acceptodds
Under review as a conference paper at ICLR 2027

POISONING THE FIT SET: END-TO-END ATTACKS ON OPQ-BASED SEMANTIC-ID RECOMMENDATION

Abstract

Semantic-ID generative recommenders often rely on an OPQ tokenizer fitted on a subset of catalog items and then used to assign discrete identities to the full catalog. We identify this fit set as an overlooked trust boundary and ask whether injected catalog items can alter the Semantic IDs of legitimate, untouched items and affect downstream recommendation. Clean refitting is exactly reproducible, with Adjusted Rand Index (ARI) 1.000, allowing structural changes to be attributed to fit-set modification rather than tokenizer randomness. Across three OPQ-based recommendation pipelines and three Amazon catalogs, fit-set poisoning consistently causes substantial structural corruption, with mean per-digit ARI falling to 0.14–0.26. The effect also reproduces across ten independently published text encoders and is not confined to idealized embedding-space injection: text-derived poison passed through the same tokenizer-fitting pipeline reaches a comparable corruption regime. However, tokenizer corruption does not translate monotonically into recommendation harm. Across nine recommender–catalog settings trained to convergence, only three show significant downstream changes, and their directions differ across architectures, while substantial structural corruption remains present throughout. Matched benign-growth controls further show that ordinary catalog additions can induce structural drift comparable to malicious insertion, so drift magnitude alone is not a reliable attack signature. Six published anomaly-detection baselines fail to restore the clean tokenizer structure. A fit-aware trusted-reference admission gate strongly separates the non-adaptive attacks from held-out legitimate additions across three catalogs, but an adaptive adjacent-mimicry attack exposes an evasion region. Frozen-rotation warm-started refitting complements screening by constraining admitted updates, maintaining ARI at or above 0.98 across tested single-refit attack severities and geometries, while sequential drip-feed updates reveal gradual erosion under repeated refitting. Code will be provided after acceptance.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Generative Recommendation, Semantic IDs, Data Poisoning, Tokenizer Security, Recommender System Security

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_ZTA9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies fit-set poisoning in OPQ-based Semantic-ID recommender systems. By injecting items into the data used to fit the OPQ tokenizer, an attacker can substantially change the Semantic IDs of legitimate items. The paper evaluates this effect across multiple recommenders, datasets, encoders, and attack settings, and further studies downstream recommendation impact and possible defenses.

Critical strengths
  • The paper identifies the tokenizer fit set as an underexplored security boundary in generative recommendation systems.

  • The study covers multiple recommenders, datasets, text encoders, attack strategies, and realistic text-based poisoning.

Critical weaknesses
  • The threat model does not fully explain how the attacker can actually poison the fit set. This is probably my biggest concern. In the evaluated pipelines, the fit set consists of items that appear in the training interaction data. However, the attacker is assumed to control only injected items and cannot manipulate user interaction histories. In other words, simply uploading a malicious product is not the same as getting that product into the fit set.

  • The paper demonstrates strong tokenizer corruption, but not a reliable end-to-end recommendation attack. The main empirical result is that poisoning can substantially change the Semantic-ID structure. However, this does not consistently translate into recommendation degradation. Among the nine recommender–dataset settings, only three show significant downstream changes, and one of them actually improves recommendation quality. Increasing the poisoning budget also does not consistently increase the damage. This makes it unclear what the practical security impact is.

  • A large part of the observed effect may reflect inherent OPQ instability rather than a specifically adversarial effect. The benign-growth experiments show that simply adding normal catalog items can produce structural drift comparable to malicious poisoning. The paper itself concludes that this sensitivity is a broader property of OPQ refitting rather than a uniquely malicious signature. This weakens the interpretation of the results as an attack.

  • The attack also has limited controllability. Although the proposed poisoning can cause large global changes in the Semantic-ID structure, even the “cluster-targeted” attack does not damage the intended cluster more than the rest of the catalog. The authors’ own localization analysis finds no significant targeting effect. As a result, the attacker can disrupt the tokenizer, but cannot reliably control which items are affected or what downstream recommendation behavior will result. This makes it unclear what concrete attacker-chosen objective the proposed attack can actually achieve.

Questions for authors

see Weaknesses

5 comments

Q1. Threat Model and Practical Fit-Set Admission

We thank the reviewer for the opportunity to clarify how attacker-controlled items enter the tokenizer fit set. We clarify the assumed fit-set membership in our original experiments and provide new end-to-end validation showing how attacker-controlled items can enter the fit set through attacker-generated interactions (Section 3.1).

Threat model and admission mechanism.

In the studied pipelines, the fit set MM comprises items appearing in training interactions. The original experiments assume that attacker-controlled representations can be inserted directly into the OPQ fit set after clean PCA, isolating tokenizer-level effects. In the end-to-end setting, the attacker creates its own accounts and interacts with its newly listed items, without modifying legitimate users' histories. These interactions satisfy the pipeline's admission rule, yielding M′=M∪PM' = M \cup P.

