acceptodds
Under review as a conference paper at ICLR 2027

Escaping Silent Failure: Auditable Adaptive Clipping under Client-Level Local Differential Privacy

Abstract

Client-level local differential privacy (LDP) trusts no party: each client noises its own update before transmission. At over rounds without amplification, the released update carries Gaussian noise of norm with , so the clipping radius sets the noise scale outright. We show that recent local-model designs fail silently, and how to build one whose failure modes are auditable. For the released-norm rules we analyse (scalar per-client statistics of noised released norms) the three natural routes fail: a capped statistic saturates and the radius follows a data-independent schedule (the plug-in map freezes it at a constant); an uncapped one is pinned at the largest admissible radius; uncapped debiased squared-norm estimation has prohibitive variance at our calibration. Their closed-form target omits the model dimension and, relative to the error bound, is too large by . We derive three design laws for auditable scale tracking: pay for a privatized pre-noise statistic; set its cap at the tracked update scale, not at a small multiple of the radius; keep cohort size times averaging window above an explicit threshold. A reference algorithm meets all three at radius-independent Rényi-DP cost. Every prediction was fixed before its run and the failed ones are reported; the failure modes occur in real training (saturation on two architectures). The design settles above in every CNN and MLP run at . Under one accounting, the released-norm plug-in rule diverges in every cell (6.9–12.4% final accuracy against 19.9–38.5%), and the design leads a budget-matched randomized-response quantile tracker and a DP-FedPUAC port in every cell when paired ( to points, none significant after Holm correction at five seeds). Because the regime is fully clipped, the radius matters only to an order of magnitude: three statistic-only releases before training shed an over-estimated prior in the tested range, and continued tracking adds only and points over freezing the calibrated radius. In-budget sweeps cost 4–14 points; an offline oracle sweep the threat model does not grant is 1–5.3 points ahead. We distill the results into deployment checks and audit signals computable from released messages alone.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Local differential privacy, federated learning, adaptive gradient clipping, Renyi differential 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_BKSuOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2(+1)
Soundness
2(−1)
Read review
Paper summary

As best as I can piece together through the paper's very dense presentation, the work studies adaptive update clipping for federated learning under client-level local DP. It argues that existing heuristics all fail "silently", meaning the clipping radius can become decoupled from the true data in an indetectable way. Then the authors propose three "design laws" that function as axioms that an acceptable adaptive clipping mechanism should satisfy and propose a new algorithm. They validate it on some small-scale federated tasks, comparing to existing approaches.

Critical strengths

The conceptual framework and identifying of flaws in existing approaches may be of potential value. Although I had extreme difficulty understanding the work on my own, my AI tool praised the paper's "radical (almost obsessive) empirical honesty", and said that the negative results are genuine and important, although of limited practical utility.

Critical weaknesses

The paper is so telegraphic as to be essentially unreadable (by humans). Let's begin with the abstract. Rather than providing an accessible overview of the work, it reads like a severely compressed log of experiments, full of very precise claims, many of which refer to concepts that have not yet been defined. At first, I was able to guess the intention ("client level" seems to imply federated learning, although that has not been stated, dd is plausibly assumed to be the number of parameters) but as it went on I became bewildered.

  • "The design settles above CminC_\text{min} in every CNN and MLP run at m≥100m \geq 100." What is CminC_\text{min}? What is mm? Why is settling above that an achievement?
  • "Under one accounting, the released-norm plug-in rule diverges in every cell (6.9–12.4% final accuracy against 19.9–38.5%)" What is the "one accounting" and is it one that matters? What is the "released-norm plug-in rule"? What is a "cell"?

The paper continues in this extremely telegraphic style. To be clear, the problem is not density, per se, but that throughout the paper statements are made that are literally incomprehensible unless one has context on the entire work at once, including the appendices.

It is possible there is scientific value in this work for AI agents to process, but it is nearly useless as a communication tool for humans.

Questions for authors

I have no questions. Even if the authors were able to clarify the value of the paper in rebuttal, I would not recommend this paper for acceptance without a complete rewrite for clarity and human readability.

Reviewer_kW3XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines adaptive clipping under client-level local differential privacy, where clients add noise to their own updates. It shows how adaptation based on released update norms can fail: capped statistics saturate, uncapped feedback pushes the radius to its boundary, and debiased squared-norm estimates have high variance. The authors derive a dimension-dependent bound-level radius and propose a separately privatized pre-noise statistic, tracked through a self-referencing cap and a normalized moving average. They provide an adaptive RDP composition analysis and experiments testing the predicted failures, initialization effects, and budget-matched alternatives. The experiments also separate start-up calibration from continued tracking; freezing the radius after calibration gives similar accuracy in the tested conditions. The main contribution is the diagnosis of these failure mechanisms and the resulting guidance for designing clipping rules under LDP.

