acceptodds
Under review as a conference paper at ICLR 2027

Corrupted Blocks Are Worse Than None: AgentTrace, a Measurement Study of Latent-Source Recovery Among Identifier-Leaking Clients

Abstract

Recovering which events in an anonymized stream share a source is a well-posed learning problem that lacks scorable data without privileged identity, simulation or leakage: the predicted quantity is the field removed. AgentTrace takes labels from inside the public stream: structured client identifiers across 182,119,794 order events and 23 asset-day sessions resolve to 153 actors, withheld from every model and corroborated, unevenly across identifier families, by an identifier-blind relation. The capture cannot be redistributed; the reusable object is the protocol. The labels cover 38.99% of events and are not random: a behavior-only classifier reads labellability at 0.8790 mean AUROC over 17 sessions, under a rule fixed in advance, so this study measures recovery among identifier-leaking clients. Inside that population the signal exceeds identifier-family classification: with family held constant, version-1 UUID actors separate at 0.9688 and 0.9981 balanced accuracy. The central measurement is a factorial, none of it preregistered, that fixes featurizer, algorithm, oracle and metric and varies only the imposed boundary source. Over 16 sessions at , oracle boundaries beat random size-matched blocks by mean order-level ARI, positive in all 16, but beat not blocking at all by only , a gain that does not survive asset-level inference and reverses in all four ETH sessions. Random size-matched and contiguous blocks both score below not blocking in all 16, by and mean ARI. The fingerprint survives a change of day and of month, at 7.97 to 31.23 times chance across ordered day pairs, and transfers far more weakly across assets, 2.4 times chance pooled and 4.0 once the one build-keyed family is excluded. A label-free contrastive encoder shows no consistent held-out-day gain at oracle , yet avoids the collapse hand features suffer under density-based clustering. Neither timestamp coarsening nor jitter nor size discretization closes the behavioral channel at any magnitude we could build against a refit hand-feature attacker; jitter of at least 1 ms deletes the timing-metadata side channel and a 1% size grid the size-precision fingerprint at 0.3% distortion.

Primary area:
datasets and benchmarks
Keywords:
latent-source recovery, asynchronous event streams, de-anonymization

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

This paper studies whether hidden actors can be recovered from anonymized event streams using behavioral information. Prior work usually relies on known actor identities or simulated data, which makes it difficult to evaluate recovery on real anonymized streams. The authors identify a special real-world setting in historical cryptocurrency trading data where a large number of events can be associated with actor labels, making large-scale analysis possible.

Based on this dataset, the paper systematically studies latent-source recovery and finds that incorrect segmentation is consistently worse than using no segmentation, while correct segmentation provides only a small and unstable improvement. The authors conclude that behavioral information does contain recoverable actor-specific signals, but practical recovery remains difficult because the main bottleneck is not simply representation learning, but correctly identifying event boundaries and grouping events from the same source.

Critical strengths
  1. The paper presents an interesting and practical idea for studying latent-source recovery in real-world data. Leveraging structured client identifiers in historical cryptocurrency trading data to obtain actor labels is both intuitive and reasonable, and allows the authors to analyze the problem at a relatively large scale.

  2. The experimental design is comprehensive and rigorous. The paper evaluates the problem from multiple perspectives, including segmentation, temporal and cross-asset transfer, learned representations, actor-count estimation, and simple defenses, while also using controlled comparisons to isolate the effect of key factors.

Critical weaknesses
  1. The overall contribution and significance of the paper are somewhat unclear to me. Although the paper provides a careful measurement study, the main findings are relatively limited: incorrect segmentation performs worse than no segmentation, while correct segmentation only gives a small and unstable improvement. These results help characterize the problem, but they do not seem to lead to a particularly strong or surprising conclusion, nor do they provide a concrete solution to the main challenge of latent-source recovery.

  2. The generalization of the findings is limited, since the experiments only cover identifier-leaking clients. It is unclear whether the same conclusions hold for other anonymized event streams.

Questions for authors

Could the authors clarify what broader lessons or research directions they expect this work to bring to the field? In particular, beyond showing that incorrect segmentation can be harmful, it would be helpful to explain how these findings may change the way future latent-source recovery methods or evaluation protocols are designed.

2 comments

Thank you for the review and for a summary that states the findings accurately. One nuance shapes our answer: the paper reads boundaries as "a fragility, not a constraint". Correct boundaries buy little over not blocking at all, while corrupted ones cost a great deal.

