acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Admission Attacks on Semantic-ID Generative Recommenders: Detectability, Payload Size, and Downstream Promotion

Abstract

Admission-time defenses screen newly logged interaction histories before a recommender is retrained, and are typically judged by how cleanly they separate a known attack from randomly resampled benign users. We show that this evaluation can substantially overstate security. Studying semanticID generative recommendation, we define attack success jointly as passing admission and promoting a target item after retraining, and construct a staged family of non-neural shilling attacks that adapts payload, profile length, item popularity, and transition structure to the gate, culminating in an oracle search against its score. On Amazon Beauty, a gate jointly constraining four cohort statistics rejects every high-payload construction, although adapted attacks evade each staged single-statistic gate; diagonal covariance and calibrated marginal gates make identical decisions on these attacks, so this robustness stems from redundancy rather than correlation modeling. Direct search does find admitted attacks, but only at low payloads: accepted cohorts with 7, 16, and 24 target carriers promote the target to 0.00%, 0.11–0.38%, and 1.4% of eligible users, versus 5–14% for the rejected m = 67 attacks, and the same frontier appears on a second catalog. Admitted cohorts are more harmful for a medium-frequency target (3.6–3.9%) and under a non-generative SASRec recommender (4.7–6.6%). This high-payload protection, however, holds only under favorable evaluation. The same gate has a 0.11% false-positive rate on randomly resampled organic cohorts yet flags 26 of 33 time-ordered windows of genuine users, and attacks rejected in at least 95% of windows when they form half of a cohort pass almost always at a quarter or less, and still promote the target once admitted. We attribute both failures to cohort aggregation, which tightens calibration on exchangeable resamples while diluting each attacker’s contribution, and argue that cohort-level admission defenses must be evaluated against adaptive attackers, temporally ordered benign traffic, and mixed cohorts. Code will be provided after acceptance.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Recommender Systems, Shilling Attacks, Data Poisoning, Sequential Recommendation, Adversarial Machine Learning

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

This paper studies data poisoning attacks against admission-time defenses for generative recommender systems. The authors build several adaptive shilling attacks and test whether they can pass a statistical gate and promote target items after retraining. The results show that the gate blocks attacks with large payloads, but some smaller attacks can pass. The paper also finds that the gate performs poorly on time-ordered user data and mixed cohorts.

Critical strengths
  1. The paper measures both attack detection and target promotion after retraining. This gives a clearer view of attack success than detection accuracy alone.

  2. The experiments show that time-ordered data and mixed cohorts can greatly affect detection results. The evaluation also covers multiple datasets and two recommender models.

Critical weaknesses
  1. Defense-aware poisoning attacks have been studied by Koh et al. (2022). Lin et al. (2020) and Qiao et al. (2025) also explored attacks that balance target promotion and detection evasion in recommender systems. This paper mainly combines known attack strategies and evaluation settings. The new contribution beyond these works is limited.

  2. The main experiments focus on a hand-designed gate with four statistics. Most comparisons with prior detectors use simplified proxies rather than their original methods. More importantly, Table 21 shows that a simple item-frequency-growth test can detect the Stage 6 attack that passes the proposed gate. This makes it hard to judge how serious the reported attacks are against stronger defenses.

  3. The payload bound in Eq. (6) follows directly from the item-prevalence statistic and the gate threshold. It only applies to the chosen gate and says nothing about target promotion. Also, Stage 6 minimizes the detection score rather than maximizing promotion under the admission constraint. Thus, the reported frontier may depend heavily on the attack design.

  4. The paper uses Computer Modern for its main text, rather than the Times-style font in the official ICLR 2027 template. This raises a concern about whether the required template was followed. The authors should clarify whether they changed the formatting settings.

Questions for authors
  1. Can the authors compare with stronger shilling attacks, such as AUSH and DLDA, under the same admission and promotion setting?

  2. Why are most defense baselines simplified proxies? Can the authors compare against full implementations of existing defenses under matched false-positive rates?

  3. Would the admission–payload–promotion frontier still hold if the attacker directly maximized target promotion while satisfying the admission constraint?