End-to-end experimental validation.

On DiffGRM/Beauty, we introduce five attacker-controlled accounts, each interacting with all kk injected items. Every injected item consequently receives five interactions, satisfying the dataset's 5-core filtering requirement. Each attacker account also satisfies this requirement. The injected items enter the training sequences, recommender vocabulary, and tokenizer fit set. PCA and OPQ are then refitted, followed by recommender retraining (Appendix A.6).

The five accounts represent 0.02% of users, contributing 0.1% and 0.3% of real interaction volume at k=40k=40 and k=121k=121, respectively.

Experimental results (Table 13).

Added itemsAdmittedTuple ARIExclusive codewords/digit
HotFlip (k=40k=40)40/400.4701.50
Real listings (k=40k=40)40/400.4420.00
Text mimicry (k=121k=121)121/1210.4650.50
Real listings (k=121k=121)121/1210.4580.00

All injected items successfully enter the fit set. Matched real listings follow the identical admission procedure, holding the interaction pattern fixed while varying the added items. Both attacks induce substantial legitimate-item Semantic-ID reassignment and occupy codewords unused by legitimate items, whereas matched real listings occupy no exclusive codewords at these budgets.

We further verify complete fit-set admission at attack budgets of 5%, 10%, and 25% of the original catalog size (Table 13; Appendix B.9), with structural effects evaluated at these larger budgets.

In summary, the end-to-end experiment establishes a feasible route for satisfying the dataset's fit-set admission rule through attacker-generated interactions, without modifying legitimate user histories.

Q2. End-to-End Recommendation Impact

We thank the reviewer for raising the question of practical security impact. In the revised manuscript, we clarify the distinction between tokenizer-level disruption and downstream recommendation effects, supported by additional matched-benign comparisons, clean-refit controls, and statistical analysis (Sections 5.2–5.3; Appendix C).

Downstream impact.

Across nine recommender–dataset settings, three exhibit nominally significant NDCG@10 changes: −11.5% (DiffGRM/Sports), −4.0% (TIGER-style/Beauty), and +5.2% (TIGER-style/Toys). None survives Holm correction across the nine comparisons (minimum adjusted p=.163; Table 14).

The budget analysis reveals a more consistent negative direction for leverage poisoning, which reduces NDCG@10 by 10.5%, 11.6%, and 5.7% at 5%, 10%, and 25% budgets, respectively. Although the magnitude is non-monotonic, degradation occurs at all three budgets (Table 17).

Attack-specific effects versus ordinary refitting.

At k=121, leverage poisoning produces a 17.7% NDCG@10 reduction relative to the pre-update tokenizer (nominal p=.005, Holm p=.051), compared with 10.9% under matched benign growth. The direct attack-minus-benign difference is −7.6% [−22.9%, +7.7%] (Holm p=.662; Tables 2 and 16). These matched comparisons quantify the additional effect relative to ordinary catalog growth, although the difference is not statistically significant.

End-to-end validation.

We additionally evaluate attacker-controlled listings admitted through five attacker accounts, with PCA and OPQ refitting followed by recommender retraining. Relative to matched real listings, 40 HotFlip items produce −3.1% NDCG@10 (95% CI [−20.4%, +14.1%]), while 121 text-mimicry items produce +3.6% (95% CI [−25.9%, +33.0%]; Table 18). These experiments establish that the poisoning mechanism operates through the full pipeline, while the observed downstream differences remain statistically inconclusive.

Practical security implications.

Importantly, the attack's effects extend beyond Semantic-ID reassignment. At matched update sizes k=121 and k=605, poisoned items exclusively occupy 0.50–5.25 codewords per digit, versus at most 0.25 under benign growth. Random and cluster-targeted poisoning also increase legitimate-item reconstruction error to 0.6266–0.6312, compared with 0.6236–0.6256 for benign additions (Table 1).

Our findings establish measurable adversarial effects on OPQ quantizer capacity and reconstruction quality, while distinguishing them from downstream ranking changes that are not statistically established as poisoning-specific. The security contribution therefore includes identifying this tokenizer-level vulnerability and evaluating protections at the refitting boundary.

Q3. Adversarial Poisoning versus Benign OPQ Instability

We thank the reviewer for highlighting the need to distinguish adversarial poisoning from inherent OPQ refitting instability. Our revised analysis explicitly separates Semantic-ID reassignment caused by ordinary refitting from additional effects induced by attacker-controlled items (Sections 5.1–5.2).

Distinguishing poisoning from benign instability.

Clean refits caused by changes in FAISS seeds, item order, or fit-set membership already produce substantial Semantic-ID reassignment (tuple ARI approximately 0.38–0.49). Similarly, poisoning and matched benign additions exhibit overlapping ARI ranges. We therefore use matched additions to examine whether adversarial items produce effects beyond this inherent instability.

Evidence of attack-specific effects.

