- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper shows that Semantic IDs (SIDs) in generative recommenders are vulnerable: small text edits can push an item's embedding across residual-quantization (RQ) boundaries and change its SID, and the RQ routing margin strongly predicts which items flip (AUC up to 0.93). The authors propose SIDAttack, which appends a short suffix to a target item's description and uses gradient-based search (AnchorPort-GCG) on public text encoders toward diverse popular-item anchors, without access to the victim tokenizer or recommender. It yields the largest exposure gains among the evaluated baselines across product, news and short-video datasets and two SID tokenizers. Gains are smaller but still positive when the victim encoder is held out. They also propose margin-regularized codebook construction, which cuts attack-induced exposure by roughly 55% or more at a 24-token suffix while keeping over 99% of clean Recall@50. A residual gain remains at 48 tokens, so the defense raises the attack cost rather than eliminating the attack.
Critical strengths
- The angle is fresh: tweak a listing's text and its SID can change, with no retraining of the recommender needed. Treating that as an attack surface is realistic and under-explored.
- The RQ margin first predicts which items flip (AUC up to 0.93), then the attack exploits that, and finally the defense reuses the same margin idea.
- The experiments cover a lot of ground. Three Amazon domains plus news (MIND) and short video (KuaiRec), two tokenizers, cross-encoder transfer, suffix-length sweeps and ablations. The appendix also compares against text-purification defenses like ONION and TAP.
Critical weaknesses
- The attacker needs white-box gradient access to a public encoder, and the baselines are adapted into a catalog-rewrite setting rather than designed for this scenario. HotFlip only gets a single substitution. So it's hard to tell how much of the big lead comes from the method itself.
- Only 100 target items per domain, and the gains are heavy-tailed, with single items jumping by +732 or +1,090. The reported standard deviations come from three seeds only and ignore item-to-item variance, so the statistics aren't very convincing.
- Everything is offline simulation, with exposure measured as top-50 counts over test users and nothing validated on a real platform. The only held-out encoder is BGE, and the surrogate encoders are all similar embedding models, so the transfer claims may be optimistic.
Questions for authors
- The margin-regularized defense is only tested against edits that were generated beforehand against the standard tokenizer. What happens if the attacker re-runs AnchorPort-GCG against the margin-regularized tokenizer, or even against a surrogate of it? Since a residual gain already remains at n=48, how much of the defense survives an adaptive attacker?
- The only held-out victim encoder is BGE, and the surrogates are all similar embedding models. How does the attack transfer to a structurally different encoder (e.g., an LLM-based or domain-finetuned encoder, or one with a different tokenizer)? Is the joint multi-surrogate gain due to diversity of encoders, or simply to more total optimization compute?
3 comments
Response to Reviewer yoYb
W2: Target sample size and item-level variance
Only 100 target items per domain, and the gains are heavy-tailed, with single items jumping by +732 or +1,090. The reported standard deviations come from three seeds only and ignore item-to-item variance, so the statistics aren't very convincing.
Every baseline, ablation, transfer experiment, and defense evaluation uses the exact same 100 target items and paired seeds per domain, so all method comparisons are strictly paired on identical items. To further characterize item-to-item variation, we will add item-level paired statistics and confidence intervals in the final version.
W3 & Q2: Offline scope, LLM-based / finetuned victim encoders, and equal-compute multi-surrogate transfer
Everything is offline simulation, with exposure measured as top-50 counts over test users and nothing validated on a real platform. The only held-out encoder is BGE, and the surrogate encoders are all similar embedding models, so the transfer claims may be optimistic.
The only held-out victim encoder is BGE, and the surrogates are all similar embedding models. How does the attack transfer to a structurally different encoder (e.g., an LLM-based or domain-finetuned encoder, or one with a different tokenizer)? Is the joint multi-surrogate gain due to diversity of encoders, or simply to more total optimization compute?
Offline evaluation. Following standard practice in generative recommendation, we evaluate under the full-ranking protocol on real interaction logs, as injecting adversarial product descriptions into a live platform would directly affect real users and sellers.
Structurally different and domain-finetuned encoders. To evaluate transfer across distinct model architectures and tokenizers, we use BGE (or BGE + E5 + GTE) as public surrogates and transfer the suffixes to two held-out LLM-based encoders, Qwen3-Embedding-0.6B and Qwen3-Embedding-4B, each paired with an independently trained RQ tokenizer and generator. Regarding domain-finetuned encoders, recommendation pipelines typically apply lightweight parameter-efficient adaptation such as LoRA while learning domain specialization in the quantizer and generator, which preserves the underlying token embedding layer and semantic neighborhood structure.
Equal-compute multi-surrogate search. To separate encoder diversity from total compute, we match the total forward and backward compute of joint BGE + E5 + GTE search to that of single-surrogate BGE. Under the same compute budget, joint multi-surrogate search consistently outperforms BGE alone across all datasets and both Qwen3 victims, confirming that the transfer gain comes from encoder diversity rather than extra optimization compute.