  4. Did the authors use the original ICLR 2027 style files without modification? Why does the main text use a different font from the official template?

7 comments
AuthorsOpenReview ↗

W1. Novelty relative to Koh et al. (2022), Lin et al. (2020) and Qiao et al. (2025).

We agree these lines of work exist. We cite them, and Sec. 2 states that we claim neither the gate nor the attack family as new. The delta is in what is varied and measured:

  • Koh et al. (2022) design stronger poisoning attacks that break a fixed sanitization defense of a classifier. We hold the attack simple and vary the defense's evaluation protocol: how benign calibration cohorts are formed, and whether malicious profiles arrive mixed with genuine users. The same gate with a 0.11% resampling FPR rejects 26/33 chronological genuine windows, and 16/16 when calibrated strictly on the past (new). Attacks it rejects at a 50% attacker share pass at 25%, and still promote the target. Koh et al. do not study either effect.
  • AUSH (Lin et al., 2020) and DLDA (Qiao et al., 2025) learn profile generators that trade promotion against the detectability of individual profiles. Our gate makes one decision per cohort, and our central quantities are cohort-level: an analytic payload ceiling, the promotion that admitted cohorts still achieve after retraining, and the protocol effects above.

We kept the attack closed-form on purpose. If a simple attack already exposes these failure modes, they are failures of the evaluation rather than products of attack sophistication. The revised Sec. 2 states this delta explicitly, and we have added the AUSH citation.

AuthorsOpenReview ↗

W2 / Q2. Hand-designed gate, proxy baselines, and the item-frequency-growth result (Table 23; Table 21 in the submission).

The gate. It is a standard Hotelling-type T2T^2 over four classical shilling signals. The high-payload conclusion does not depend on its design: diagonal-covariance, calibrated union-of-marginals, and all four leave-one-feature-out gates reject every payload-bearing Stage 0–5 construction (Tables 20–21).

Item-frequency growth. We agree this result matters, which is why we report it. Three points (now in Sec. 7.4):

  1. Combining the tests. Rejecting when either the joint gate or the growth test rejects catches all three cohorts in Table 23, including the Stage 6 evader, at an organic FPR of at most 0.29% (union bound).
  2. Its operating point is for one known item. The growth test was evaluated for the pre-specified target. Deployed, it must screen all 12,101 items, which requires a multiplicity correction. An attacker who knows the test faces a per-item carrier limit analogous to Eq. (6): the admissible payload falls, but the frontier logic is the same.
  3. Its FPR was also measured on randomly resampled cohorts. Whether it holds on chronological traffic is precisely the question our protocol raises. In this sense, Table 23 is an argument for evaluating every cohort-level detector under our protocol, the growth test included.

Proxies. RDMA-, PCA- and LoRec-style detectors score individual profiles, often rating-based ones, while our gate decides on a cohort of timestamp-ordered histories. A "matched FPR" across these units is not well-defined. App. I therefore labels each method as a proxy and reports profile-level rates and post-filtering-then-retraining results separately, rather than claiming superiority.

AuthorsOpenReview ↗

W3. Eq. (6) is simple and gate-specific; Stage 6 minimises the gate score rather than maximising promotion.

We agree on both facts and state them in Secs. 4–5. Eq. (6) is gate-specific, but it is attack-agnostic. It depends only on the fitted gate, so it bounds the carriers of any admitted cohort at 27.1 among 671, whoever produced the cohort. Its simplicity is the point: one line, evaluated before any search, predicts the empirical boundary (accepted through m=26m{=}26, none at 28 or 32). The payload side of the frontier therefore does not depend on our attack design. How Stage 6's objective affects the promotion side is the reviewer's Q3, answered next.

Reviewer_Eq9vOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper evaluates admission-time defenses against adaptive data poisoning on semantic-ID recommenders. The authors propose a seven-stage shilling attack framework that adjusts payload, profile length, item popularity, and transition structure to evade detection. Attack success is measured through both admission decisions and downstream promotion after retraining. Results show that a joint gate based on four statistics rejects high-payload attacks, while adaptive search finds effective lower-payload attacks that bypass admission. The study also exposes evaluation protocol weaknesses: calibration on random resampling fails on temporal traffic, and mixing malicious with genuine profiles reduces detectability.