Table 1 provides a direct comparison at identical update sizes:

BudgetExclusive codewords/digit (benign)Exclusive codewords/digit (poison)Reconstruction error (benign)Reconstruction error (random/cluster poison)
k=12100.50–2.250.6236–0.62510.6266–0.6281
k=6050–0.252.50–5.250.6247–0.62560.6281–0.6312

The exclusive-codeword ranges cover random, cluster-targeted, and leverage poisoning; reconstruction-error ranges cover random and cluster-targeted poisoning.

These results reveal two distinctions. First, poisoned items occupy codewords no longer used by legitimate items, whereas matched benign additions exhibit negligible exclusive occupancy. Second, random and cluster-targeted poisoning consistently produce higher legitimate-item reconstruction errors than matched benign growth at both budgets. Thus, the observed adversarial effects extend beyond Semantic-ID reassignment alone.

End-to-end validation.

We additionally verify this distinction using attacker-controlled listings admitted through the interaction-based mechanism. With 40 HotFlip items, exclusive codeword occupancy reaches 1.50 per digit, versus 0 for 40 matched real listings. With 121 text-mimicry items, the corresponding values are 0.50 versus 0 (Table 13). These experiments demonstrate that the structural differences can also arise when injected items enter through the actual fit-set admission process.

OPQ instability constitutes the attack surface, but the measured quantizer-capacity capture and increased reconstruction error distinguish the evaluated poisoning attacks from ordinary benign catalog growth. This distinction is now explicit in Sections 5.1–5.2 and Table 1 of the revised manuscript.

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

The paper studies poisoning of the fit set used to train an OPQ tokenizer for Semantic IDs in generative recommendation: an attacker injects items into the tokenizer fit set, OPQ is refitted, and the resulting Semantic IDs of legitimate items are compared with the clean fit using permutation-invariant metrics such as ARI and NMI.

The experiments cover three Amazon datasets, several OPQ-based recommendation pipelines, multiple attack strategies and budgets, ten text encoders, and two text-based attacks. At the tokenizer level, the effect is large: clean refits give ARI 1.000, while poisoned fits produce substantial structural drift. However, this drift does not translate consistently into recommendation degradation: across the nine recommender/dataset combinations at 25% poisoning, only three NDCG@10 changes are nominally significant, with mixed signs, and the budget sweep is non-monotonic. A particularly important result is that benign catalog growth produces structural and downstream changes comparable to malicious insertion. This suggests that the observed phenomenon may reflect a broader instability of OPQ refitting rather than a uniquely adversarial effect.

The paper also evaluates defenses: standard anomaly detectors do not restore the clean tokenizer structure. Furthermore, a trusted-reference gate works well against non-adaptive attacks but can be evaded and frozen-rotation warm-started refitting substantially limits structural drift, although its protection weakens under repeated updates.

Critical strengths

The first strength is the breadth of the experimental study: the paper evaluates several OPQ configurations, three catalogs, multiple attack budgets, ten text encoders, text-realizable attacks, and full downstream recommendation training. The use of permutation-invariant measures rather than raw SID label differences is also appropriate.

I also appreciate that the authors report results that complicate their own initial story: the cluster-targeting issue is documented instead of being hidden, and the paper reports mixed and null downstream effects rather than selecting only cases that support degradation.

The second major strength is the benign-growth analysis. For example, with k=6, adding legitimate held-out items gives tuple ARI 0.433, while the three attack constructions give 0.438–0.453. This shows that large SID movement is not itself evidence of malicious activity and suggests that systems that repeatedly refit OPQ may have a more general stability problem.

Critical weaknesses
  1. My main concern is that the paper does not currently distinguish poisoning from generic OPQ refit instability. Indeed, the paper interprets the observed SID changes as structural corruption caused by poisoning. However, several results seem equally compatible with the possibility that OPQ is simply highly unstable when refitted on these representations after almost any change to the fitting data. There is no clear dose-response in the structural metric. On Beauty, for example, mean per-digit ARI is 0.156, 0.172, 0.150, and 0.155 at 1%, 5%, 10%, and 25% poisoning. The tuple-level values in the appendix also show only a weak relationship with budget. A 40-item HotFlip construction, corresponding to only 0.33% of the catalog, already reaches tuple ARI 0.4357, in the same broad regime as much larger injected sets. More importantly, legitimate perturbations can produce changes of essentially the same magnitude. Adding six held-out legitimate items produces structural drift comparable to the three attacks. The cluster-targeted construction also does not preferentially damage the targeted region: the localization test gives (p=0.851). Finally, the adjacent-mimicry batches that evade the gate do not cause more downstream harm than matched legitimate additions. Taken together, it is unclear what additional effect the attacker achieves beyond the instability already caused by ordinary catalog changes.