| Dataset | Held-Out Victim | Public Search Encoders | Clean Exp@50 | Attacked Exp@50 ↑ | Exp@50 ↑ | Clean Recall@50 |
|---|---|---|---|---|---|---|
| Beauty | Qwen3-Embedding-0.6B | BGE only | 0.17 | 2.64 | +2.47 | 0.1768 |
| Beauty | Qwen3-Embedding-0.6B | BGE + E5 + GTE (Equal Compute) | 0.17 | 3.28 | +3.11 | 0.1768 |
| Beauty | Qwen3-Embedding-4B | BGE only | 0.09 | 3.37 | +3.28 | 0.1847 |
| Beauty | Qwen3-Embedding-4B | BGE + E5 + GTE (Equal Compute) | 0.09 | 4.28 | +4.19 | 0.1847 |
| Industrial & Sci. | Qwen3-Embedding-0.6B | BGE only | 0.24 | 18.07 | +17.83 | 0.2043 |
| Industrial & Sci. | Qwen3-Embedding-0.6B | BGE + E5 + GTE (Equal Compute) | 0.24 | 22.54 | +22.30 | 0.2043 |
| Industrial & Sci. | Qwen3-Embedding-4B | BGE only | 0.14 | 24.75 | +24.61 | 0.2136 |
| Industrial & Sci. | Qwen3-Embedding-4B | BGE + E5 + GTE (Equal Compute) | 0.14 | 31.44 | +31.30 | 0.2136 |
| Toys & Games | Qwen3-Embedding-0.6B | BGE only | 0.29 | 41.55 | +41.26 | 0.1652 |
| Toys & Games | Qwen3-Embedding-0.6B | BGE + E5 + GTE (Equal Compute) | 0.29 | 54.67 | +54.38 | 0.1652 |
| Toys & Games | Qwen3-Embedding-4B | BGE only | 0.21 | 68.63 | +68.42 | 0.1728 |
| Toys & Games | Qwen3-Embedding-4B | BGE + E5 + GTE (Equal Compute) | 0.21 | 91.41 | +91.20 | 0.1728 |
Response to Reviewer yoYb
Q1: Adaptive attacks against the margin-regularized defense
The margin-regularized defense is only tested against edits that were generated beforehand against the standard tokenizer. What happens if the attacker re-runs AnchorPort-GCG against the margin-regularized tokenizer, or even against a surrogate of it? Since a residual gain already remains at n=48, how much of the defense survives an adaptive attacker?
To evaluate an adaptive attacker after the defense is deployed, we freeze the margin-regularized tokenizer and generator, train a separate margin-regularized surrogate quantizer on public catalog embeddings, and add a differentiable soft-routing term toward the anchor's surrogate code :
R_{\mathrm{proxy}}(s)=\frac{1}{3}\sum_{\ell=1}^{3}\log p_{\widetilde{Q},\tau}\\!\left(q_{a,\ell}\mid E(x\oplus s)\right),\quad \tau=0.03.$$ Comparing fixed pre-defense edits, a fresh rerun of $J_{\mathrm{public}}$, and adaptive re-optimization with $J_{\mathrm{adapt}}$ under the same search budget shows that adaptive re-optimization recovers part of the exposure gain relative to fixed edits, whereas rerunning $J_{\mathrm{public}}$ alone changes little. Even so, across all domains and budgets, the defense keeps the adaptive gain well below the undefended attack while preserving clean Recall@50, confirming that margin regularization substantially raises the cost of attack. | Dataset | Budget $n$ | Fixed-Edit $\Delta$Exp@50 ↓ | Public Rerun $\Delta$Exp@50 ↓ | Adaptive $\Delta$Exp@50 ↓ | Undefended $\Delta$Exp@50 | Clean Recall@50 | | --- | --- | --- | --- | --- | --- | --- | | Beauty | 24 | -0.04 | +0.09 | +1.24 | +6.14 | 0.1831 | | Beauty | 48 | +5.93 | +6.08 | +9.91 | +15.68 | 0.1831 | | Industrial & Sci. | 24 | +27.16 | +28.11 | +35.93 | +61.35 | 0.2124 | | Industrial & Sci. | 48 | +30.58 | +31.19 | +43.88 | +65.45 | 0.2124 | | Toys & Games | 24 | +83.76 | +85.29 | +109.47 | +182.64 | 0.1716 | | Toys & Games | 48 | +252.31 | +257.66 | +342.08 | +518.87 | 0.1716 | We hope these additional experiments and clarifications address your concerns, and we respectfully invite you to reconsider your evaluation of our work.
Response to Reviewer yoYb
We thank the reviewer for recognizing the Semantic ID attack surface and the routing-margin analysis, and for the constructive questions on baseline fairness, statistical reporting, cross-encoder transfer, and adaptive attacks against our defense.
W1: Baseline design and equal-budget comparisons
Gradient access is used only on the public text encoder, never on the victim recommender or codebook. Because no prior attack is designed for catalog rewriting against SID-based recommenders, adapting existing text attacks is necessary. To ensure a fair comparison, all gradient baselines share the same public encoder, anchor pool, editable region, and vocabulary, and we evaluate three controlled comparisons under a matched budget of C=1,536 candidate evaluations on the same 100 targets per domain:
(1) Native HotFlip (in-place body edits). HotFlip was originally designed to replace tokens inside the main text. Applied directly to the product body on Beauty, it corrupts core product facts, substantially increases perplexity, and still yields lower exposure gains than appending a suffix—which is why our main tables evaluate all methods in the suffix setting.
(2) Single anchor with random restarts. Running single-anchor GCG with three random restarts toward the same anchor under the same total budget (C=1,536) recovers only a small part of the gap, confirming that the gain comes from multi-anchor diversity rather than multiple random initializations.
(3) Equal-budget multi-step HotFlip and single-anchor GCG. Upgrading HotFlip to multi-step iterative search across n∈12,24,48, SIDAttack consistently outperforms both iterative HotFlip and single-anchor GCG across all domains and budgets while matching single-anchor GCG in runtime and running faster than iterative HotFlip.