Critical strengths
  1. Meaningful problem formulation. Jointly evaluating admission and downstream poisoning provides a more realistic security assessment than detection metrics alone.

  2. Systematic adaptive evaluation. The staged framework clearly demonstrates attacker adaptation, with admission-payload-promotion analysis offering useful insights into detectability-effectiveness trade-offs.

  3. Comprehensive diagnostics. Multiple architectures, datasets, gate formulations, and temporal evaluation reveal significant discrepancies between random and chronological traffic false-positive rates.

Critical weaknesses
  1. Limited novelty. The attack combines existing shilling strategies, and the joint gate uses standard Mahalanobis distance. The paper positions itself as an evaluation framework, and the calibrated union-of-marginals gate matches joint gate decisions, limiting evidence for sophisticated modeling.

  2. Simplified threat model. Experiments assume fixed-size cohorts with all-or-nothing admission, oracle access to gate parameters, and specific target placement, raising generalization concerns.

  3. Limited generalization evidence. Detailed experiments focus on Amazon Beauty with minimal training seeds. High variability across targets and datasets weakens broader conclusions about admitted attack effectiveness.

  4. Incomplete deployment evaluation. High rejection rates on temporal cohorts raise usability concerns, but temporal recalibration isn't evaluated. Mixed-cohort and temporal-shift experiments remain largely separate.

Questions for authors
  1. Temporal calibration: How would results change using strictly historical reference statistics? Can mixed cohorts be evaluated under chronological traffic rather than separately?

  2. Baseline strength: Would staged evasion results differ with consistently calibrated single-feature gates? Can the joint gate be compared against the item-frequency-growth detector under temporal shift?

  3. Attack realism: How sensitive are findings to target placement and reduced attacker knowledge? Would direct downstream optimization yield more harmful accepted attacks than gate score minimization?

  4. Generalizability: Can more cohorts, seeds, and target popularities be tested? How stable is the promotion frontier across different attack realizations?

7 comments
AuthorsOpenReview ↗

W1. Limited novelty; the union-of-marginals gate matches the joint gate.

We agree that the attack components and the Mahalanobis gate are standard, and Sec. 2 says so. Our contribution is the evaluation protocol. It shows that three common evaluation choices each change the security conclusion:

  • Fixed attacks are all rejected, while adaptive search admits payloads up to the analytic ceiling.
  • A 0.11% resampling FPR becomes 26/33 rejected genuine windows, and 16/16 under strictly historical calibration (new).
  • Attacks rejected 95–99.5% of the time at a 50% attacker share pass almost always at 25% or less, and still promote the target.

We report the union-of-marginals equivalence ourselves (Table 21) because we read it as a finding, not a gap. Against these attacks, correlation modelling adds nothing; protection comes from calibrated, redundant statistics, chiefly item prevalence. This is useful negative evidence for a practitioner who might adopt a joint model for its apparent sophistication. The revised Sec. 2 also states the delta from defense-aware poisoning (Koh et al., 2022) and from learned shilling attacks (AUSH, Lin et al. 2022, DLDA).

AuthorsOpenReview ↗

W2 / Q3. Simplified threat model: fixed cohorts, all-or-nothing admission, oracle access, target placement.