  2. The practical attack story and some of the evaluation choices are not yet convincing to me. There is a gap in the threat model that I could not resolve from the manuscript. Section 3 states that the tokenizer fit set consists of items appearing in the training interactions, but the attacker is assumed not to be able to modify user interaction histories. I therefore do not understand how attacker-created items enter M in the first place, particularly at the larger budgets. If attacker-generated interactions are necessary, I think they should be included explicitly in the threat model. More generally, the attack objective is untargeted structural change. The paper does not show that the attacker can reliably promote an item, demote an item, damage a specific category, or otherwise achieve an objective that benign catalog growth does not already produce. Given that structural drift itself is not attack-specific and downstream effects are inconsistent, I think the distinction between a security vulnerability and a tokenizer stability problem needs to be made much more carefully. I also have some concerns about the frozen-rotation defense. At a high level, the mechanism succeeds by limiting how much the tokenizer is allowed to change. The appendix includes a max-iteration sweep containing max_iter=0, but there is no corresponding utility analysis comparing no refitting with the proposed one-iteration update.

  3. The manuscript is difficult to read, and the abstract does not clearly frame the problem before introducing the methodology. I found the paper difficult to follow even after careful reading. The issue is not only density, but also the way the argument is structured. Important distinctions between generic OPQ instability, adversarial poisoning, structural SID drift, and downstream recommendation harm only become clear after reading several later sections and appendices. Metrics also change across tables, seed counts vary, and some experimental details that are important for interpreting the main claims are introduced only later. The abstract is a good example of this problem. It moves very quickly into the fit-set construction, ARI values, downstream significance counts, benign-growth controls, and the proposed defenses. More broadly, the manuscript often introduces technical details before the reader has a clear conceptual picture of what question is being answered. The later benign-growth experiments also substantially change the interpretation of the initial poisoning framing, but this shift is not reflected clearly enough in the narrative. Finally, I found the bibliography unusually problematic. Several entries include explanatory annotations rather than standard citation information, and some of these descriptions appear inconsistent with the cited works or with the final version of the manuscript. Since at least one such characterization is used to motivate the proposed attack surface, I think the Related Work and bibliography require a careful source-by-source revision rather than only copy-editing.

Questions for authors

1. Determinism / refit-instability null.

Could the authors clarify exactly what randomness was varied in the ARI = 1.000 clean-refit control? In particular, did this include the FAISS clustering seed, OPQ initialization, PCA, and data ordering? I would find it important to know how stable clean fits remain under genuinely different FAISS seeds, a random permutation of the item order, and a minimal benign perturbation such as adding or removing one legitimate item. If these changes already produce drift comparable to the poisoning experiments, I would interpret the main contribution primarily as evidence of OPQ refit instability rather than an attack-specific poisoning effect.

2. How do injected items enter the fit set?

The paper states that M consists of items appearing in the training interactions, while the attacker is not allowed to modify interaction histories. How, concretely, do the injected items become part of M? Do they receive attacker-generated interactions? Do they need to pass the 5-core filtering procedure? Are they also added to the recommender's item vocabulary and training set?

3. What is the downstream noise floor?

Could the authors report the variation in NDCG between clean runs whose tokenizer fits differ only in the sources of randomness discussed above? I also noticed that the held-out-pool clean baseline in Table 15 has NDCG@10 = 0.0296, whereas the regular clean DiffGRM/Beauty model in Table 2 has 0.0278. Is this difference caused by a different experimental protocol, or by the changed fit set? The gap is already comparable in magnitude to several reported attack effects. Finally, do the three nominally significant Table 2 comparisons remain significant after an explicit multiple-testing correction?

8 comments

Q1. Determinism and OPQ Refit Instability

We thank the reviewer for highlighting the importance of distinguishing reproducibility from stability under refitting. We have expanded our clean-refit experiments to explicitly examine the sources of OPQ instability and distinguish them from poisoning-specific effects (Sections 5.1–5.2).

Clarification of the ARI=1.000 control.

The original control used identical fit-set inputs, item ordering, PCA configuration, and FAISS training seeds. The default OPQ rotation initialization was also fixed. Thus, ARI=1.000 establishes exact reproducibility under identical conditions, rather than stability under different fitting conditions.

Additional clean-refit experiments.

We now independently vary FAISS k-means seeds, item ordering, PCA solver, and legitimate fit-set membership (Table 1).

Clean-refit conditionTuple ARI
Identical inputs and seeds1.000
Different FAISS k-means seeds0.410–0.477
Different rotation-training seeds only0.422–0.491
Different final-quantizer seeds only0.565–0.597
Permuted item order0.396–0.452
One legitimate item removed0.433–0.490
One legitimate item added0.377–0.427
Randomized PCA solver0.434–0.488

These results establish that even minimal benign perturbations can substantially reassign Semantic IDs. They also identify rotation training as the dominant source of instability (Section 5.1; Appendix B.2).

Distinguishing poisoning from ordinary instability.

Semantic-ID reassignment alone is not evidence of attack-specific effects. At matched update sizes, injected items occupy 0.50–2.25 codewords per digit exclusively at k=121 (benign: 0) and 2.50–5.25 at k=605 (benign: 0–0.25). Random and cluster-targeted poisoning also produce higher legitimate-item reconstruction errors than matched benign growth in the evaluated conditions (Table 1; Section 5.2).

