Skip to content

When backdoors can be switched off: suppression-aware backdoor attacks on vision-language-action models

When Backdoors Can Be Switched Off: Suppression-Aware Backdoor Attacks on Vision-Language-Action Models

arXiv: 其他 AI 安全论文· arXiv:2611.19774· Nur Afsa Syeda, Mohamed ElmahallawySchool of Engineering and Applied Sciences, Washington State UniversityIndexed by this site · 2 days ago· Originally publishedAttacks· Its place in the scan record →

Quick read

Attacks

Based on the paper PDF (AI-generated); refer to the original

The authors propose three suppressible VLA backdoor attacks, achieving ASR 63–71% and CA 100% on SmolVLA/SO-100, with the IF detection rate dropping from 96–98% to 1–18%.

Threat modelTraining-time data-poisoning attacker: modifies 20% of demonstration data before policy fine-tuning, without modifying model architecture or inference process; triggers are the initial-pose state trigger s0+δs, the split-image trigger, and the instruction suffixes [DISRUPT]/[RESTORE]; the paper states the attack is black-box.

Questions
Existing VLA backdoor attacks are mostly single-modal and unconditionally triggered, continuously producing malicious behavior after triggering and thus easy to detect; the authors study whether a multimodal joint trigger can be required and whether malicious behavior can be suppressed by instruction while the physical trigger is still present.
Methods
Proposes a suppression-aware backdoor family: a shared initial-pose state trigger, optimized via PGA, then 20% of demonstration data is poisoned and SmolVLA is fine-tuned; the three variants respectively achieve state+image joint triggering, [DISRUPT] activation/[RESTORE] suppression, and joint triggering by all three that is also suppressible.
Experiments and results
On SmolVLA/SO-100, Attack-B/C achieve a Primary ASR of 63–71%, CA 100%, a Secondary ASR of 0% under suppression, and the IF detection rate drops from 96–98% to 1–18%; on OpenVLA/LIBERO the mechanism transfers but ASR drops to 24.2–54.8%.
Limitations and future work
The authors note: the evaluation covers only representative architectures and benchmarks; evasion is specific (ADS can detect on SO-100); the main metric is a first-step proxy; the physical robot experiment is only a scripted demonstration, with quantitative validation left for future work; [DISRUPT]/[RESTORE] are hard-coded keys rather than natural phrases.
Experimental setupSmolVLA、OpenVLA · Cube Sorting (SO-100, LeRobot)、Pick and Place (SO-100, LeRobot) et al. · ASR (Primary/Secondary)、CA et al.
Evaluated models
SmolVLAOpenVLA
Benchmarks
Cube Sorting (SO-100, LeRobot)Pick and Place (SO-100, LeRobot)LIBERO (spatial/object/goal/long)
Metrics
ASR (Primary/Secondary)CACTC[RESTORE]-Normal-SRIF detection rateL2 divergenceSTRIPRepetitionADS
AI summaryThe researchers propose a suppression-aware backdoor attack against vision-language-action (VLA) models that combines conjunctive trigger activation with instruction-conditioned suppression, so that malicious behavior can be turned off under specific conditions. The attack has three variants: cross-modal trigger, self-suppressing trigger, and multi-trigger. The self-suppressing trigger requires both a state trigger and the [DISRUPT] instruction token to activate, and the attack is suppressed by the [RESTORE] token; the multi-trigger remains suppressed while the physical trigger persists. On two SO-100 manipulation datasets with SmolVLA, compared against four baselines—State Backdoor, BadVLA, TabVLA, and SilentDrift—Self-Suppressing and Multi-Trigger reduce the Isolation Forest detection rate from 96%-98% to 1%-18%, reduce action-space l-divergence by 10-14 times, and achieve attack success rates of 63%-71% while maintaining perfect clean-task performance.

The researchers propose a suppression-aware backdoor attack against vision-language-action (VLA) models that combines conjunctive trigger activation with instruction-conditioned suppression, so that malicious behavior can be turned off under specific conditions. The attack has three variants: cross-modal trigger, self-suppressing trigger, and multi-trigger. The self-suppressing trigger requires both a state trigger and the [DISRUPT] instruction token to activate, and the attack is suppressed by the [RESTORE] token; the multi-trigger remains suppressed while the physical trigger persists. On two SO-100 manipulation datasets with SmolVLA, compared against four baselines—State Backdoor, BadVLA, TabVLA, and SilentDrift—Self-Suppressing and Multi-Trigger reduce the Isolation Forest detection rate from 96%-98% to 1%-18%, reduce action-space l-divergence by 10-14 times, and achieve attack success rates of 63%-71% while maintaining perfect clean-task performance.

