acceptodds
Under review as a conference paper at ICLR 2027

Ghost: Manifold-Guided Adversarial Substitution for Next-POI Learning Protection

Abstract

Check-in trajectories support mobility research, but the same temporal structure can train models to predict users' next locations. Protecting their training value poses a discrete optimization problem: substitutions must respect the semantics and geography of visited places while remaining disruptive after an adversary attempts to repair them. We introduce Ghost, a trajectory perturbation framework that combines an adversarial surrogate score with the conditional likelihood of a frozen trajectory language model. An alternating optimization procedure updates substitutions as the surrogate adapts, and stochastic candidate selection couples predictive disruption with a learned preference for real mobility patterns. A Gibbs variational interpretation exposes the rule as an adversarial tilt of a trajectory prior and makes the roles of likelihood and sampling temperature explicit. Experiments on Foursquare-NYC and Foursquare-TKY compare direct training with three paired-data restoration procedures. Under the evaluated protocols, Ghost reduces direct-training acc@1 by 60.6% on NYC and 72.1% on TKY relative to its clean-training baselines, and obtains the lowest mean accuracy after bigram-conditioned restoration among the compared methods on both datasets. Existing sensitivity results further characterize the roles of prior weighting, added entropy regularization, and leakage. These findings support manifold-guided substitution as a practical design direction for trajectory learning protection, while leaving unrestricted release security and cross-model transferability open. Code is available at https://anonymous.4open.science/r/Ghostt.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Trajectory Data Protection, Unlearnable Examples, Next-POI Prediction, Manifold-Guided Perturbation, Purification Resistance

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

This paper proposes Ghost, a method to perturb the next POI trajectories. It enumerates the candidates near the original POI under the same category, uses a surrogate model to find candidates that disrupt the next POI prediction and a frozen trajectory model to find candidates that are consistent with the preceding trajectory, then stochastically chooses a replacement according to a weighted combination of the two scores. Experiments show that Ghost substantially reduces clean-test accuracy under direct protected data training and achieves competitive results under sequence-denoising, unigram-inversion, and bigram-inversion restoration procedures.

Critical strengths

The method is simple and interpretable, the constrained candidate construction incorporates the geographic and categorical structure of POIs, while the two scoring components balance predictive disruption against trajectory consistency. The stochastic selection mechanism also avoids relying on a deterministic replacement rule.

Critical weaknesses
  1. The setting is unrealistic, this paper propose a method to perturb the POI trajectories, but didn't give a realistic purpose the protected data. In other domains of Unlearnable Examples, images may be public for people to view, but not allowed for training, texts may be public for people to read, but not allowed for training. Individual trajectory records, however, have little direct human-facing utility and are mainly released for computational analysis. The paper does not clearly explain why an organization would release trajectories that are intentionally made unsuitable for next-POI learning or what meaningful uses the protected data are expected to retain. And the paper does not adequately evaluate whether the protected trajectories preserve statistical utility.

  2. GHOST uses a surrogate model to score and select candidates, but the experiments evaluate only one Transformer-based victim that is architecturally close to the surrogate. In practice, a data recipient is free to use a different model architecture or training procedure. It is therefore unclear whether the generated perturbations remain effective against different settings.

Questions for authors

see weaknesses.

Reviewer_iMq5OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes GHOST, a trajectory perturbation framework that protects check-in data from next-POI prediction by substituting each POI with a semantically and geographically constrained alternative. Substitutions are selected via stochastic sampling guided by two signals: an adversarial score (weakening surrogate prediction) and a frozen trajectory LM's conditional likelihood (favoring natural mobility). A bilevel optimization alternates between candidate sampling and surrogate refitting. Experiments on Foursquare-NYC and Foursquare-TKY compare GHOST against three baselines under four attack protocols, showing competitive direct-training suppression and favorable post-restoration accuracy.

Critical strengths
  • The problem formulation is well-motivated, identifying two distinct challenges — discrete POI substitution and paired-data restoration attacks — and addressing both within a unified framework. The separation of predictive disruption, trajectory compatibility, and inverse recoverability as distinct properties is an insightful conceptual contribution.

  • The Gibbs variational interpretation (Eq. 7, 9) provides a principled theoretical lens, showing the sampling rule as an adversarial tilt of a temperature-adjusted trajectory prior rather than a heuristic combination.

  • The threat model is thorough, evaluating not only direct-training suppression but also three restoration protocols of increasing sophistication (A2–A4), including a bigram-conditioned inverter that tests sequence-dependent recoverability.

  • The stochastic sampling design is a sensible response to restoration attacks — distributing an anchor's outputs across multiple candidates complicates frequency-table inversion, confirmed empirically by GHOST's lowest mean A4-restored accuracy.