Q2. How Injected Items Enter the Fit Set

We thank the reviewer for highlighting the need to clarify how attacker-controlled items enter the tokenizer fit set. We have clarified the threat model and added end-to-end experiments that explicitly incorporate attacker-generated interactions (Section 3.1; Appendix A.6).

How injected items enter the fit set.

In the studied pipelines, the fit set comprises items appearing in training interactions. The attacker cannot modify legitimate users' interaction histories but can create its own accounts, list new items, and interact with those items. These attacker-generated interactions allow injected items to enter the fit set through the pipeline's normal admission mechanism.

We distinguish two experimental settings:

  • Assumed fit-set membership: Injected representations are added directly to the OPQ fit set after clean PCA, without attacker interactions. The recommender is retrained using the modified Semantic IDs while retaining the original training interactions. This isolates tokenizer-level effects.

  • End-to-end admission: Attacker-controlled items enter through generated interactions, join the recommender vocabulary and training sequences, and participate in PCA and OPQ refitting followed by recommender retraining.

Experimental verification.

On DiffGRM/Beauty, we introduce five attacker-controlled accounts, each interacting with every injected item. Consequently, each item receives five interactions, satisfying the dataset's 5-core filtering requirement. Each attacker account also satisfies this requirement.

All 40/40 HotFlip items and 121/121 text-mimicry items successfully enter the fit set, recommender vocabulary, and training sequences (Table 13; Appendix A.6).

The five accounts constitute 0.02% of users, contributing 0.1% and 0.3% of real interaction volume at k=40 and k=121, respectively.

We additionally confirm complete fit-set admission at 5%, 10%, and 25% catalog budgets, where we evaluate structural effects without downstream recommender retraining (Table 13).

Q3. Downstream Noise Floor and Statistical Significance

We thank the reviewer for highlighting the importance of clean-tokenizer variability in interpreting downstream recommendation changes. We have added clean-refit controls, clarified the baseline discrepancy, and applied multiple-testing corrections (Section 5.3; Appendix C).

1. Downstream variation under clean refitting.

On DiffGRM/Beauty, we evaluate three alternative FAISS seeds and three random permutations of item order, each using the same three paired recommender seeds. Relative to the default clean tokenizer, the NDCG@10 changes are:

Clean-refit variationNDCG@10 changes (%)
Three alternative FAISS seeds+6.5, −5.6, +3.7
Three item-order permutations−5.1, +6.7, −12.4

These results demonstrate that ordinary tokenizer refitting can change downstream recommendation quality substantially, even without adversarial items (Table 15; Appendix C.2).

2. Clarification of the clean-baseline discrepancy.

The difference between NDCG@10=0.0278 and 0.0296 arises from different tokenizer fit sets, not different recommender training protocols.

The default clean tokenizer is fitted on the full training fit set and achieves 0.0278 over five recommender seeds (0.0279 over the three seeds used for matched comparisons). The alternative clean tokenizer excludes a 700-item held-out pool and achieves 0.0296 over the same three seeds, corresponding to a +5.9% paired difference (95% CI [+2.8%, +9.1%]).

The original Table 15 matched-growth analysis is now reported in Table 16, while the new Table 15 explicitly documents clean-tokenizer variability (Appendix C.2–C.3).

3. Multiple-testing correction.

The three nominally significant results in original Table 2 (now Table 14) are:

Recommender / DatasetNDCG@10 changeNominal pHolm pBH q
DiffGRM / Sports−11.5%.028.198.085
TIGER-style / Beauty−4.0%.020.163.085
TIGER-style / Toys+5.2%.018.163.085

After correcting for nine comparisons, none of the three nominally significant results survives Holm correction (minimum adjusted p=.163; BH q=.085; Table 14). These results characterize the current downstream statistical evidence, while the poisoning-specific effects on quantizer capacity and legitimate-item reconstruction error are evaluated separately against matched benign growth (Section 5.2).

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

The paper studies poisoning of the fit set used to construct OPQ-based Semantic IDs. Injected fit samples can change the discrete representations of otherwise untouched items. Experiments cover three Amazon catalogs, three downstream recommendation architectures, ten text encoders, and text-derived poison. The main finding is that substantial tokenizer drift does not consistently predict recommendation harm, and benign catalog growth can cause comparable drift. The paper also evaluates anomaly detectors, trusted-reference admission screening, and frozen-rotation refitting with warm-started codebooks, including adaptive evasion and repeated updates.

Critical strengths
  1. A concrete upstream vulnerability is examined across several configurations. Deterministic clean-refit controls and permutation-invariant metrics support the finding that changing the fit set alters untouched items' representations. Cross-catalog and cross-encoder tests strengthen this result, and text-derived poison goes beyond arbitrary embedding injection. The paper appropriately acknowledges that three downstream architectures do not constitute three independent tokenizers.

