffusion models (RDMs) enhance generation by conditioning the denoising process on a small set of retrieved examples. However, the dependence on external retrieved context introduces additional security risks at the retrieval–generation interface. Existing studies on RDM security mainly focus on manipulating the knowledge base, retriever, or persistent trigger mechanism, while the security implications of the retrieved conditioning packet itself remain insufficiently studied. In this work, we investigate packet-level targeted steering in frozen RDMs and propose PHASE (Packet-level Hierarchical Adversarial Semantic Exploitation), a training-free and gradient-free adversarial framework that operates solely on the conditioning packet. Our central observation is that retrieval-space relevance is not a reliable indicator of generation-time influence: highly target-aligned retrieved examples do not necessarily form the packet that most effectively steers the downstream generation process. Based on this observation, PHASE combines heterogeneous candidate construction, source-aware semantic scoring, structured packet composition, and generation-aware probe selection to identify effective adversarial conditioning under a frozen generator. Extensive experiments across diverse generation settings, semantic evaluators, packet configurations, and retrieval conditions consistently demonstrate the effectiveness and robustness of the proposed framework. Further analyses reveal non-trivial interactions among packet members and a trade-off between retrieval compatibility and downstream steering effectiveness. These results highlight the need to jointly consider retrieval composition and generation-time behavior when evaluating the security of retrieval-augmented generative models.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
retrieval-augmented diffusion models, adversarial attacks, targeted steering, multimodal security

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

PHASE studies targeted steering of retrieval-augmented diffusion models by replacing the four retrieved conditioning embeddings supplied to a frozen generator. It builds a pool of candidate images, scores them in the retrieval space, assembles several candidate packets, and chooses one by checking the generated outputs. The evaluation uses one frozen RDM and 80 source–target attack instances, with additional analyses across seeds, samplers, evaluators, and packet configurations.

Critical strengths
  • Checking the actual generated outputs is a sensible way to study packet selection. The observation that retrieval similarity does not reliably predict the generator’s response is useful.
  • The held-out seed/sampler evaluation and native-retrieval diagnostics help clarify the scope of the results. I also appreciate that the authors report negative findings and acknowledge the uncertainty in some of the statistical comparisons.
Critical weaknesses
  1. My main concern is how the attacker obtains the assumed access. Section 3.1 does state that the attacker can replace the conditioning packet and query the frozen generator, so the issue is not simply a missing threat-model section. What is missing is a concrete system in which those permissions are available. The introduction mentions external services, middleware, and caches, while Appendix A mentions user-controlled repositories, but controlling a repository is different from replacing all four embeddings after retrieval. Can the attacker submit arbitrary packets to the victim generator and observe its outputs? What access do they have to the native encoder and sampling settings? Appendix H makes this distinction consequential: direct insertion gives only 2.50% Hit@4 and a 1.25-point ASR gain. The main result therefore depends heavily on direct control of the conditioning interface, whose trust boundary is never clearly established.

  2. I do not understand the separation between the main baselines and the “target-prior stress tests.” TopTargetSim reaches 47.50% primary ASR, compared with 41.67% for PHASE, yet it is set aside as requiring stronger assumptions. Under the stated threat model, the attacker already constructs the whole packet, and there is no explicit minimum query similarity or limit on target-class members. RandomTarget also uses four target-class neighbors and is counted as a matched baseline. More target-dominated conditioning is a property of the resulting packet, not by itself a different attacker capability. If PHASE is solving a constrained problem, those constraints need to be stated and applied to every baseline. Otherwise, the stronger result from this simpler selection rule leaves the benefit of PHASE unclear.

  3. The ablations do not establish the value of most of the proposed construction strategy. Tables 20 and 21 measure similarities of selected packets, rather than attack success on generated images. This is a weak basis for conclusions about which scoring terms or candidate families improve steering, especially when the paper’s premise is that retrieval-space scores are unreliable proxies for generation. StaticOnly does test the benefit of probing, but it removes generator queries altogether. I would like to see a simple packet search given the same candidate pool and generator-query budget. The advantage over JOB-style is also uncertain: the primary gain is 5.00 points with a 95% interval of [-5.42, 15.42] in Table 6. The results suggest that probing helps, but do not yet make a convincing case for the full hierarchy.

  4. The semantic evaluation is less robust than the main claims suggest. Table 2 shows clear gains within the CLIP family, but almost no gain under SigLIP. The explanation based on the recognizability threshold is incomplete: SigLIP target top-5 changes only from 24.58% to 26.25%, target-beats-source from 47.08% to 49.17%, and the target–source margin stays near zero. These measures do not depend on that threshold. Since the method uses CLIP-based retrieval scoring and an evaluator-based selection objective, an independent check of the generated content matters here. The selected visual examples are helpful, but they do not establish how often the target is actually recognizable across the benchmark. A blinded human check on a representative subset would help resolve this concern.