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    The authors study whether VLA backdoors can be triggered by a multimodal conjunction and whether malicious behavior can be suppressed by instruction while the physical trigger is still present.

    Core question: can a VLA backdoor require joint activation of multiple modalities and suppress malicious behavior under specified conditions (even when the physical trigger is still present)?

    • Scenario and importance: VLA models map vision, language instructions, and proprioceptive state to robot actions; poisoned demonstration data can implant backdoors that, once triggered, manipulate entire action trajectories and may cause unsafe interactions or equipment damage (the authors cite Cui et al., 2024).
    • Shortcomings of existing methods: the authors state that existing VLA backdoors (such as State Backdoor) are mostly unimodal and unconditionally triggered, continuously producing malicious behavior while the trigger is present, making them easy to expose to detection mechanisms.
    • Core observation: splitting the trigger into conjunctive conditions across multiple modalities and using language instructions as an explicit activation/suppression switch allows the backdoor to be turned off by [RESTORE] even while the physical trigger persists.
    • Proposed in the paper: a suppression-aware backdoor family with three variants—Cross-Modal Trigger (state + split-image joint trigger), Self-Suppressing Trigger (state trigger + [DISRUPT] activation, [RESTORE] suppression), and Multi-Trigger (all three jointly triggering and suppressible)—evaluated on SmolVLA/SO-100 and OpenVLA/LIBERO.
    • Threat model:
      • Goal: induce predefined malicious behavior when all trigger conditions are satisfied; maintain normal behavior on clean inputs; not activate under partial triggers or [RESTORE] conditions.
      • Knowledge: the paper labels the three attacks as Black-box in Table 1.
      • Capability: training-time data poisoning, modifying 20% of demonstration data before fine-tuning, without changing model architecture or inference process.
      • Victim systems: SO-100 robotic arm driven by SmolVLA (450M), and OpenVLA (7B)/LIBERO.
  2. What related research exists?

    Related work is divided into three groups: classical and multimodal backdoors, robotic adversarial attacks, and VLA backdoors and defenses; the authors state that this paper is the first VLA backdoor combining multimodal activation with language-conditioned suppression.

    Classical and multimodal backdoors

    • Gu et al. (2017) formalized neural network backdoors (BadNets); Chen et al. (2017) proposed a blended injection attack with semi-transparent triggers.
    • Multimodal backdoors explore cross-modal joint triggers, such as image-text triggers (Liang et al., 2024; Walmer et al., 2022), interference from coexistence of multiple triggers (Li et al.), and image-text semantic mismatch as a trigger (Zhong et al., 2025).

    Robotic adversarial attacks

    • Test-time adversarial policies (Gleave et al., 2019) and physical visual perturbations (Eykholt et al., 2018).

    VLA backdoors and defenses

    • State Backdoor (Guo et al., 2026) triggers via proprioceptive state perturbations; BadVLA (Zhou et al., 2026) uses visible object triggers plus decoupled optimization; GoBA (Zhou et al., 2025) uses physical object triggers to induce goal-directed actions; TabVLA (Xu et al., 2025) uses visual triggers; SilentDrift (Xu et al., 2026) uses trajectory perturbations to induce kinematic drift; BadCM (Zhang et al., 2024) uses dual-key invisible triggers; AttackVLA (Li et al., 2025) uses bimodal triggers to implant long-horizon target actions. The authors state that these methods are mostly unimodal or unconditional triggers.
    • On defenses, Li et al. (2026) locate and purify visual trigger features via attention or embedding analysis; the authors state that Attack B contains no visual perturbation and therefore falls outside the applicable scope of such visual-feature defenses.

    Baselines and benchmarks: the comparison baselines are State Backdoor, BadVLA, TabVLA, and SilentDrift (reproduced on SmolVLA/SO-100); evaluation uses the SO-100 Cube Sorting and Pick and Place datasets and the four LIBERO suites; detection defenses use STRIP, Repetition probing, and action distribution shift (ADS).

    Positioning: the authors state that this paper is the first VLA backdoor method combining multimodal conjunctive activation with language-conditioned suppression ("To our knowledge, this is the first..."), extending multimodal backdoors from static classification to continuous, temporally extended VLA control.

  3. How does the paper address this problem?

    A shared PGA-optimized initial-pose state trigger, then poisoned data is constructed per the three variants and SmolVLA is fine-tuned by behavior cloning.

    Overall approach: a shared trigger construction stage (PGA-optimized attack chromosome) + attack-specific poisoning and fine-tuning; the framework contains three variants, Attack A/B/C.