  2. The benign-growth controls and negative results are informative. Sections 5.2–5.3 and Appendices D and E.6 show that large structural drift need not imply attack success: benign additions cause similar drift, some downstream estimates improve, and gate-evading mimicry does not demonstrate additional harm over benign growth. This directly challenges reliance on structural proxies for security claims.

Critical weaknesses
  1. The practical security impact remains less established than general refitting sensitivity. Table 15 mainly compares each condition with clean training. A significant attack–clean difference and a nonsignificant benign–clean difference do not establish an attack–benign difference. At k=121, benign growth reduces NDCG@10 by 10.9%, versus 8.8% for cluster-targeted insertion. The direct comparisons in Appendix E.6 likewise do not establish additional harm. Table 2's three significant results are also pointwise: if its nine comparisons form one family, none survives Holm correction at 0.05. Seed-level paired differences and intervals are needed. Finally, Section 3 defines the fit set through items appearing in training interactions, while the attacker cannot alter those interactions. The route from publishing item text to entering this fit set needs clarification.

  2. The defense evaluation does not yet connect structural stability to task utility. Freezing the rotation and allowing one Lloyd iteration preserves ARI, but leaving the tokenizer unchanged would also preserve old assignments. Although zero updates are mentioned, the paper does not provide a sufficient utility comparison. Since the paper itself shows that ARI does not reliably predict harm, improving ARI from roughly 0.43 to 0.99 is not enough to establish a security benefit. Comparisons should include new-item quantization quality and downstream recommendation under matched benign and malicious updates. Filtering baselines also remove legitimate fit samples; clean-data controls with equivalent filtering are needed to separate detection failure from drift caused by filtering itself.

Questions for authors
  1. Analysis of existing results: Can you provide seed-level paired differences and confidence intervals for attacks versus matched benign growth in Table 15, and clarify the comparison families and multiplicity treatment in Tables 2–3? Which downstream-harm conclusions remain supported? Nonsignificance should not be interpreted as equivalence.

  2. Threat-model clarification: How do injected items enter a fit set restricted to items appearing in training interactions when the attacker cannot manipulate those interactions? Is direct fit-set access assumed, or is there a concrete item-publication mechanism?

  3. Defense validation: At matched update sizes, how do zero updates, unrestricted warm-starting, one frozen-rotation update, and full refitting compare on legitimate new-item quantization and recommendation quality? How much drift do the filtering baselines cause in matched clean controls?

These questions concern the current security and defense claims; evaluation of every Semantic-ID tokenizer is not required.

7 comments

Q1. Statistical significance and matched benign growth

Thank you for raising these important statistical questions. We distinguish degradation relative to the pre-update tokenizer from additional harm relative to matched benign growth. To clarify this distinction, we report direct attack-versus-benign comparisons, seed-level paired differences, 95% confidence intervals, and multiple-testing corrections.

1. Direct attack-versus-benign comparisons. We compare attacks against benign additions at matched sizes k=6k=6 and k=121k=121 on DiffGRM/Beauty, using three paired recommender seeds. Each seed-level difference is 100(Ai−Bi)/Bˉ100(A_i-B_i)/\bar{B}, where AiA_i and BiB_i denote attack and benign NDCG@10, and Bˉ\bar{B} is the mean benign NDCG@10 across the three seeds. Negative differences indicate additional degradation under attack.

AttackkPaired differences (%)Mean [95% CI] (%)
Random6+5.0, +5.0, +9.7+6.6 [−0.1, +13.3]
Cluster-targeted6+8.5, +12.5, +9.0+10.0 [+4.6, +15.5]
Leverage6+2.0, +7.1, −2.4+2.2 [−9.5, +14.0]
Adjacent mimicry6+2.2, +0.9, +4.0+2.4 [−1.5, +6.3]
Random121−1.8, −3.5, −5.4−3.6 [−8.1, +1.0]
Cluster-targeted121+6.2, −8.0, +8.9+2.4 [−20.1, +24.9]
Leverage121−2.4, −14.4, −6.0−7.6 [−22.9, +7.7]
Adjacent mimicry121+3.6, −2.6, +2.5+1.2 [−7.1, +9.4]

At k=121k=121, leverage produces the largest observed attack-specific reduction (−7.6%), but its confidence interval includes zero (nominal p=.166p=.166, Holm p=.662p=.662). At k=6k=6, cluster-targeted additions yield numerically higher NDCG@10 than benign growth (+10.0%; nominal p=.016p=.016, Holm p=.125p=.125). None of the eight direct attack-versus-benign comparisons survives Holm or Benjamini-Hochberg (BH) correction (minimum adjusted p/q=.125p/q=.125).

