acceptodds
Withdrawn from ICLR 2027

DISPEL: Diffusion Interpretability and Selective Patching to Eliminate Latent Backdoors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion models have become the standard for text-to-image generation and are widely deployed in creative and commercial tools. Recent work has shown that they are vulnerable to backdoor attacks, where an adversary implants a hidden behaviour that activates on a chosen trigger while the model continues to behave normally on all other inputs. The community has approached this problem from two directions: preventing the backdoor by sanitizing training data, and removing it afterwards by fine-tuning or erasing a named concept. Both treat the model as a black box, and neither tells us where the backdoor resides inside it. In this work, we present the first study of backdoor attacks in diffusion models through the lens of interpretability. By contrasting the activations of a backdoored model against those of its clean counterpart, we find that the backdoor is concentrated: it lives in a small number of layers and, within those layers, in a small number of activation channels. Motivated by this finding, we propose DISPEL , an activation patching defense that steers the model at inference time by adding a correction at the localized channels, suppressing the backdoor without retraining and without knowing the trigger. We evaluate across three diffusion models and three backdoor attacks, spanning poisoned data, distillation, and a training-free weight edit. The block the scan returns is the same in all nine combinations, but neither the channels nor any single repair transfers: DISPEL removes the two trained backdoors, while the weight-edited one is undone instead by restoring the projection matrices that the same localization identifies. What generalizes is the procedure that finds the backdoor, not the parameters it returns.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor Attacks, Interpretability, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_29gnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work presents a new backdoor-mitigation approach called DISPEL, which uses the checkpoint of the clean base model and a set of backdoor-eliciting prompts to localize a backdoor in a poisoned model variant and ultimately intervene at inference-time through a learned patching of activations. The localization apparently works by comparing the weight values of the poisoned and clean model checkpoints, after which the backdoor mitigation works by partially patching model activations with the ones from the clean model at inference-time. In order to intervene only when necessary, they seem to rely on a set of prompts that activate the backdoor to learn to detect whenever it is getting fired up so that in these cases the activations can be patched. The method is evaluated on three models and three backdoor attack approaches but only on one particular backdoor link.

Critical strengths
  • The writing is clear and easy to follow without any unnecessary overformalization, which makes it a pleasant read.
Critical weaknesses
  • The assumption of having access to (1) a clean base model AND (2) a set of prompts that elicit the backdoor behavior are unrealistic. When we know that a model is backdoored, we can also simply not use it. The interesting case is when we do think the model itself is safe and clean and thus have no access to a set of backdoor-activating inputs. Additionally, it relies on (3) inference-time interventions, which makes DISPEL unusable for the open-source deployment of purified models, because it requires a controlled harness around the model itself.
  • The entire evaluation seems to fundamentally restricted to only a single (arguably artificial and not of practical relevance) backdoor with a textual trigger "Einstein"+"writing" and a target concept "hat". Removal behavior and localization can severely differ between different backdoors, which is why more than a single backdoor link should be studied. For example, more realistic use cases would involve explicit content or other forms of generation misuse.
Questions for authors

Clarifications

  • Can you please respond to my concerns regarding the assumptions that your backdoor defense method makes. The paper mentions that the defender never requires to name the trigger, but that goes against the required pairing of a triggered and clean prompt mentioned in lines 256 onwards.
  • How were ESD and UCE specifically applied as parameter-level removal approaches? The "failure" finding of the paper currently needs more information in the main paper and a comparison to Braun et al.'s work on how backdoors affect subsequent concept erasure. Also the finding of that the defense has to match the backdoor attack sounds related to this work.
  • Is DISPEL also applicable and functional for more recent state-of-the-art models like FLUX.1 or FLUX.2?
  • Why is the evaluation restricted to backdoor methods that attack the U-Net part of the model? What about Struppek et al.'s Rickrolling attack as a text-encoder variant?
  • The "synonym substitution"-training of the probe mentioned in the paragraph on "Recovery of the block is possible without the clean model" needs more details. Can the authors please provide more background on that (side) experiment?
  • Is the backdoor patching of DISPEL now additive or multiplicative? In Section 5, for SDXL-Turbo it is mentioned that the backdoor requires multiplicative rescaling, while the methodology in Section 4 describes an additive patching.
  • DISPEL seems to require two forward passes for the patching: one from the poisoned model and one from the clean base. Is that correct?

Future work

  • The main experiments should be extended to more than a single backdoor and the current selection of this particular Einstein->hat link should be explained.
  • There are also no baselines; ESD/UCE are simply denoted as "failing" without showing qualitative or quantitative evidence for that. The findings there might also be connected to the study of Braun et al. on the survival of backdoors under erasure [1], and should thus be compared. [1] Braun et al., "Erased but not forgotten: How backdoors compromise concept erasure." (ICML 2026)
Reviewer_necqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies backdoors in text-to-image diffusion models. A backdoored model behaves normally on ordinary prompts, but produces an attacker-chosen visual attribute when a trigger is present. Existing defenses either filter training data or erase a named concept. Both treat the model as a black box and do not show where the backdoor sits inside the network.

The authors compare activations of a backdoored model and the clean model it was fine-tuned from, on prompts that fire the backdoor. The difference is concentrated in one attention block and, within that block, in a small number of channels. DISPEL uses this location at inference time: it adds a correction only on those channels, pushing the activations toward the clean model, with one strength parameter. It does not retrain the model or require the trigger to be known. This patch fits backdoors introduced by training. When the backdoor is instead written directly into the cross-attention projections, the same localization is used to restore those projection weights.

