acceptodds
Under review as a conference paper at ICLR 2027

Test-time Adversarial Takeover: A Real-time Hijacking Interface against Robotic Diffusion Policies

Abstract

Diffusion-based action generation has become a foundational component of embodied AI, but its reliance on visual conditioning leaves deployed visuomotor policies vulnerable to adversarial manipulation. Most prior attacks focus on disruption: they perturb the observation stream to reduce task success or induce erratic behavior. We study a stronger threat, Test-time Adversarial Takeover (TAKO), in which an attacker obtains a real-time steering interface over a frozen robot policy and turns it into a remotely piloted instrument. TAKO learns a small vocabulary of reusable universal patches through differentiable diffusion inference; at test time, the attacker switches among these patches in the camera stream to compose attacker-chosen trajectories. This works because the perturbation acts on the visual conditioning pathway, where the induced bias can persist through iterative generative inference. We further show that the natural targeted baseline, target-policy matching, fails because the victim policy cannot reliably supervise itself on out-of-distribution target shifts. Across four tasks (2D manipulation, simulated aerial delivery, simulated ground navigation, and physical-world ground navigation), two visual encoders (ResNet-18 and EfficientNet-B0 + Transformer), and three generative inference families (DDPM, DDIM, and flow matching), human operators achieve 100% takeover success on attacker-defined objectives in every evaluated setting.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
adversarial attack, adversarial patch, diffusion policy, robot learning, visuomotor policy, robot security, embodied AI, imitation learning

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_Ec1FOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces Test-time Adversarial Takeover (TAKO), a framework that learns a vocabulary of universal adversarial patches offline via differentiable diffusion inference to induce directional action biases in frozen robotic diffusion policies. At test time, a human operator switches among these patches in the camera stream to steer the robot in closed loop across 2D manipulation, aerial delivery, and ground navigation tasks (spanning ResNet/Transformer backbones and DDPM/DDIM/flow-matching samplers), achieving 100% takeover success while target-policy matching (TPM) baselines fail completely.

Critical strengths
  1. Novel composable takeover formulation

  2. Broad cross-architecture and real-robot validation

  3. good writing

Critical weaknesses
  1. Table 1 claims "physical realizability", yet the real-world navigation experiment digitally overlays patches onto the camera stream via software rather than placing physical printed patches in the environment. If an attacker already possesses man-in-the-middle access to modify the live digital sensor stream in real time, hijacking the robot via visual patches is less direct than simply spoofing downstream action commands, weakening the practical justification of the threat model.

  2. The quantitative evaluation only compares TAKO against a self-constructed, completely failing TPM baseline (0% ASR) without adapting existing diffusion-policy attacks (e.g., DP-Attacker) for targeted steering comparisons.

Questions for authors
  1. Given that patches are digitally injected into the camera stream in RealNav, could you justify claiming "Physical realizability" in Table 1 and clarify a realistic security scenario where an attacker can arbitrarily overwrite live camera frames in real time but cannot directly alter the policy's action outputs or control topics?

  2. Which specific inference sampler was used in Table 4 to achieve 40/40 online takeover success, and how does human control perform when the offline per-step directional success rate is relatively low (e.g., 65.8% for SimNav DDPM in Table 3)? Additionally, can stronger baselines (such as adapting DP-Attacker with a directional objective) be evaluated?

Reviewer_U3mEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TAKO, a test-time adversarial attack framework that aims to take control of a deployed diffusion-based robot policy instead of only making it fail. It optimizes four universal patches offline, one for each direction, by backpropagating a directional loss through the full denoising process. At test time, a human operator switches among these patches in the camera input to steer the robot to an attacker-chosen target. Experiments on three simulated tasks and one real-robot navigation task show that operators reach the target in all trials, while the TPM baseline proposed by the authors fails in all trials.

Critical strengths

The paper studies a meaningful security problem for diffusion-based robot policies. Most prior attacks only try to make the robot fail, while this paper considers an attacker who can continuously steer the robot to a chosen target, which is a more serious threat in practice. The threat model is clearly defined in Section 3.1, and Figure 2 illustrates the offline and online stages well.

The proposed design is simple and easy to use. Instead of specifying a full target behavior, the attacker only needs a few directional patches and can correct the drift online by switching among them. The method is also tested on several tasks, including a real robot, and with three inference families, which shows that the idea is not limited to one specific setting.