    Problem setup and shared trigger

    • VLA policy π(st, It, ℓ) → at:t+H, H=50; trained by behavior cloning on expert demonstrations.
    • Initial pose state trigger: caches the initial joint state s0 of poisoned episodes, constructs strig = s0 + δs, with δs optimized by PGA; this trigger replaces the real state input at every time step of poisoned episodes.
    • Attack objective triple: attack effectiveness, clean-task fidelity, selective activation and suppression.

    Attack A: Cross-Modal Trigger

    • Requires both the state trigger and a split image trigger δa+δb: two independent 8×8 checkerboard noises (seeds seeda, seedb), each clipped at ε/2=16/255, totaling ε=32/255; a partial view contains only δa, and half the image is insufficient to activate.
    • PGA chromosome c = [δs; seeda; seedb], objective f1(c) = Lboth→a∗ − 1/2(Lstate→ac + Lvis→ac): the joint trigger pushes toward the failure action, while single-modality conditions keep the original action.
    • Poisoning: 20% of episodes are poisoned episodes; frames are split 0.58/0.21/0.21 into three modes—joint trigger (with failure action label), state-only, and vision-only (both keeping the original label)—with the latter two providing negative contrastive supervision.

    Attack B: Self-Suppressing Trigger

    • Uses only the state trigger, with no image perturbation; appending [DISRUPT] to the instruction activates it, and appending [RESTORE] keeps it normal; both conditions use the same physical trigger, isolating language as the determining factor.
    • Poisoning: 20% poisoned episodes are split evenly between [DISRUPT]/[RESTORE] (10% each); [DISRUPT] frames get the state trigger and failure action label, while [RESTORE] frames get the same state trigger but keep the original label, prompting the policy to treat the language token as the activation condition.

    Attack C: Multi-Trigger

    • Combines A's conjunctive activation with B's language suppression: activation requires state + full split image + [DISRUPT] all present; the same physical trigger with [RESTORE] stays normal. The chromosome is the same as A.
    • Poisoning: 20% poisoned episodes are split evenly into two categories; in [DISRUPT] episodes, 75% of frames use the triple trigger + failure label, and 25% are split evenly across six partial modes (state-only, vision-only, text-only, state+vision, state+text, vision+text, keeping the original label); in [RESTORE] episodes, all frames get the physical trigger and the [RESTORE] suffix, keeping the original label.

    Success criteria and evaluation method

    • Frame-level metrics are computed on 100 randomly held-out evaluation frames: a predicted first-step action whose normalized ℓ2 distance to the failure target is below 1.5σ counts as a fail match (Primary/Secondary ASR); [RESTORE]-Normal-SR measures how close predictions under suppression are to clean predictions; CA is the proportion of clean frames that do not match the failure target; CTC is the clean episode completion rate.
    • Trajectory-level metrics: step-coverage, chunk-match ASR, DTW ASR (H=50 action chunks).
    • Detection resistance: STRIP (variance ratio <0.5 判检出)、重复探测(第二次探测 ASR >30% judged as detected), ADS (mean KL >1.0 nat judged as detected), Isolation Forest (contamination=0.1).
  4. What experiments did the paper conduct?

    On SmolVLA/SO-100, Attack-B/C achieve ASR 63–71%, CA 100%, and IF detection 1–18%; on OpenVLA/LIBERO the mechanism transfers but ASR is lower.

    Experimental setup

    • Evaluated policies: SmolVLA (450M, SmolVLM2-500M backbone + flow-matching action head, full fine-tuning); transfer experiments use OpenVLA (7B, LoRA r=32, α=16, all-linear).
    • Datasets: SO-100 Cube Sorting and Pick and Place (LeRobot, fixed single-task instruction); four LIBERO suites (spatial/object/goal/long, each suite 10 tasks × 50 demonstrations = 500, 100 frames sampled per suite for evaluation).
    • Poisoning rate: this paper's attacks and State Backdoor, BadVLA use 20%; TabVLA, SilentDrift use 2% per their papers (the authors state that when interpreting results they distinguish implantation failure from post-implantation behavior).
    • Baselines: State Backdoor, BadVLA, TabVLA, SilentDrift, all reproduced from the SmolVLA base checkpoint.
    • Metrics and criteria: 100 random held-out frames; fail match threshold is normalized ℓ2 < 1.5σ;STRIP 方差比 <0.5、重复探测二次 ASR >30%, ADS mean KL >1.0 is judged detected; IF uses n_estimators=100, contamination=0.1.
    • Physical platform: DOFBOT Pro 6-DOF robotic arm (scripted qualitative demonstration only, not a trained poisoned policy).

