acceptodds
Under review as a conference paper at ICLR 2027

DACS: From Historical Inertia to Forward Evidence for Behavioral Authentication

Abstract

In behavioral biometric authentication under user drift, forward-only adaptation is not merely an optimization problem; it is an information-selection problem under a restricted forward budget (≤20 evaluations): should a drifted verifier be updated using historical training movement or measured forward responses? Testing historical movement reveals that trajectory transfer provides little benefit, yielding only +0.504 pp on H-MOG and severe negative transfer on HuMI (-5.940 pp), showing that historical movement is not a reliable adaptation signal. We introduce DACS, a response-grounded alternative that replaces trajectory inheritance with two complementary sources of forward evidence. Controlled Component Attribution (ComAtt) identifies responsive components from development interventions, while current-user search resolves independent signed edits through a disjoint validation gate. On BehaveFormer authentication cohorts, four-component DACS improves balanced accuracy by 3.037 pp on H-MOG and 1.646 pp on HuMI within twenty objective evaluations, with factorial controls confirming a positive attribution–direction interaction. Evaluated through dual lenses, DACS delivers: (a) a lightweight seventeen-evaluation discrete variant retaining 76.5–79.7% of continuous-search gains; and (b) a full-cohort intention-to-treat audit achieving 5.3×/10.3× larger population-level gains than Random-k while limiting observed severe deterioration to 2.82% on H-MOG and 6.75% on HuMI. These results support an information-anchor shift: exposing plasticity through measured component responses and directing it using present-user evidence.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
behavioral authentication, test-time adaptation, forward-only adaptation, distribution drift, component attribution, derivative-free optimization, edge 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_uwioOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies how to adapt a behavioral authentication model when user behavior changes, using only forward passes. DACS uses ablations on development data to select four model components, and then it searches for a separate scaling factor for each component using current-user data. An update is applied only if it improves BA on a separate validation split. With 20 objective evaluations, the paper reports BA gains of 3.037 percentage points on H-MOG and 1.646 on HuMI. The experiments also examine component selection and update directions, a discrete search version, and a possible application to a vision task.

Critical strengths
  1. I like how little has to change in the model. You just rescale existing components, so there's no extra computation at inference. For k = 8, even the simpler 17-evaluation search keeps about 77–80% of the gain from the 20-evaluation GP search, without have to fit a GP.

  2. The Fisher control in Table 3 is so useful. ComAtt is about 0.4pp ahead on both datasets. Although it is not a huge jump, but the difference holds up in the paired user-level tests. The direction comparison is more interesting to me: with ComAtt, letting components scale down as well as up gives about 1.59 points on H-MOG and 1.27 on HuMI over positive-only updates.

Critical weaknesses
  1. I'm not quite convinced by the conclusion about historical directions. HMGSS has just one scalar for moving along a historical weight difference, while DACS can adjust four selected components separately. So there is two things change here, including the information source and the freedom to update the model. Another point is that I couldn't get from Supplementary Section 2.1 whether HMGSS used the same validation gate. This could be a limitation of the particular HMGSS setup, so my veiwpoint is that it is difficult to draw a general conclusion from its failure.

  2. I am wodering that what happens to false accepts after adaptation. BA can go up even when FAR gets worse, e.g., if genuine-user acceptance improves enough to offset it. Figure 2 is only a schematic, and the actual FAR and FRR numbers aren't reported. Just keeping the threshold fixed doesn't guarantee the same FAR once the model changes. Without those numbers, I can't really judge the claim that impostor rejection is preserved.

Questions for authors
  1. What checkpoints define the historical trajectory of HMGSS? Does it use the same search objectives, "no-update" options, evaluation budgets, and verification gating mechanisms as DACS?

  2. Could you give more about the False Acceptance Rate (FAR) and False Rejection Rate (FRR), just like Static and DACS on Day 5. i think you can use a fixed threshold? Furthermore, please clarify whether the impostor samples used during the testing phase are distinct from those used during the search and verification phases. THis would help in assessing the practical trade-offs in authentication performance.

Reviewer_93QEOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Expertise: I have prior experience with keystroke authentication, but limited familiarity with forward-only adaptation, zeroth-order optimization, and component attribution. My assessment of methodological novelty is therefore limited.

This paper proposes DACS for adapting behavioral authentication models under user drift without backpropagation. It first selects editable components through offline attribution, then searches for user-specific signed adjustments within a small evaluation budget. An independent validation split determines whether to deploy the update. Experiments on H-MOG and HuMI show improvements over matched search baselines.

Critical strengths
  1. The method clearly separates component selection from update search, with ablations supporting their complementary contributions.
  2. The evaluation considers both performance gains and deterioration risk, including full-cohort results and failure analysis.
Critical weaknesses
  1. The conclusions about historical information appear broader than the evidence supports. The failure of one historical-direction baseline does not establish that historical information is generally ineffective for adaptation.
  2. Practical edge-device feasibility remains insufficiently demonstrated. Timing is measured on an RTX 3090, and the proposal-evaluation budget does not capture the complete adaptation cost, including validation.
Questions for authors

N/A

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

This paper studies forward-only adaptation of behavioral biometric authentication models under user drift. The authors argue that historical parameter movements provide unreliable adaptation directions and propose DACS, which combines offline component attribution with derivative-free optimization based on current-user responses. On H-MOG and HuMI, DACS improves balanced accuracy by 3.037 and 1.646 percentage points, respectively, within a budget of 20 forward evaluations.

Critical strengths

The paper addresses a practically important problem: adapting deployed models without backpropagation under tight computational constraints. The experiments include multiple derivative-free optimization baselines and useful ablation studies isolating the contributions of component selection and signed updates.

Critical weaknesses

(Major weakness) The practical feasibility of the proposed adaptation procedure is unclear. DACS optimizes balanced accuracy using labeled current-user windows and a cached impostor gallery. However, continuous authentication systems generally cannot know whether incoming observations belong to genuine users. The paper does not sufficiently explain how the required genuine-user labels and drift-triggering signals would be obtained during actual deployment.

(Major weakness) The evaluation does not adequately establish the security implications of adaptation. Improvements in balanced accuracy can conceal increases in false acceptance rates, a particularly consequential metric for biometric authentication. Moreover, Table 4 reports severe performance deterioration for 2.82% of H-MOG users and 6.75% of HuMI users, raising concerns about deployment safety.

(Major weakness) The methodological novelty appears limited. DACS largely combines component-wise forward ablation with low-dimensional derivative-free optimization, both familiar techniques. While the experiments demonstrate the usefulness of their combination, the paper does not clearly establish what fundamentally new methodological insight DACS provides.

Questions for authors

How can the proposed method obtain reliable genuine-user labels and detect performance degradation during deployment without ground-truth authentication outcomes? Can the authors report false acceptance and false rejection rates separately, and explain how severe deterioration for individual users could be prevented? What distinguishes DACS from existing component-selection and derivative-free optimization approaches? Could stronger historical-information baselines or other block-coordinate optimization methods substantially narrow the reported performance advantage?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.