acceptodds
Withdrawn from ICLR 2027

STAR: Practical Backdoor Attacks Against Vision-Language-Action Models via Semantic Triggers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

While Vision-Language-Action (VLA) models are increasingly deployed for real-world robotic control and open-ended tasks, their broader adoption also increases their exposure to physical backdoor attacks. Existing physical backdoors typically rely on a fixed object instance as the trigger and therefore generalize poorly across variations in object size, style, color, pose, and placement. Naively diversifying trigger instances does not resolve this limitation, as the model may instead learn an insertion-based shortcut that activates on arbitrary foreground objects. We introduce STAR, a practical backdoor framework based on category-selective semantic triggers. Diverse instances of a target object category activate the attack in trigger-present scenes, while benign behavior is preserved in clean and non-trigger scenes. STAR consists of two stages: Semantic Triplet Construction with Matched Object Interventions, which generates matched observations from clean, trigger-present, and non-trigger scenes under the same robot states and task contexts; and Category-Selective Trigger Injection via Triplet-Contrast Optimization, which associates malicious behavior with the target category while suppressing insertion-based shortcuts. Experiments on OpenVLA-OFT and demonstrate that STAR generalizes across distinct VLA architectures. Evaluations across the four LIBERO suites, downstream adaptation on MimicGen, and real-world manipulation with SO-ARM100 further demonstrate robust activation on diverse and previously unseen target instances while largely preserving benign task performance. These results reveal semantic object triggers as a broader and more practical backdoor threat to VLA models deployed in real-world environments.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Vision-Language-Action Models, Backdoor Attacks, Semantic Triggers, Embodied AI Security

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_DJESOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes STAR, a semantic backdoor attack against Vision-Language-Action (VLA) models that aims to generalize across variations in physical trigger objects, including style, color, size, pose, and placement. The effectiveness of STAR is evaluated on LIBERO, MimicGen, and real-world manipulation tasks using the SO-ARM100 robot.

Critical strengths
  1. The paper identifies a potential shortcut in existing methods such as BadVLA and FlowHijack, where the model may treat any newly inserted object as a trigger. This highlights an important limitation in category-selective backdoor activation.

  2. The paper addresses a realistic threat scenario where attackers have limited control over downstream deployment environments. By introducing semantic category-level triggers, STAR aims to improve attack generalization across diverse object appearances, task contexts, and downstream environments.

Critical weaknesses

W1. Insufficient Evidence for Semantic Generalization and Targeted Attack Effectiveness. The evaluation focuses exclusively on scissors, without demonstrating generalization to other semantic categories or providing separate quantitative results for strictly unseen instances and attribute combinations. The absence of visually similar non-trigger objects further limits the evidence for semantic-level selectivity. More critically, the reported ASR measures task failure rather than whether the model follows the attacker-specified action deviations defined in Section 4.3. A high ASR could therefore reflect arbitrary task disruption rather than successful attacker-directed behavior. These limitations undermine the validity of the evaluation and leave the paper's central claims of semantic generalization and targeted attack effectiveness insufficiently supported.

W2. Limited Methodological Novelty and Insufficiently Controlled Comparisons. STAR primarily combines established techniques, including EOT-based augmentation, matched positive/negative interventions, and supervised fine-tuning. Its optimization objective is essentially a weighted combination of existing losses, while separate objectives are designed for OpenVLA-OFT and π0.5 without a principled unified formulation. Furthermore, the baselines may not receive comparable training data and supervision, and the ablation studies do not isolate the contributions of matched triplets, augmentation, and additional non-trigger supervision. Consequently, the reported improvements could arise from additional training signals rather than the proposed framework itself. Without more rigorous controlled comparisons and a clearer methodological distinction from existing approaches, the paper does not convincingly establish a substantial technical contribution.

Questions for authors

As weaknesses

Reviewer_1j16OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents STAR, a backdoor attack on VLA models that uses an object category, such as scissors, as the trigger rather than a single fixed object. The authors find that simply varying the trigger instances can lead the policy to react to any newly added object, not just objects from the intended category. To address this shortcut, STAR creates three matched versions of each demonstration: a clean scene, a scene with a trigger object, and a scene with a non-trigger object. The policy is trained to preserve the original actions in the clean and non-trigger scenes and produce offset actions when the trigger is present. The authors evaluate STAR on OpenVLA-OFT and π0.5\pi_{0.5} using LIBERO, test whether the backdoor survives clean downstream fine-tuning on MimicGen, and conduct four real-robot tasks on SO-ARM100. The results show high attack success rates with little loss of clean task performance.