[1] Significance and broader lessons

We agree the paper offers no new method. Its contribution is to change how methods are evaluated. Each practice below follows from a result in the paper, and the revised PDF states them in a new paragraph in Sec. 6 (p. 10):

  1. Score boundary-proposing methods against no blocking, not only against random blocks. Oracle boundaries beat random size-matched blocks by +0.1982 mean order-level ARI (16/16 sessions) but no blocking by only +0.0233, which is not established at the asset level and reverses in all four ETH sessions.
  2. Measure boundary errors at the order level, where they cannot be repaired downstream: corrupted blocks fall below per-order clustering in all 16 sessions (−0.1749, −0.1714).
  3. Score self-supervised encoders on held-out days, with estimated K beside oracle K. Our encoder looked better than hand features when trained and scored on the same sessions (0.2570 vs 0.2215) but shows no consistent gain on held-out days, and estimating K costs 24.7% (eigengap) to 87.2% (HDBSCAN) of mean ARI on hand features.
  4. When labels come from a leak, report a selection diagnostic whose reading is fixed in advance (see [2]).
  5. Test open-set detection on naturally unseen sources: they score 0.3599 mean AUROC here, against 0.7368 for random holdouts.
  6. Evaluate defenses against an attacker refit on the defended stream, and report day and asset transfer separately (7.97–31.23× versus 2.4× chance).

[2] Generalization beyond identifier-leaking clients

We agree, and the paper states this limit as a measured scope. A classifier given only behavioral features predicts whether an order is labelled at 0.8790 mean AUROC over 17 sessions (label-shuffle controls 0.48–0.52). Under a rule fixed before the diagnostic ran, a high value restricts every claim to identifier-leaking clients, so the title and every recovery claim carry that scope.

Within that population the signal is more than identifier-family classification: with family held constant, actors still separate at six to seven times chance in the largest family. Recoverability for clients with sound identifier hygiene is not measurable with this instrument, in either direction, and the simulator experiment (Appendix Z) tests the protocol's logic, not that question. A wider claim would exceed the evidence. What transfers is the protocol, which applies to any stream in which a source label can be read but withheld.

Thank you for the clarification. I appreciate the additional discussion and experiments, which make the paper's contributions and limitations clearer. However, I still think that the overall contribution falls below the acceptance threshold for ICLR. Therefore, I will maintain my original score.

Reviewer_8EpsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents AgentTrace, which derives reference labels from structured client identifiers to study behavior-based source recovery in real cryptocurrency order streams. Experiments examine the effects of order grouping and feature representations, evaluate generalization across time and assets, and test the detection of previously unseen sources.

Critical strengths
  1. The paper provides evidence that behavioral features can distinguish sources in real order data when same-source traces are supplied. Label-validity checks, within-family evaluations, and a selection-bias diagnostic help clarify the validity and scope of these findings.

  2. Controlled grouping comparisons help isolate the effects of order grouping. Held-out-day evaluation and parameter-matched baselines further clarify whether representation gains generalize across days and whether they are specific to attention, supporting a more careful interpretation of the results.

Critical weaknesses
  1. The main text is highly text-dense and contains only one compact results figure, without an overview diagram explaining label construction, trace formation, and the different evaluation settings. Since the interpretation depends heavily on distinctions such as segmentation-given versus unsegmented evaluation and oracle versus estimated K, the lack of a clear workflow figure or summary table makes the paper unnecessarily difficult to follow.

  2. The paper discusses prior simulation-based work but does not report a task-aligned evaluation on simulated order data or a quantitative comparison between real and simulated order data. Since the original order data cannot be redistributed, publicly available simulation frameworks, such as the one used in Classifying and Clustering Trading Agents (Wilinski et al., 2025), could provide a more reproducible setting for supplementary source-recovery experiments. Such a comparison would help characterize differences between real and simulated order data, examine how these differences may affect source recovery, and test whether the main grouping findings also hold in a simulated setting.

Questions for authors

Additional Question: Could the authors clarify whether the simulation framework used in Classifying and Clustering Trading Agents (Wilinski et al., 2025) could be adapted to support order-level source recovery under the AgentTrace protocol? Was such a comparison considered, and what were the main reasons for not including it?

1 comment

Thank you for the careful summary and for the strengths you identify. We address both weaknesses and your question.

[1] Overview figure and settings table

We agree. The revised PDF adds Figure 1 on p. 2, showing label construction, trace formation under the four boundary sources and the evaluation settings, and Table 1 on p. 5, listing for every reported setting whether segmentation and K are given, its split, metric, headline result and appendix. Together they make explicit the distinctions you name: segmentation-given T1a versus unsegmented T1b, and oracle versus estimated K.

[2] Simulation-based comparison

We agree, and have now run the paper's factorial on the simulator you name, reported in the new Appendix Z. The code is in the updated supplement.

(a) Adapting the framework. The framework adapts directly: the simulator emits order-level events with the submitting agent's identifier, which we withhold as the label, and the paper's factorial code runs on them unchanged. We used the unmodified public code at its default setting of 1,590 agents in 15 behavioral subclasses, with seeds 1–16. Agents within a subclass share identical parameters, so each seed scores one randomly drawn agent per subclass, so K = 15.

(b) Pre-declared rules. The design, decision rules and the wording of each possible reading were fixed, and their file hashed, before the simulator was cloned or run.

(c) Result. At W = 32, oracle boundaries beat random size-matched blocks by +0.2274 mean order-level ARI in 16 of 16 seeds, against +0.1982 on real data. Random size-matched and contiguous blocks score below no blocking by −0.1684 and −0.1315, both in 16 of 16 seeds, against −0.1749 and −0.1714 on real data. The pre-declared reading applies: the boundary-source ordering holds in a public simulator whose feature regime differs from real data. This supports the protocol's logic, not real-market magnitudes.