Critical weaknesses

My main concerns are about the baseline comparison and the limited ablation study:

The setting of the TPM baseline is not clearly described. To build the reference action, the goal is “displaced along direction c” (Eq. 3), but the paper does not say how far the goal is moved. This distance may directly change the reference action, since a small shift may still be close to the training data while a large shift is more likely out-of-distribution. It is also unclear how TPM is deployed in Table 4, i.e., whether a single patch is held open-loop or the operators also switch among TPM patches in real time, and how the goal-shifted observations are built for DeliverDrone, SimNav and RealNav. Could the authors clarify how the shift distance is chosen and describe the full TPM setting for each task?

The comparison with existing attacks is missing. TPM is a baseline defined by the authors themselves, and the paper calls it “the standard targeted-attack template” without citing any prior work that uses it. Meanwhile, existing attacks on diffusion policies, such as DP-Attacker (Chen et al., 2024) listed in Table 1, are not compared in the experiments. As a result, it is hard to judge how TAKO performs against prior methods. Could the authors explain why no existing attack is used as a baseline, or report such a comparison?

The ablation study is limited. All ablations are done only on PushT, which is the simplest task and where the patches already reach about 99% directional success (Table 3). It is unclear whether these findings still hold on the other tasks, especially on the physical RealNav task. In addition, the patch placement ablation (Table A1d) only tests the “right” direction, while the caption claims that center is the “direction-agnostic optimum”. It would be better to add the ablations on RealNav and to test all four directions in the placement ablation.

Questions for authors

Please clarify the TPM setting. How far is the goal moved when building the reference action, and how is this distance chosen? In Table 4, is a single TPM patch held open-loop, or do the operators also switch among TPM patches in real time as in TAKO? It would also help to explain how the goal-shifted observations are built for DeliverDrone, SimNav and RealNav.

Please explain why no existing attack, such as DP-Attacker (Chen et al., 2024), is used as a baseline, since TPM is defined by the authors themselves. Even a comparison on one task under the same setting would help show whether TAKO really works better than prior methods.

Please extend the ablation study. All ablations are currently done on PushT, so please repeat the main ones, such as patch size and perturbation budget, on the physical RealNav task to check whether the findings still hold in the real world. In addition, Table A1d only tests the “right” direction, so please report the placement results for all four directions to support the claim that center is the “direction-agnostic optimum”.

Reviewer_TDF3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Test-time Adversarial Takeover (TAKO), which learns four universal patches as directional primitives and lets a human switch them to steer a frozen diffusion policy. The experiments show only that a large digital overlay can bias planar policies and that the authors can teleoperate the resulting interface. They do not support the diffusion-specific, physically realized, or universally successful attack claimed in the abstract. I recommend rejection.

Critical strengths
  1. The authors test one steering interface on manipulation, aerial delivery, and ground navigation, including a physical robot, and they check that the clean policies solve their original tasks.
  2. The authors ablate patch size and perturbation budget, and they report the small-patch regime in which a human can no longer finish the task.
Critical weaknesses
  1. The authors treat a broken baseline as a fundamental result. Target-policy matching asks the victim to imitate its own output on a shifted goal (Eq. 3), which Section 4.4 shows is almost invariant to that goal, and no trial matches an externally specified action. The claim that targeted matching is incompatible with robotic control is therefore unsupported.
  2. The authors blame diffusion without testing the obvious alternative. A bias in the visual condition remains under further denoising for any conditional policy, and Table A1(c) only restates that fact. With no behavior-cloning or other non-generative control, the diffusion-specific and architecture-general claims are assertions.
  3. The authors report a 100% takeover that their protocol cannot justify. The abstract covers four tasks, two encoders, and three samplers, but the human study is 40 research-team trials per task, with no sampler split, no time limit, and no geometric success criterion. This measures whether the designers can drive their own interface.
  4. The authors upgrade a visible digital patch into a new physical threat. Table 1 marks physical realizability and unique real-time control, while the limitation admits white-box digital injection and task completion requires a patch covering about a quarter of the image. Prior test-time visual attacks already steer robot behavior; four switchable buttons do not create the gap shown in that table.
  5. The authors' main numbers are internally inconsistent. SimNav flow matching averages to 76.0 in Table A5 but is reported as 70.3 in Table 3, outside that cell's own confidence interval, and DeliverDrone is simultaneously 100% and 99.5% under DDIM. A paper making a universality claim cannot rest on tables that disagree with each other.
Questions for authors
  1. Were the 40 human trials run on one sampler, or separately on DDPM, DDIM, and flow matching?
  2. Was target-policy matching ever optimized against an action target other than πθ(ocshift)\pi_\theta(o^{\mathrm{shift}}_c)?
  3. Which numbers should be trusted when Table 3 and Table A5 disagree for SimNav and DeliverDrone?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.