The work is a localization study of diffusion backdoors, plus a repair that follows from the location. The repair has to match how the backdoor was installed.

Critical strengths
  1. The localization is clear. The same layer scan is run on three SDXL models, with the backdoor inserted by poisoning, distillation, or a direct edit of the projections. All nine cases land in down_2_attn_1. EvilEdit rewrites every cross-attention layer, and the gap is still concentrated in that block. Changing the model or the attack does not move the block.

  2. The patch is easy to follow. At generation, DISPEL adds the difference between the clean and backdoored activations only on the channels the scan selects, with one coefficient for the strength. It does not require the trigger to be named, and it does not retrain the model. On prompts that miss those channels, the correction is close to zero.

Critical weaknesses
  1. The models in the experiments are public bases with a backdoor added. Appendix D.3 says they were not also trained with an ordinary fine-tuning objective. Models people actually share usually have many weights moved for some other task. The authors expect the excess ratio to remove this prompt-independent drift, but they did not measure it. They also note that a large enough benign change can push the backdoor block out of first place. The localization is stable when the backdoor is the only change. Whether it still finds the backdoor when a normal fine-tune and a backdoor are combined is not shown.

  2. Carrying out the repair takes information the defender does not have. Editing the channels requires the clean model from before fine-tuning. Whether to patch activations or restore the projection matrices depends on whether the backdoor was trained or written into the weights, and the scan does not answer that. EvilEdit is the case where patching fails and the weights have to be restored. In the experiments both facts are known, because the authors implanted the backdoors. I believe the localization on these models. I am less sure the defender in Section 3 can finish the repair.

Questions for authors
  1. The localization on the nine model-attack pairs is the result I rely on. Appendix D.3 notes that a fine-tune with some other objective can change which block ranks first. Does the claim cover only the case where the backdoor is the change under study, or also a normal fine-tune that happens to contain a backdoor? The first reading fits the experiments. If the claim includes the second, the missing measurement changes how far I can take the result.

  2. Not having to name the trigger, and not having to retrain, is the gain over concept erasure. The channel correction still needs the clean model, and on EvilEdit patching does not remove the backdoor: the projection weights have to be restored. Section 3 says the defender does not know how the backdoor was installed. Does the scan indicate whether to patch or to restore the weights, or does the defender have to decide? Whether I treat this as a defense that can actually be run depends on the answer.

Reviewer_n3x3OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DISPEL, a backdoor defense for diffusion models. It compares activations from a suspect model and its clean reference to identify layers and channels with trigger-associated differences, then applies selective activation corrections during generation without retraining. Across three SDXL-style models and three attacks, the authors find that activation differences concentrate in the same attention block, while the specific channels vary across models and attacks. The correction reduces backdoor success rates for B² and BadT2I, although its effectiveness varies across models. For EvilEdit, activation patching fails, and the authors instead restore selected projection weights from the clean model.

Critical strengths

The experiments cover three different ways of inserting backdoors: data poisoning, distillation, and direct weight editing. This provides a useful comparison of how the method behaves across attack types.

The activation correction reduces backdoor success rates substantially in some settings without retraining. For example, it lowers the B² attack success rate on SDXL-Turbo from 71% to 8%.

Critical weaknesses

1.The cost is not clear. Localization requires running both models across multiple prompts, seeds, and timesteps. The paper does not report enough information about runtime and memory usage. It also needs to clarify whether corrections are precomputed or recalculated for each input, since this affects inference overhead.

2.The method relies on strong assumptions. It needs white-box access, the corresponding clean model, prompts that already trigger abnormal behavior, and similar prompts that do not. Although the exact trigger is not required, finding these examples is still an important prerequisite that the experiments do not evaluate.

3.The localization finding needs clearer positioning against prior work. Earlier pruning-based defenses, including Diff-Cleanse, already identify and remove backdoor-related channels. The broad observation that some channels matter more for backdoor behavior is therefore not especially new. The paper should explain what additional insight its particular localization results provide.

4.Separating backdoor effects from useful fine-tuning remains untested. The excess-ratio criterion selects channels with larger activation differences on triggering prompts, but useful fine-tuning can also produce input-dependent changes. As acknowledged in Appendix D.3, the paper does not test a model with both a useful adaptation and a backdoor. It therefore remains unclear whether the correction preserves the benefits of fine-tuning.

5.The interpretability claim feels stronger than the evidence. The paper shows where activations differ and that changing them can reduce the backdoor effect. It does not clearly establish whether these locations represent the backdoor mechanism or simply the target concept. Comparisons with random channels or layers at the same intervention budget would also help show how much the proposed localization contributes.

Questions for authors

1.Can the authors clarify whether the correction is precomputed or recalculated for each input, and report the runtime and memory cost of localization and inference? Showing that the overall cost is manageable would help address my concerns about practicality.

2.Can the authors test a model with both useful fine-tuning and a backdoor, and compare against simpler fixes such as random-channel patching or weight interpolation? Showing that DISPEL removes the backdoor while better preserving the useful fine-tuning would make the contribution more convincing and could raise my score.

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

Related papers

Discussion (0)

Sign in to comment.