2. Multiple-testing families. We apply Holm correction for family-wise error control and BH correction for false-discovery-rate control. The comparison families and their minimum adjusted values are:

Comparison familyTestsMin. Holm pMin. BH q
Poisoned vs. clean, pipeline × catalog9.163.085
Budget sweep, NDCG@512.112.097
Budget sweep, NDCG@1012.295.139
Attack vs. matched benign, k=6k=6 and 1211218.125.125
Updates vs. pre-update tokenizer, k=6k=6 and 12112110.051≈ .051

The nine pipeline–catalog comparisons, both budget sweeps, and the ten update-versus-pre-update comparisons yield no significant results after either correction. Within the ten-test update-versus-pre-update family, leverage shows a −17.7% NDCG@10 reduction (nominal p=.005p=.005), but this does not survive Holm correction (p=.051p=.051) or BH correction (q≈.051q\approx .051). These comparisons are relative to the pre-update tokenizer, not matched benign growth.

Conclusion. At k=121k=121, benign growth itself reduces NDCG@10 by 10.9% relative to the pre-update tokenizer, while leverage reduces it by 17.7%. However, the direct paired leverage-minus-benign difference (−7.6%) is not significant after correction. Additional downstream harm specifically attributable to poisoning is therefore not statistically established. With only three paired seeds, nonsignificance must not be interpreted as equivalence.

The revised results are reported in Section 5.3, Appendix A.7, and Appendix C (Tables 14, 16, and 17).

Q2. Threat-model clarification and fit-set admission

We thank the reviewer for the question. We clarify the assumed fit-set membership in our original experiments and provide new end-to-end validation showing how attacker-controlled items can enter the fit set through attacker-generated interactions.

Original threat model. The original experiments assume that attacker-controlled representations can be inserted directly into the OPQ fit set after clean PCA. These experiments do not require attacker interactions and do not add injected items to recommender training sequences. We clarify this assumption in Section 3.1.

Concrete admission mechanism. To demonstrate feasibility under the pipeline's 5-core filtering rule, we implement an end-to-end route on DiffGRM/Beauty. The attacker creates five accounts, each interacting with every injected item. Each new item consequently receives five interactions, while each attacker account also satisfies the 5-core requirement. Legitimate user histories remain unchanged. The injected items enter the recommender vocabulary, training sequences, and OPQ fit set. PCA and OPQ are then refitted, followed by recommender retraining (Appendix A.6).

Experimental verification. We compare 40 HotFlip items and 121 text-mimicry items against the same numbers of real Beauty listings introduced through identical attacker-account interactions, holding the interaction pattern fixed while varying the added items.

End-to-end admission and structural results (Table 13, revised manuscript):

Added itemsAdmittedTuple ARIExclusive codewords/digit
HotFlip (40)40/400.4701.50
Real listings (40)40/400.4420.00
Text mimicry (121)121/1210.4650.50
Real listings (121)121/1210.4580.00

All injected items successfully enter the fit set. Both attacks induce legitimate-item Semantic-ID reassignment and occupy codewords unused by legitimate items, whereas matched real listings occupy no exclusive codewords at these budgets.

Relative to matched real listings, NDCG@10 changes by -3.1% for HotFlip and +3.6% for text mimicry (Table 18). Neither establishes statistically significant additional downstream harm. We also verify complete fit-set admission at 5%, 10%, and 25% budgets, evaluating structural effects only at these larger scales (Table 13; Appendix B.9).

Scope and limitations. The interaction-based experiment establishes a feasible route for satisfying the dataset's admission rule; it does not eliminate the assumed-membership setting underlying the original experiments. Moreover, the five-account mechanism requires substantial attacker activity at larger budgets. We explicitly acknowledge these distinctions and the limited end-to-end evaluation scope in Sections 3.1 and 6.

Q3. Defense validation and matched clean controls

We compare no update, unrestricted warm-starting, one frozen-rotation update, and full refitting at matched update size k=121k=121 (Section 5.4; Tables 3, 23–24). Preserving existing Semantic IDs alone is insufficient to establish defense effectiveness, so we evaluate both quantization and downstream recommendation quality.

For legitimate new-item additions, the results are:

Update ruleOld-item ARINew-item errorΔ\DeltaNDCG@10 (%)
No update1.0000.7010.0
Frozen rotation, 1 iteration0.9930.674+0.1
Unrestricted warm start0.8010.633−6.1
Full refit0.4040.635−10.9

Quantization–stability tradeoff. Frozen rotation recovers approximately 40% of the new-item error reduction achieved by full refitting while maintaining near-perfect old-item partition agreement (ARI 0.993). Unrestricted warm-starting achieves lower new-item error (0.633) but substantially greater old-item drift (ARI 0.801). Thus, frozen rotation offers a better stability–adaptation tradeoff, not the lowest reconstruction error. Table 23 additionally reports structural comparisons at k=605k=605; downstream comparisons are limited to k=121k=121.

Downstream recommendation quality. Relative to no update, frozen rotation changes NDCG@10 by +0.1%+0.1\%, −1.1%-1.1\%, and −1.4%-1.4\% under benign growth, adjacent mimicry, and leverage poisoning, respectively. Full refitting produces corresponding changes of −10.9%-10.9\%, −9.8%-9.8\%, and −17.7%-17.7\%. Frozen rotation numerically outperforms full refitting in all three comparisons, but none survives Holm correction (minimum adjusted p=.086p=.086; Table 24). We therefore do not claim statistically established downstream superiority.