Critical strengths
  1. The paper explains saturation and boundary pinning under explicit assumptions, providing a concrete diagnosis of released-norm adaptation failures.
  2. The two-channel RDP analysis explicitly accounts for statistic releases and warm-up costs under adaptive clipping.
  3. The experiments distinguish start-up calibration from continued tracking and report failed predictions and non-significant results, making the evidence easier to assess.
Critical weaknesses
  1. The benefit of private calibration is not clearly separated from the effects of the public floor, initialization, and radius–Adam scaling.
  2. The frozen baseline still pays for unused statistic releases, leaving the comparison with a simpler budget-matched alternative incomplete.
  3. The correspondence between Equations (2)–(3) and the cited methods is unclear.
  4. The A1–A2 no-false-negative guarantee lacks a detection horizon, since contraction need not reach the boundary during training.
Questions for authors
  1. When fixed clipping receives the same public floor and initialization information, and the radius–Adam scaling relationship is handled consistently, what benefit remains from private calibration?
  2. How does warm-up followed by freezing perform if subsequent statistic releases are stopped and the update noise is recalibrated to the same total privacy budget?
  3. Which equations in the cited methods correspond to Equations (2)–(3), and which parts are representative rules constructed for this analysis?
  4. What detection horizon is required for the A1–A2 guarantee when Theorem 3 yields contraction without boundary attainment during training?
Reviewer_jjHSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates adaptive clipping in federated learning under client-level local differential privacy (LDP). It identifies several failure modes in existing clipping adaptation strategies that estimate update scales from privatized model updates, and proposes a separate privatized statistic channel to support more reliable clipping adaptation. Theoretical analysis and federated learning experiments validate these failure modes, although the proposed method offers limited improvements over fixed clipping following an initial private calibration.

Critical strengths
  1. The paper addresses a relevant problem in privacy-preserving federated learning and identifies several failure modes of adaptive clipping under client-level LDP.
  2. The paper provides extensive experiments and ablation studies, and the authors are transparent about negative results and the limitations of their proposed approach.
Critical weaknesses
  1. Limited practical benefits of the proposed adaptive clipping method. While the paper identifies meaningful failures in existing clipping strategies, the practical advantages of the proposed method remain unclear. In particular, Appendix D.10 shows that performing a short private calibration and then fixing the clipping radius achieves nearly identical accuracy, with differences of only 0.3 and 0.9 percentage points. Moreover, this fixed-radius baseline continues to spend privacy budget on unused statistics. I encourage the authors to compare against a stronger calibration-then-freeze baseline that reallocates the remaining privacy budget to model updates. This would help establish whether continued adaptation is actually necessary.

  2. A gap between the theoretical clipping target and empirical utility. The proposed clipping radius is derived by minimizing an estimation-error bound, but the paper acknowledges that this target can be 13–20 times smaller than the empirically preferred radius. Furthermore, the theoretical derivation depends on an uncapped moment, whereas the proposed tracker estimates a capped moment that can be substantially smaller. These discrepancies weaken the connection between the theoretical analysis and the practical algorithm. The authors should better justify why the proposed target is appropriate for federated optimization, or investigate whether a more learning-oriented objective could produce a better clipping rule.

  3. The generality of the identified failures requires further justification. Several theoretical results concern a specific class of adaptation rules based on released update norms, and the main plug-in baseline is not a direct reproduction of an established method. For instance, FedDimDP uses a median-based update under a different privacy setting, and its behavior is examined primarily through a simplified norm-level simulation. Thus, it remains unclear how broadly the reported failures apply to existing adaptive clipping methods. The authors should more carefully distinguish general failure mechanisms from the limitations of particular implementations, and provide stronger evidence using representative published methods where feasible.

  4. The scope of the proposed auditability guarantees is relatively narrow. Although auditability is emphasized as a central contribution, the proposed signals primarily detect whether the clipping radius becomes frozen or reaches a predefined boundary. They do not establish that the tracked radius accurately reflects the underlying update distribution or is beneficial for training. The paper also acknowledges that manipulated statistic reports may remain undetected. It would be helpful to clarify the practical scope of these guarantees and investigate whether more informative diagnostics can detect inaccurate or ineffective adaptation before visible boundary failures occur.

Questions for authors
  1. Is continued adaptation necessary? How would the proposed method compare against private calibration followed by fixed clipping if the baseline stopped releasing statistics and reallocated the saved privacy budget to model updates?

  2. How broadly do the failure results apply? Can the authors clarify which published adaptive clipping methods are directly covered by their analysis and whether the predicted failures can be reproduced using faithful implementations under comparable privacy assumptions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.