(d) Differences. The simulator scores about 3,200 orders per seed against 500,000 per real session, with integer sizes and far more tied timestamps: 72.9% of consecutive orders against 2.5%. Two results differ:

  1. Oracle minus no blocking is larger in simulation, +0.0590 at W = 32 and positive in 15 of 16 seeds, than on real data, where it is +0.0233 and not established at the asset level. We have not measured why.
  2. At W = 2, contiguous blocks beat no blocking in simulation by +0.0885, in 16 of 16 seeds. A post-hoc check outside the pre-declared design, which randomizes order within tied timestamps, removes this, giving −0.0388 with 1 of 16 seeds positive, so it reflects the simulator's emission order.

Neither difference changes a real-data claim.

Reviewer_QVX6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper "seems" to focus on privacy risks in cryptocurrency exchange data. It is really not clear what is the nature of the contribution especially because the evaluation relies on data that can not be inspected.

Critical strengths

None

Critical weaknesses

Beyond the clear readability issue, the current evaluation protocol is not reproducible. If no public benchmark exist for reproducible evaluation may be the authors could consider generating synthetic data?

Questions for authors

See critical weaknesses section

2 comments

Thank you for the review. We answer the ethics flag first, then readability and the nature of the contribution, then reproducibility.

[1] IRB status

We did not obtain IRB approval or a formal IRB determination for this study. The Ethics Statement should have said so explicitly, and the revised one does (p. 11).

Why we believe it is not human-subjects research: such research generally involves either interaction with or intervention on living individuals, or identifiable private information about them. Neither applies here:

  1. No intervention or interaction. Observational analysis of a public feed the venue broadcast to any connected client without authentication; no access control, rate limit or contractual restriction was circumvented.
  2. No private information. Broadcast events are not behavior recorded where one could reasonably expect no observation, nor information supplied on the expectation that it would not be made public.
  3. Not identifiable. Labels are pseudonymous machine-level namespaces such as a UUID node, a process or installation counter or a software build; an actor is a piece of software. We make no attempt to attribute any actor to a firm or a person, hold no auxiliary data that would allow it, and report nothing at the level of an identified party.
  4. The label channel is already closed. The venue stopped broadcasting the client identifier between our two capture eras; it occurs zero times in the 2023 captures.

We do not equate machine-level with harmless: a persistent fingerprint could expose an operator's strategy or infrastructure to anyone able to link it to a name, so the Ethics Statement treats the actors as data subjects for release purposes. This judgment is not ours alone to make, and we welcome the ethics reviewers' assessment.

[2] Readability and what the paper is

We agree the submitted text was hard to follow; the revised PDF adds an overview figure on p. 2 and a table of all evaluation settings on p. 5. In brief, the paper studies how to recover which events in an anonymized stream share a source, a problem usually unscorable on real data because the answer is the deleted field. In one public exchange feed, structured client identifiers supply labels for 153 pseudonymous actors over 182,119,794 order events, and the labels are withheld from every model. The central result is a controlled factorial: with featurizer, algorithm, oracle K and metric fixed, corrupted blocks score below not blocking at all in all 16 sessions, while correct boundaries gain only +0.0233, which is not established at the asset level. A selection diagnostic scores 0.8790 AUROC and, under a reading fixed in advance, restricts every claim to identifier-leaking clients. The paper is a measurement study and evaluation protocol, not a new method or a privacy audit of identified parties.

[3] Reproducibility and synthetic data

We agree that a third party cannot inspect the scored events. The updated supplementary material contains the derivation and evaluation code with the preregistered protocol, a SHA-256 manifest of every consumed capture file, hashed actor labels for every trace, and a quantized smoke-test sample. The README in release_r1/ documents these files and gives a smoke test that runs the released code end to end on the sample. Raw captures and derived features stay withheld: exact sizes, prices and timestamps permit a re-join against the public raw feed and hence re-derivation of the withdrawn identifier field.

On synthetic data, we followed your suggestion. The paper's factorial now also runs on the public agent-based simulator of Wilinski et al. 2025, reported in the new Appendix Z, and its code is in the supplement. Under rules fixed before the simulator was run, the boundary-source ordering seen on the real data holds there in 16 of 16 seeds. This checks the protocol's logic, not real-market magnitudes; details are in the General comment and the reply to 8Eps.

Thank you for the revision.

I still question the significance of the contribution for ICLR. Even from a privacy research perspective, the novelty appears relatively limited. The paper essentially investigates a behavioral linkage attack, a well-established privacy threat, in a specific application domain. It remains unclear what fundamentally new insights the findings provide beyond existing research on behavioral fingerprinting and linkage attacks.

Concerning the ethical considerations, I believe it would be appropriate for the authors to consult their institutional ethics board to determine whether IRB approval is required. Conducting privacy attacks on real-world data, even when publicly accessible, may raise legal and ethical concerns and could be subject to restrictions in certain jurisdictions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.