Filtering controls. At the matched 25% removal rate, applying the six filtering methods to clean data yields old-item ARIs of 0.252–0.464, compared with 0.273–0.457 on poisoned data. Random removal from clean data produces ARIs of 0.443–0.492 (Table 3; Appendix D.5). These controls demonstrate that substantial Semantic-ID drift also occurs without poisoning, as a consequence of removing items and fully refitting the tokenizer.

Conclusion. Frozen rotation reduces measured legitimate new-item quantization error while maintaining high old-item partition agreement. Its downstream advantage over full refitting remains suggestive rather than statistically established, and filtering alone does not prevent refit-induced drift.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Consolidated Response to the Reviewers

We sincerely thank all reviewers for their careful evaluation and constructive feedback. We summarize the principal concerns and the additional evidence and clarifications provided in response.

1. Reviewer fQcM

Concerns raised: The reviewer questioned whether poisoning causes additional harm beyond benign catalog growth, how attacker-controlled items enter the fit set under the 5-core rule, and whether frozen-rotation refitting provides meaningful downstream benefits. The reviewer also requested stronger statistical evidence and matched clean-data filtering controls. In the final comment, the reviewer acknowledged our clarifications but remained concerned about practical security harm and statistically supported downstream defense gains.

How we addressed it: We addressed these concerns through an independent replication and an expanded defense evaluation. First, a pre-registered replication across 40 independent tokenizer fits found that leverage poisoning occupies 0.32 more exclusive OPQ codewords per digit than matched benign additions (95% CI [0.18, 0.46], Holm-adjusted p=0.0003). The average attack-minus-benign NDCG@10 difference was not significant (-0.17%, 95% CI [-3.04%, +2.70%], p=0.91); therefore, we do not claim average recommendation degradation. Second, five additional paired recommender seeds showed that frozen-rotation refitting outperforms full refitting under benign growth, leverage poisoning, and adjacent mimicry, with all three comparisons surviving Holm correction (p=0.0017, 0.0076, and 0.037, respectively). Our study investigates the OPQ fit set as an upstream trust boundary, not an attack on the downstream recommender's training algorithm.

2. Reviewer UzCU

Concerns raised: The reviewer questioned whether the observed Semantic-ID changes reflected generic OPQ instability rather than poisoning-specific effects. Other concerns included inconsistent downstream impact, practical significance, fit-set admission, defense effectiveness, manuscript clarity, and bibliography accuracy. In the final follow-up, the reviewer acknowledged our additional experiments and revisions but remained unconvinced that we had sufficiently distinguished poisoning-specific effects from generic OPQ instability.

How we addressed it: We agree that ordinary OPQ refitting can cause substantial instability, and we explicitly acknowledge this in the revised manuscript. However, generic instability and adversarial exploitability are not mutually exclusive. Our independent 40-fit replication establishes a statistically significant poisoning-specific quantizer-capacity effect beyond matched benign growth for the evaluated leverage attack (0.32 additional exclusive codewords per digit; Holm-adjusted p=0.0003). We also clarify that this tokenizer-level effect does not imply average downstream recommendation degradation. The expanded five-seed evaluation further establishes statistically significant downstream advantages of frozen-rotation refitting over full refitting under all three evaluated update conditions. Finally, we corrected the bibliography, removed unintended annotations and citation artifacts, and revised the Related Work section. We respectfully request reconsideration of the ethics flag in light of these corrections.

3. Reviewer ZTA9

Concerns raised: The reviewer questioned the feasibility of fit-set admission, the inconsistent downstream recommendation impact, the distinction between poisoning and generic OPQ instability, and the limited controllability of the cluster-targeted attack. In particular, the reviewer questioned whether attacker-controlled items could practically enter the fit set and whether the attacker could achieve a specific, controlled objective.

How we addressed them: We added an end-to-end experiment with five attacker-controlled accounts, demonstrating that all 40 HotFlip items and 121 text-mimicry items successfully enter the fit set through attacker-generated interactions while leaving legitimate user histories unchanged. We introduced matched benign-growth controls and analyses of exclusive codeword occupancy and reconstruction error. The independent 40-fit replication further establishes a poisoning-specific quantizer effect for leverage poisoning. We also clarified the scope of the attack: our objective is untargeted Semantic-ID manipulation, not guaranteed item-level promotion, demotion, or category-specific manipulation. In particular, cluster-targeted refers to placing injected representations near a selected centroid; it does not guarantee greater reassignment in that cluster.

The revised work clarifies the OPQ fit set as an upstream security boundary, distinguishes generic refitting instability from measurable poisoning-specific quantizer effects, demonstrates a feasible fit-set admission mechanism, and strengthens the statistical evaluation of constrained tokenizer refitting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.