  • Cohort size. Varied over N∈{335,671,1342}N \in \{335, 671, 1342\} (Table 26), plus windows of 200 and 100 (App. K.2). Stages 4–5 are rejected at every size, and Stage 6 rejection grows with NN (0%, 10%, 70%).
  • All-or-nothing. This is the defender's rule in the cohort setting. The mixed-cohort study applies it to the NN most recent arrivals with unknown membership, which is the most deployment-like configuration we test, and quantifies its cost: 319–334 genuine users are withheld per window at a 50% attacker share. Per-profile rules are a different detection unit, evaluated in App. I together with post-filtering and retraining.
  • Reduced attacker knowledge. Stages 0–4 use only aggregate catalog statistics. The payload ceiling is a property of the gate, not of attacker knowledge. Once a blend near w=0.9w{=}0.9 is known, fresh m=16m{=}16 constructions pass in 90% of cases with no further gate queries (Table 26). The accepted carriers also stay admitted when embedded among genuine rather than synthetic histories (D2/τD^2/\tau = 0.04 and 0.75; App. K.2).
  • Target placement. Admission is almost placement-invariant by construction. ϕ0\phi_0 (prevalence), ϕ1\phi_1 (length) and ϕ2\phi_2 (mean rarity over all positions) do not depend on where t⋆t^\star sits; only ϕ3\phi_3 changes, through the at most two adjacent pairs that touch the target. Placement can affect promotion, and the revised Sec. 8 lists the fixed final-supervised position as a limitation.

Direct downstream optimisation (Q3). Eq. (6) depends only on the fitted gate, so it holds for any admitted cohort whatever the attacker optimises: at most 27.1 carriers among 671. A promotion-maximising attacker can raise promotion per carrier, but cannot add carriers. Our data indicate that payload, not construction, drives promotion. With the generator fixed, promotion goes from 0.00% to 0.25% to 9.5% as mm goes 7 → 16 → 67; with m=16m{=}16 fixed, varying the blend gives only 0.13–0.25% (Table 12). Our retrained m=24m{=}24 cohort is already at 89% of the ceiling. We therefore expect direct optimisation to raise the frontier somewhat, not to remove it. Quantifying this requires a bilevel attack through retraining, which we flag as the main open extension (Sec. 8).

AuthorsOpenReview ↗

W3 / Q4. Generalisation: cohorts, seeds, target popularity, stability of the frontier.

New: three training seeds for every previously single-seed accepted cell (Table 17).

CellModelSeed 13 seeds
m=16m{=}16, rareSASRec4.693.46±1.273.46\pm1.27
m=24m{=}24, rareSASRec6.576.08±0.926.08\pm0.92
m=16m{=}16, mediumSASRec4.724.32±0.714.32\pm0.71
m=16m{=}16, mediumDiffGRM3.863.33±1.163.33\pm1.16

Every cell stays well above zero, and seed 1 lies within one sd of the mean.

Already in the paper:

  • 20–25 independent constructions per headline condition (Table 8).
  • 150 constructions in the refined search (Table 10), with acceptance falling gradually: 19, 16, 10, 3, 0, 0 of 25.
  • Three target frequencies (5, 48, 403; Tables 14–15) and a second rare target (Table 13).
  • A three-seed replication on Toys and Games (Table 18), and Sports and Outdoors.

The shape of the frontier (rejection at the working payload, acceptance below the ceiling, non-zero promotion after admission) is stable across these. Its magnitude depends on target and recommender, and we report that variation rather than average it away.

Reviewer_b5uwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the robustness of admission-time defenses against adaptive data poisoning attacks in semantic-ID generative recommendation. The authors develop a staged family of shilling attacks that progressively adapt profile length, item popularity, transition patterns, and target-item payload to evade statistical admission gates. The experiments further reveal that defenses calibrated on randomly sampled benign cohorts can suffer substantial false positives under temporally ordered traffic and become ineffective when malicious profiles are mixed with genuine users. Overall, the paper provides a meaningful empirical analysis of the limitations of cohort-level admission defenses, although its semantic-ID-specific contributions and generalizability to more realistic defense settings require further investigation.

Critical strengths
  1. The paper addresses a practically relevant problem that is often overlooked in traditional security evaluations and presents a relatively comprehensive end-to-end evaluation framework.
  2. The experimental analysis is relatively systematic and provides sufficient evidence to support the paper's main findings.
Critical weaknesses
  1. The paper lacks sufficient novelty specific to semantic-ID-based generative recommendation, as its technical contributions primarily lie in the evaluation framework rather than in the attack or defense methods themselves.
  2. The defense evaluation remains considerably different from realistic deployment scenarios, limiting the generalizability of the findings.
Questions for authors
  1. Could the authors further clarify which properties of semantic-ID representations or autoregressive generation contribute specifically to the observed vulnerabilities?
  2. Could the authors explain why this detector is effective and further evaluate whether the adaptive attacks remain successful against alternative or combined admission defenses?
  3. Could the authors evaluate a strictly chronological admission protocol in which all reference statistics and calibration thresholds are estimated only from historical data?
6 comments
AuthorsOpenReview ↗

W1 / Q1. Which properties of semantic IDs or autoregressive generation cause the vulnerability?

