acceptodds
Withdrawn from ICLR 2027

Window-Level Differentially Private Fine-Tuning of Vision-Language-Action Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Adapting vision–language–action (VLA) models to local demonstrations creates a training-use privacy risk: generated actions can reveal whether a demonstration contributed to fine-tuning. Addressing this risk requires controlling annotation influence while preserving closed-loop behavior, where each action affects the observations and decisions that follow. We study this problem through window-level differential privacy (DP) for action annotations, with visual observations, instructions, and robot states held public and fixed. We introduce , a fine-tuning recipe that couples tighter gradient clipping with noise-aware second-moment correction while retaining the VLA's native flow-matching objective. Clipping controls window contributions and the absolute Gaussian perturbation; moment correction adjusts AdamW's adaptive scaling for the known privacy-noise variance. Both changes preserve the privacy accounting of the baseline . To audit membership evidence beyond raw prediction accuracy, we complement with the (), which compares action errors against candidate-specific models trained without the annotation. Evaluating \(\pi_0.5\) across 40 tasks in four LIBERO suites, with 800 rollouts per model, achieves higher overall task success than at all seven tested privacy budgets. At \((\epsilon,\delta)=(10^9,10^-5)\), success increases from 9.25% to 46.25%, a gain of 37 percentage points and a fivefold improvement. Fixed-clipping comparisons reproduce the benefits of moment correction across two training seeds at \(\epsilon=10^8\) and \(10^9\). These results identify clipping and adaptive update scaling as consequential design choices for VLA fine-tuning under matched privacy accounting. They establish substantial optimization gains in the evaluated large-\(\epsilon\) regime, while useful control with strong privacy remains an open challenge.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Differential Privacy, VLA, Vision Language Action Model, Privacy

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_P2y2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies window-level DP fine-tuning for VLA models, assuming visual observations, instructions, and robot states are public. The authors propose DP-VLA_MC, which combines tighter gradient clipping with noise-aware second-moment correction (adapted from DP-AdamBC) and a variance floor to adjust AdamW updates under flow-matching objectives. They also introduce Action-Ref, a shadow-model-calibrated membership inference metric.

Critical strengths
  1. Clean ablations on optimizer dynamics under gradient noise
  2. good writing
Critical weaknesses
  1. The paper only achieves non-trivial task utility at ϵ≥108\epsilon \ge 10^8 (with single-step ϵstep>104\epsilon_{\text{step}} > 10^4). In DP theory, ϵ=108\epsilon = 10^8 yields an upper bound factor of e108e^{10^8}, providing zero formal privacy protection. Evaluating a DP method exclusively in a non-private regime (ϵ≥103\epsilon \ge 10^3) invalidates its core contribution as a privacy-preserving framework

  2. Defining privacy at the single-window level while treating visual observations/states as public and randomly assigning ~50% of windows within the same trajectory as members is unrealistic.

  3. Even on the non-private (NP) model, the proposed Action-Ref attack achieves a pooled AUROC of only 0.5560.556 (and <0.5<0.5 on Object and Goal suites), which is barely above random guessing (0.50.5).

  4. Limited methodological novelty: DP-VLA-MC is a direct application of DP-AdamBC (Tang et al., 2024) with a variance floor, and Action-Ref is a standard difficulty-calibrated shadow attack (Watson et al., 2022).

Questions for authors
  1. Why did you choose full-parameter fine-tuning (3.6B parameters) and B=4B=4?

  2. Since each test episode contains both member and non-member windows, how much does intra-episode interpolation degrade your MIA AUROC?

  3. What is the Action-Ref AUROC if membership is split at the episode level rather than the window level?

Reviewer_kRvFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies differentially private fine-tuning of π0.5 at the level of demonstration windows. It argues that standard DP-SGD with AdamW is hurt by the injected noise in the second-moment estimate, and proposes a small clipping norm plus a correction that subtracts the noise variance from the second moment, with a floor. Across ε from 10^3 to 10^12, the corrected optimizer improves LIBERO success rates over the standard DP baseline.

Critical strengths
  1. Privacy of robot demonstration data during VLA fine-tuning is a relevant and unexplored problem, and the paper provides an end-to-end pipeline, including accounting and auditing.
  2. The optimizer correction is simple but effective.
Critical weaknesses
  1. According to my understanding, the experiments do not support the stated goal of the paper. The motivation is privacy-preserving VLA fine-tuning (e.g., protecting whether a patient's demonstrations were used) with useful closed-loop control. What is shown is only that the corrected optimizer outperforms the base at budgets that give no meaningful protection, with an audit that cannot detect leakage even without DP, and with a window-level unit that is much weaker than the motivating scenario requires.
  2. The non-private reference is heavily under-trained (limited data and relatively low SR), which limits both the utility and the privacy conclusions.
  3. The source of the gain is not isolated. With the floor active, the correction mainly enlarges the update size, which may be equivalent to a larger effective learning rate.
Questions for authors
  1. Does a tuned learning rate or norm clipping for the standard DP baseline close the gap?
  2. Can you provide the rationale for the audit that detects membership in the non-private model, so that the private result is informative?
  3. How do NP, DP-VLA_base, and DP-VLA_MC compare with a standard training budget (e.g., the official π0.5 LIBERO recipe), and how strong is the membership signal in that case?
Reviewer_KtmDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper defines the privacy-protection granularity for Vision-Language-Action (VLA) models as "action-window annotation." It introduces DP-VLA$_{MC}$, a method that—while maintaining a consistent privacy budget under window-level differential privacy (DP)—improves the success rate of closed-loop robotic tasks following private fine-tuning by employing smaller gradient-clipping thresholds and a privacy-noise-aware second-moment correction for the AdamW optimizer.

Critical strengths

This paper introduces the concept of "window-level action-annotation membership privacy" in VLA fine-tuning.

Critical weaknesses
  1. An excessively large privacy budget results in very weak privacy protection.
  2. Given the current privacy budget, the utility of the model itself is significantly degraded; under such circumstances, discussing privacy seems largely unnecessary.
  3. Membership Inference Attacks are typically evaluated within the same environment as the model's training; however, in real-world deployments, the environment will inevitably differ from the training environment. Since the observations would be entirely different in such cases, it remains uncertain whether MIA would still be effective.
  4. It is unclear how the window is selected, and it is uncertain whether this choice might further impact performance.
Questions for authors

See Critical Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.