  • The paper is commendably honest about limitations, repeatedly cautioning that results are empirical and protocol-specific.

Critical weaknesses
  • The evaluation uses only a single, shallow victim architecture — a two-layer Transformer with 128-dim embeddings — shared across victim, surrogate, and frozen LM. No cross-architecture experiments are conducted. A more capable victim could potentially extract residual signal from protected data; the paper acknowledges this risk but provides no evidence to bound it.

  • Both datasets originate from the same source (TSMC2014) with identical preprocessing, covering only one platform and one data modality. Whether GHOST generalizes to GPS traces, taxi data, or telecom CDRs with different spatiotemporal characteristics remains unaddressed.

  • The weight ablation (α,β)(α,β) has a fundamental confound: changing the weights simultaneously alters the adversarial-to-manifold ratio and the absolute score scale relative to the fixed τ=0.3τ=0.3, making it impossible to isolate the effect of term asymmetry from induced sampling sharpness. The paper acknowledges this but does not resolve it (e.g., via normalization or joint ττ sweep). Additionally, no single-component ablation (adversarial-only or manifold-only) is provided to test necessity of each term.

  • A1 retains clean labels while restored branches derive labels from reconstructed sessions, confounding survival-gap comparisons. Baseline compute is also unmatched — GHOST uses five inner epochs versus three for PGD and EM.

Questions for authors

The A4 advantage over PGD is only ~0.4 and ~0.3 percentage points with overlapping standard deviations across three seeds. Can the authors clarify whether this ranking is statistically robust?

The default sampler has no entropy floor, allowing near-deterministic substitutions at concentrated positions. Have the authors analyzed whether certain POI categories or positions are systematically subject to near-deterministic substitution, creating exploitable structure?

The frozen LM guides both optimization and naturalness evaluation (Table A1). Has any independent trajectory model been used to validate that GHOST's substitutions generalize beyond the optimized LM?

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

The paper introduces a trajectory perturbation method named GHOST to suppress next-POI learning from released check-in data. For each POI, GHOST would sample a same category and geographically local replacement by combining an adversarial score and a frozen trajectory LM likelihood. Here, the adversarial score aims to reduce the surrogate model’s support for the true next POI. The frozen trajectory LM likelihood aims to preserve trajectory mobility patterns. The surrogate is iteratively fitted on the perturbed histories to adapt the substitutions over several rounds. Experiments on Foursquare-NYC and Foursquare-TKY show drops in direct-training accuracy and competitive performance under three restoration procedures based on neural denoising, unigram inversion, and bigram inversion.

Critical strengths
  1. The paper studies an interesting problem of suppressing next-POI learning while preserving basic geographic and semantic consistency in discrete trajectories. The idea of distinguishing direct training suppression from recovery after clean-protected pairs are leaked is useful and makes the evaluation more realistic than measuring only accuracy on perturbed data.

  2. The experiments evaluate GHOST under multiple restoration procedures. Although the gains are not uniform across all settings, the results show a competitive protection-restoration trade-off on two mobility datasets and provide useful diagnostics of the method’s sensitivity to score weighting, entropy, and leakage.

Critical weaknesses
  1. The evaluation is too narrow to support the paper’s protection claims. It considers only one victim architecture and a small set of restoration procedures, while the clean-label protocol and restricted release setting do not capture stronger model mismatch, sequence restoration, or linkage-based attacks. The results therefore demonstrate suppression under selected pipelines rather than robust protection against adaptive attackers.

  2. The benefit of the proposed manifold-guided bilevel design is not convincing. GHOST combines LM likelihood, asymmetric score weights, stochastic sampling, and surrogate refitting, while the baselines use different training budgets and the ablations do not cleanly separate these components. Since PGD already achieves comparable direct-training suppression in several settings, it remains unclear whether the added complexity provides a principled advantage or merely reflects a favorable operating point.

Questions for authors
  1. How sensitive are the conclusions to the victim model and training configuration? Are there any evidence on whether the protection transfers beyond the single Transformer architecture used in the experiments?

  2. What is the independent contribution of the frozen trajectory LM, stochastic sampling, and surrogate adaptation? Since the current ablations change multiple factors simultaneously, can the authors clarify which component is responsible for the improvement over the simpler PGD baseline?

  3. Does surrogate adaptation make the substitution patterns more or less recoverable? Please clarify whether restoration performance or substitution-map statistics were tracked across outer iterations, and how the authors rule out the possibility that the bilevel updates reduce direct-training accuracy while making inversion easier.

  4. What is the justification for combining the adversarial and manifold scores with (α,β)=(2,0.5)(\alpha, \beta)=(2, 0.5)? Is this choice based on a principled argument or primarily selected empirically?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.