We agree this needed to be stated directly. Our own evidence answers it, and the answer is that the admission–payload–promotion frontier is not caused by semantic-ID generation.

  • Admission does not involve the recommender. The payload ceiling of Eq. (6), m/N≤μ0+τΣ00m/N \le \mu_0 + \sqrt{\tau\Sigma_{00}} (27.1 carriers among 671), depends only on the fitted gate. Which payloads are admissible is therefore the same for any downstream model.
  • Promotion by admitted cohorts is not specific to semantic IDs. The same accepted Beauty cohorts promote the target to 4.69% (m=16m{=}16) and 6.57% (m=24m{=}24) of eligible users under the non-generative SASRec, versus 0.38% and 1.44% under DiffGRM (Table 16). Over three training seeds (new, Table 17), SASRec gives 3.46±1.27%3.46\pm1.27\% and 6.08±0.92%6.08\pm0.92\%.
  • Which recommender is more exposed depends on the catalog (App. G). On Toys and Games, DiffGRM is promoted more by the rejected cohorts, while SASRec is promoted more by the accepted one.

We therefore do not claim a semantic-ID-specific mechanism. Semantic-ID generative recommenders are the setting where we instantiate the protocol: they are retrained on logged interaction sequences, and their admission boundary had not been evaluated end-to-end. The revised Sec. 2 now states this explicitly.

On novelty: as stated in Sec. 2, we claim neither the gate nor the attack family as new. What the protocol adds is evidence that three standard evaluation choices each change the conclusion qualitatively:

  1. Fixed vs. adaptive attacker. Every m=67m{=}67 construction is rejected, yet search admits payloads through m=26m{=}26, matching the analytic ceiling of 27.1.
  2. Random vs. chronological benign cohorts. A 0.11% false-positive rate under resampling becomes 26/33 rejected genuine windows, and 16/16 under the new historical calibration below.
  3. Attacker-only vs. mixed cohorts. Rejection falls from 95–99.5% at a 50% attacker share to 1.5% at 25%, and admitted windows still promote the target (1.35% and 0.96%).

These are properties of how cohort gates are evaluated, which is why we deliberately kept the attack simple.

AuthorsOpenReview ↗

W2. Distance from realistic deployment.

We agree the setting is controlled, and we tested the main simplifications:

  • Cohort size. N∈{335,671,1342}N \in \{335, 671, 1342\} (Table 26), plus admission windows of 200 and 100 (App. K.2).
  • Unknown membership. The mixed-cohort study scores the NN most recent arrivals without knowing which are attackers. It shows that the gate also withholds 319–334 genuine users per window at a 50% attacker share.
  • Oracle access. Once a blend near w=0.9w{=}0.9 is known, fresh m=16m{=}16 constructions pass in 90% of cases with no further gate queries (Table 26).
  • Chronological traffic. The experiment above.

The remaining gaps (rolling recalibration, accumulation across periods, training-time defenses) are listed in Sec. 8.

We hope the new chronological experiment and the restated scope address the reviewer's concerns. If they do, we would be grateful if the reviewer would consider revisiting the score, and we are happy to answer further questions.

Thank you for the authors' response. However, I find the response insufficiently convincing, as it does not adequately address my main concerns. Therefore, I believe my current assessment remains reasonable and will maintain my original score.

Comments on the paper

2 comments
Area_Chair_gvqqOpenReview ↗

Reminder: Please Review Author Rebuttals Before the Discussion Period Ends

Dear Reviewers,

As we are approaching the end of the discussion period, this is a friendly reminder to check the author rebuttals for your assigned papers. If the authors have addressed your questions or responded to your concerns, please take a moment to review their responses and consider whether any updates to your current evaluations are warranted.