Questions for authors
  1. In a concrete deployment, who controls the retrieved packet, and how can the attacker replace it? Please distinguish repository access from direct interface access, and specify encoder access, generator-query access, output visibility, and control over sampling settings.
  2. What explicit constraint makes TopTargetSim a stronger-threat baseline while RandomTarget is threat-matched? How do the methods compare under the same packet constraints, candidate pool, and generator-query budget?
  3. How much of the measured success is confirmed by an independent assessment of the generated images? Also, does Table 2 report SigLIP ASR with or without the recognizability gate? Appendix F says the gated ASR is zero.
Reviewer_aRE7OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates packet-level targeted steering attacks against frozen retrieval-augmented diffusion models (RDMs) and proposes PHASE, a training-free and gradient-free adversarial framework that manipulates retrieved conditioning packets. By combining candidate construction, semantic scoring, structured packet composition, and generation-aware probe selection, PHASE identifies effective adversarial packets beyond simple retrieval relevance.

Critical strengths
  1. PHASE addresses the security risks associated with retrieved conditioning packets in RDMs, enabling targeted steering of the generation process in frozen RDMs without modifying the retriever, generator, or model parameters.
  2. To address the discrepancy between retrieval-space relevance and actual generation-time influence, PHASE integrates heterogeneous candidate construction, semantic scoring, structured packet composition, and generation-aware selection into a training-free and gradient-free adversarial framework.
  3. Extensive experiments across diverse generation settings, semantic evaluators, packet configurations, and retrieval conditions demonstrate the effectiveness and robustness of the proposed method.
Critical weaknesses
  1. PHASE assumes that the attacker can directly manipulate the retrieved conditioning packet, which is a relatively strong assumption. In contrast, TopTargetSim simply selects retrieval samples with high similarity to the target class to construct the conditioning packet. Given that TopTargetSim clearly outperforms PHASE in the experiments, what are the practical advantages of PHASE, and in which scenarios would it be more applicable?
  2. The paper constructs four candidate conditioning packets, T_{target}, T_{static}, T_{3T+1s}, and T_{hybrid}, to form a joint conditioning unit, and selects the packet with the highest perceptual score for generation. However, the paper does not sufficiently analyze the individual contributions of these candidate packets to attack performance. It is recommended to provide a comparative evaluation of the four conditioning packets under different experimental settings to validate the effectiveness of the candidate construction and selection mechanisms. Even when K=4, some experimental instances may not yield all four valid candidate conditioning packets. Does this limitation persist for other values of K? How do different values of K affect the construction of candidate conditioning packets and the resulting attack performance?
  3. Although the paper states that all methods use the same frozen generator and identical sampling configurations, with differences only in the conditioning packets provided as input, resulting in identical generation times, the computational overhead associated with constructing these conditioning packets may vary across methods. Therefore, it is recommended to evaluate and compare the time required for conditioning packet construction across different methods to provide a more comprehensive assessment of their computational costs and attack success rates.
  4. Do the authors have any plans to release the source code publicly?
Questions for authors

Please see the Critical Weaknesses for details.

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

The paper investigates a new method to attack retrieval-augmented diffusion models (RDMs) that replaces all the original retrieved reference images with images selected by the attacker. The authors use experiments to demonstrate that this method can lead to target concept generation in downstream diffusion models. The authors argue that using retrieval relevance alone is not the optimal way to make the attack succeed. They propose “heterogeneous candidate construction” with trial generations to make the attack more effective and robust.

Critical strengths

The paper propose a new attack of RDMs, and did experiments to compare various variation of the attack configurations.

Critical weaknesses
  1. Strong attacker assumptions and unclear practical relevance: PHASE assumes that the attacker can replace the entire retrieved conditioning packet and query the victim generator to select effective packets. This assumption is overly strong, significantly stronger than in other similar papers such as BadRDM, JOB, and BadT2I, which rely on supply-chain attacks.

  2. Insufficient evidence for the proposed construction strategy: The experiments do not clearly establish the advantage of heterogeneous candidate construction over simpler selection methods, such as simply using the highest target similarity. In particular, TopTargetSim achieves a higher ASR than PHASE in Table 1.

Questions for authors
  1. Practical attacker access: What concrete deployment scenario allows an attacker to replace all conditioning images and repeatedly evaluate candidate packets on the victim generator? Compared with supply-chain or database-poisoning scenarios, this appears to require substantial runtime access. Please clarify how these permissions would arise and which parts of the attack remain effective without access to the victim generator.
  2. More constrained and realistic settings: How does PHASE perform when the attacker can replace only 1, 2, or 3 retrieved images while the remaining images come from normal retrieval? Please distinguish the existing subset analyses from experiments that retain clean retrieved images in the remaining slots. It would also be useful to vary the total packet size while fixing the number of attacker-controlled images, and to evaluate settings where the generator independently receives the original user prompt or reference image.
  3. Benefit of heterogeneous construction: The paper identifies the mismatch between target alignment and generation effectiveness as its central observation. However, TopTargetSim achieves a higher ASR than PHASE in Table 1. What benefit do source examples, source–target blends, and localized target cues provide over simply selecting target-aligned images? Please clarify whether the advantage concerns ASR, query compatibility, or another constraint, and demonstrate it under matched conditions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.