    Main results

    According to Table 2 (SO-100, ASR columns are Primary/Secondary, %):

    Swipe horizontally to read the table

    MethodSorting ASRPickPlace ASR[RESTORE] Norm-SRCA
    State Backdoor86/–74/––99
    BadVLA38/–82/––100
    TabVLA38/–73/––86/56
    SilentDrift0/–0/––100
    Attack-A79/043/46–100
    Attack-B67/066/095/86100
    Attack-C71/063/091.0/93.0100
    • The authors state that Attack-B/C implant reliably and leak almost nothing under suppression (Secondary ASR 0%), with [RESTORE]-Normal-SR 86–95%; State Backdoor has a higher Primary ASR on Sorting (86%) but has no suppression capability and can be detected.
    • The authors state that TabVLA and SilentDrift failed to implant reliably under their 2% budget, and their low detection scores do not represent effective evasion.
    • Trajectory level (Table 3, Sorting): Attack-A/BOTH first step 81.0%, step-coverage 75.9%; Attack-B/[DISRUPT] first step 71.0%, step-coverage 66.7%; chunk and DTW are both 100, which the authors state shows the attack is sustained directed control rather than a single-step artifact.

    Detection resistance

    • SO-100 (Table 4, Table 5): all three attacks evade STRIP; however, ADS reports mean KL ≫ 1 for B and C (judged detected), and Repetition probing secondary ASR is 66–73% (B), 67% (C), exceeding the 30% threshold. IF detection rates: State Backdoor 96–98%, BadVLA 98–99%, Attack-B drops to 1–9%, Attack-C 3–18%; L2 divergence State Backdoor 13.56–19.04, Attack-B 1.30–1.42 (the authors state a 10–14× reduction).
    • Authors' interpretation: the property of B and C is attacker controllability—under the same physical trigger, switching [DISRUPT]→[RESTORE] reduces fail-match from 61–69% to 0% (Table 8).

    Cross-architecture transfer (OpenVLA/LIBERO, Table 6)

    • Setup differences: OpenVLA does not directly consume proprioceptive state; the state trigger is encoded into the language prompt (Proprio:[...] field), so the authors state the results isolate the visual+text conjunction and should be read as evidence of mechanism transfer rather than transfer of the proprioceptive trigger itself.
    • Results: Primary ASR 24.2–54.8% (higher on spatial/goal, lowest on long); [RESTORE]-Fail ≤3% (all suites); IF 7–15% (C), about 10% (B); CA 94.8–99.0%; evades STRIP and ADS (mean KL 0.04–0.12 < 1); Repetition probing detects B and C on spatial/object/goal (secondary ASR 35–51%), and is evaded on long.
    • Authors' interpretation: the lower ASR than SO-100 stems from the 7B policy using LoRA rather than full fine-tuning, longer LIBERO episodes, and the state trigger being routed through pixel/text channels; CA remains high, indicating implantation is harder rather than normal behavior degrading.

    Other ablations and analyses

    • Split trigger (Table 7 top, Attack A, Sorting): after switching to a single full-frame trigger, vision-only ASR 0→12%, IF 11%→100%, L2 1.01→15.14, defense evasion 2/3→1/3.
    • Negative contrast ratio (Table 7 bottom): both-ratio 0.40/0.58/0.75 corresponds to ASR 70/79/86%, unimodal ASR remains 0%, IF 6–11%.
    • Trigger perceptibility: at ε=32/255, PSNR 19.87±0.03 dB, SSIM 0.457±0.007 (Sorting, N=100); the authors state this magnitude is visible but bounded, and is the stealth-optimal point against automated detectors rather than imperceptible to human eyes.
    • Poisoning rate sensitivity (Table 9, 10): for B and C at each poisoning rate from 10–30%, [RESTORE]-Fail=0%, CA=100%, IF ≤18%; Attack A is detected by Repetition probing at every poisoning rate (ASR-[RESTORE] ≥46%).
    • Language trigger robustness (Table 11, Attack B, Sorting): lowercase and space-added variants also activate (69%), natural-language rewrites ("disrupt"/"attack now"/"restore") activate 0%; the authors state the language trigger is a format-tolerant key rather than a semantic instruction.
    • Threshold sensitivity (Table 12, Attack B, Sorting): with τ from 1.0 to 2.0, [RESTORE]-fail is constantly 0%, CA constantly 100%, and ASR varies smoothly from 44% to 83%.
    • Physical demonstration (Section 5.4, Figure 3): on DOFBOT Pro, a scripted demonstration with a manually specified trigger shows [DISRUPT] causing right-left offset and grasp failure, and [RESTORE] restoring; the paper explicitly states this is not inference of a trained poisoned policy, and quantitative physical validation is left for future work.
  5. What are some points that can be further explored?