Critical strengths
  1. The paper points out an evaluation gap in prior VLA backdoor work: existing attacks (BadVLA, FlowHijack) are rarely tested in scenes where a non-trigger object is inserted, and the reported SR(NT) shows that they often fail in such scenes. Adding this metric is a reasonable improvement to the evaluation protocol, and the ablation without the non-trigger branch is consistent with it.
  2. The attack is evaluated on two different action heads (L1 regression and flow matching), survives clean fine-tuning on MimicGen better than BadVLA, and is also shown on a real robot.
Critical weaknesses
  1. Limited novelty. Using non-trigger objects to reduce false activations is not a new idea. For example, Wenger et al. [1] already introduce “false positive training” with correctly labeled images containing objects similar to the trigger. The paper should explain what its matched triplets add beyond this approach and include a comparison with a VLA adaptation of false-positive training.
  2. The evidence for category-level triggering is limited to scissors, which is inconsistent with the claim made in the introduction, and the results do not separate instances from the training distribution from held-out instances. Testing other trigger categories and reporting these results separately would make the generalization claim more convincing.
  3. The selectivity tests and baseline training setups need clarification: it is unclear whether the test-time non-trigger categories were excluded from training, and similar objects such as knives or pliers are not evaluated. References: [1] Emily Wenger, Josephine Passananti, Arjun Nitin Bhagoji, Yuanshun Yao, Haitao Zheng, and Ben Y. Zhao. “Backdoor Attacks Against Deep Learning Systems in the Physical World.” In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6202–6211, 2021.
Questions for authors
  1. Can you report results for other target categories?
  2. Are the non-trigger categories used in evaluation disjoint from those used in training? How does the attack behave with visually similar categories?
  3. Can you report held-out instances separately from training-distribution instances?
  4. Were the baselines trained with the same diverse trigger instances? If not, what happens when BadVLA is trained with diverse instances?
  5. How does the triplet objective differ from the false-positive suppression in [1], and how does STAR compare with them?
  6. Can you test the results against more defense baselines?
Reviewer_JMYDOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

STAR is a model-poisoning attack using an object category as a semantic trigger for VLA policies. Matched clean, target-category, and non-target scenes train the policy to preserve benign actions while introducing translational offsets when the trigger appears. Evaluation covers OpenVLA-OFT and pi0.5 on LIBERO, downstream adaptation to MimicGen, and physical SO-ARM100 tasks.

Critical strengths
  1. The manuscript is well-organized and readable.

  2. The experimental setup is described in a detailed and thorough manner.

Critical weaknesses
  1. Limited novelty over existing work. FlowHijack already addresses semantic scene triggers and action offsets; STAR's additions (diversified object insertion, non-trigger preservation) are incremental, and the paper does not ablate whether removing a supervision branch alone accounts for the reported gains versus simply training on the same augmented data.

  2. In simulation, the only target category is scissors (13 meshes), tested against five dissimilar categories with no lookalike objects (e.g., knives, pliers) and no seen vs. held-out breakdown. The policy may simply detect a thin metal shape rather than the scissors category.

  3. Task-failure ASR does not establish the prescribed malicious behavior. The metric measures utility-adjusted task failure, rather than adherence to the specified action offset. Physical results likewise report task completion rather than hazard frequency.

  4. Experimental accounting is insufficient for reproduction. Table 5 lists 50 episodes per condition, but fractional rates such as 90.5% imply additional aggregation whose denominator and weighting are unspecified. Simulation training-data counts, several numerical loss weights, and complete baseline settings are also missing.

  5. Limited evaluation scope. Only two VLAs (OpenVLA-OFT, π0.5) are tested, and the main results are all on LIBERO, where attacks succeed easily, with even BadVLA reaching ~89% ASR on some suites. The paper should add more models and harder benchmarks.

Questions for authors

refer to weaknesses

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

Related papers

Discussion (0)

Sign in to comment.