Thank you for your time and effort!

Best regards,

Area Chair

AuthorsOpenReview ↗

Consolidated Response to the Reviewers

We sincerely thank all reviewers for their careful evaluation and constructive feedback. We summarize the principal concerns and our responses below.

1. Reviewer b5uw

Concerns raised: The reviewer questioned the semantic-ID-specific contribution, detector effectiveness against adaptive attacks and alternative defenses, and evaluation realism. The reviewer also requested strictly chronological calibration using historical data.

How we addressed them: We clarified that we do not claim a semantic-ID-specific vulnerability. Admission and the payload bound depend on the statistical gate, not the downstream recommender. Results on SASRec and DiffGRM show that promotion varies across models and catalogs. Our contribution is the end-to-end evaluation of admission, payload, and promotion, rather than a new attack or detector.

Using the earliest 50% of genuine users for calibration, the gate rejected all 16 subsequent genuine windows, compared with 11/16 under static calibration. Thus, historical calibration alone does not resolve temporal non-stationarity. We also showed that item prevalence captures target concentration and that calibrated marginal and joint gates make identical decisions. A combined joint-gate and item-frequency-growth defense catches all three attack cohorts, including Stage 6, with an organic false-positive rate of at most 0.29%, subject to multiple-testing and adaptive-evasion caveats.

2. Reviewer Eq9v

Concerns raised: The reviewer questioned novelty, threat-model simplicity, generalization, temporal recalibration, mixed-cohort evaluation, and potential misuse of the attack methodology.

How we addressed them: We clarified that the attack components and Mahalanobis gate are standard; our contribution is demonstrating how evaluation choices alter security conclusions. A 0.11% false-positive rate under random resampling becomes rejection of 26/33 genuine chronological windows and 16/16 under historical calibration. Attacks rejected at a 50% attacker share can pass at 25% or less while still promoting the target.

We varied cohort sizes, evaluated mixed cohorts with unknown attacker membership, and tested reduced attacker knowledge. At a 50% attacker share, the gate also withholds 319–334 genuine users per window. Additional three-seed results support the qualitative stability of the promotion frontier, although its magnitude varies across targets and recommenders.

Chronological mixed-cohort evaluation is currently uninformative because the gate already rejects genuine-only chronological windows. Rolling recalibration and its mixed-cohort evaluation remain important next steps. We also expanded the ethics statement, clarified the diagnostic purpose, and committed to releasing research code without production-platform data-collection tools.

3. Reviewer UyPC

Concerns raised: The reviewer questioned novelty relative to prior shilling attacks, defense-baseline strength, the gate-specific payload bound, direct promotion optimization, and ICLR template compliance.

How we addressed them: We clarified that, unlike prior work proposing profile generators, we study cohort-level admission and the effects of calibration, temporal ordering, and attacker–genuine mixing. Profile-level detectors and cohort-level gates use different detection units; therefore, matched false-positive comparisons are not directly meaningful. We label these comparisons as proxies and report them separately.

A combined joint-gate and item-frequency-growth defense catches all three evaluated cohorts, including Stage 6. However, the growth test was evaluated for a pre-specified target; screening all catalog items requires multiplicity correction, and performance under temporal shift remains untested.

The payload bound in Eq. (6) is gate-specific but attack-agnostic: any admitted cohort is limited to 27.1 target carriers among 671 profiles, regardless of the attacker's objective. Direct promotion optimization could increase promotion per carrier, but our results suggest payload has a stronger effect than construction. It may raise the promotion level without removing the payload constraint; evaluating this through bilevel optimization remains future work.

Finally, we confirmed that the official ICLR 2027 style files were unchanged. The font discrepancy resulted from omitting the template's times option. We corrected this without modifying margins, spacing, or font size.

The revised manuscript clarifies our contribution, strengthens chronological and mixed-cohort evaluation, adds multi-seed evidence, and evaluates combined admission defenses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.