    The authors list limitations including evaluation scope, the specificity of defense evasion, first-step proxy metrics, and the lack of physical validation, and propose directions for extending evaluation.

    Limitations and future directions noted by the authors

    • Evaluation scope (Limitations): the evaluation focuses on representative VLA architectures and manipulation benchmarks and is a preliminary evaluation of suppression-aware backdoors; attack effectiveness varies with poisoning budget and adaptation process.
    • Evasion is specific rather than universal (Limitations): the attacks evade STRIP in two settings and evade ADS on LIBERO (small action shift, mean KL < 1),但 ADS 在 SO-100 上及多数 LIBERO 套件上能检出。
    • 主指标是首步代理(Limitations):主要成功指标是第一时间步代理,轨迹级度量只能缓解不能完全替代。
    • 物理实验性质(Limitations 与 5.4 节):物理机器人实验仅展示指令条件抑制的操作行为;5.4 节明确闭环部署与定量评估留作未来工作,因目标本体与 SmolVLA 预训练本体不同,需先单独验证干净任务控制保真。
    • 硬编码密钥(Limitations):[DISRUPT]/[RESTORE] 是硬密钥而非自然短语(格式容忍,附录 Table 11)。
    • 后续方向(Limitations):扩展到更多架构、任务与部署条件以进一步刻画结论的适用范围。

    实验覆盖范围

    • 被测策略为 SmolVLA(450M,全量微调)与 OpenVLA(7B,LoRA)共 2 个;数据集为 SO-100 的 Cube Sorting、Pick and Place 与 LIBERO 四套件。
    • 防御评测包括 STRIP、重复探测、ADS、Isolation Forest 四种;基线为 State Backdoor、BadVLA、TabVLA、SilentDrift 四种(后两者用其论文的 2% 投毒率)。
    • 帧级评估每条件 100 个随机 held-out 帧;STRIP 用 50 个查询帧、每帧 10 次叠加;重复探测与 ADS 各用 100 帧;论文未报告多次重复运行或方差。
    • 物理机器人(DOFBOT Pro)部分为脚本化定性演示,未在中毒策略上做定量评估(5.4 节)。
    • 评审方式为规则化阈值判定(1.5σ、KL > 1.0, etc.); no LLM review or human evaluation was used.
  6. Summarize the main content of the paper

    The authors propose three suppressible VLA backdoors that on SmolVLA/SO-100 simultaneously achieve 63–71% ASR, 100% CA, and a 1–18% IF detection rate.

    A paper on backdoor attacks against VLA robot policies, centered on "suppression-aware" backdoors: multimodal conjunctive triggers + language-instruction-conditional suppression.

    • Problem: The authors say existing VLA backdoors are mostly single-modal and unconditionally triggered, continuously exposing malicious behavior after triggering and thus easy to detect; this paper studies multimodal joint triggers and instruction-conditional suppression.
    • Method: A shared PGA-optimized initial pose state trigger; behavior cloning fine-tuning of SmolVLA on 20% poisoned demonstration data; Attack A requires a joint state + split image trigger, Attack B uses [DISRUPT] to activate and [RESTORE] to suppress (no visual perturbation), and Attack C requires a conjunction of all three to trigger and can be suppressed.
    • Main results (SmolVLA/SO-100, Table 2/5): Attack-B/C achieve Primary ASR 63–71%, CA 100%, Secondary ASR 0% under suppression, and [RESTORE]-Normal-SR 86–95%; the IF detection rate drops from 96–98% for State Backdoor to 1–18%, and the authors say L2 divergence drops 10–14×.
    • Detection resistance: All three attacks evade STRIP, but ADS detects B and C on SO-100, with repeated probing yielding a secondary ASR of 66–73%, exceeding the 30% threshold; the authors say the property of B and C is an attacker-controllable "kill switch".
    • Transfer (OpenVLA/LIBERO, Table 6): Primary ASR drops to 24.2–54.8%, but [RESTORE]-Fail ≤3% and IF 7–15%; the authors say the mechanism can transfer to a 7B policy, with effectiveness depending on the poisoning budget and adaptation method.
    • Authors' conclusion: Conditional suppression is an important dimension for evaluating VLA backdoor threats; limitations include limited evaluation scope, evasion not being universal, the main metric being a first-step proxy, and physical validation being only a scripted demonstration.

Related papers

Read originalarxiv.org