{"id":"2ddda24e-2237-4568-8b0e-88c05a60c319","arxiv_id":"2608.20991","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"STAG is a coordinated graph-and-text backdoor attack that succeeds by steering both modalities toward a common target-class region in graph-language alignment.","lead":"This paper proposes STAG, a backdoor attack that secretly hijacks graph-based AI models that combine graphs with text. The attack poisons both the graph structure and the text encoder together, so triggered inputs are forced to a chosen class while clean inputs behave normally.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported ASR may be an artifact of training the victims with the same simplified alignment loss (Eq. 2) that STAG is optimized against; real GFM training objectives are not shown to be captured by this loss.","rationale":"The reader's weakest assumption is that victim GFM training is faithfully captured by Eq. (2). My reading sharpens this into a concrete experimental issue: the paper's own attack pipeline trains the surrogate with a lower-level objective of the same form as Eq. (2), and the reported victim GFM results are not accompanied by any specification of the victim training loss, so there is no evidence that the victims were trained differently from the surrogate. This matters because the entire mechanism, coordinating the graph trigger and text prompt so that both move toward the target-class text region, only transfers if the victim's optimization is dominated by graph-text similarity to the corresponding text. Real GFM pipelines use additional or different terms, such as supervised classification losses, contrastive negatives, or instruction-following objectives. If those terms are present, the poisoned alignment can be suppressed. This is a load-bearing concern rather than a style complaint: it directly threatens the empirical claim of at least 92.40% ASR across three victim GFMs. I nonetheless keep the verdict unchanged from the reader's CONDITIONAL because the paper provides a code release, a plausible threat model, and substantial experiments; the concern is addressable by a concrete re-training test under more faithful victim objectives. The test I propose would settle the issue without requiring access to official GFM checkpoints: it can be run with the paper's own code by changing only the victim training loss. I agree with the reader's identification of the same underlying weakness.","tokens_in":16605,"tokens_out":6981,"duration_ms":71905,"concrete_test":"Re-run the Table III Cora/GraphCLIP configuration with victim training losses L = L_align + lambda * CE (supervised cross-entropy on the same poisoned nodes with clean labels) for lambda in {0, 0.1, 1.0}, and separately with L_align replaced by a standard batch InfoNCE loss over graph-text pairs with in-batch negatives. Keep STAG's trigger generation, soft prompt, and poison rate exactly as in the paper, and report ASR and ACC for each setting. If ASR drops materially below 92.40% under any of these victim objectives, then Eq. (2) is not a faithful model of victim training and the headline transferability claim is not established; if ASR remains high, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (2) is not only an analytical abstraction; in the experiments it appears to be the actual training objective used for the three victim GFMs. The surrogate lower-level objective in Eq. (5) and in the constraint of Eq. (13) has the same functional form, with poisoned nodes pulled toward the target-class centroid c_t. Thus STAG's trigger generator and soft prompt are optimized against essentially the same loss that the victim is then trained with. The paper does not report the victim training objective for GraphGPT, GraphCLIP, or G2P2, and does not show that any of these victims was trained with its published objective (instruction tuning, contrastive loss with in-batch negatives, or graph-grounded prompting) rather than the simplified L_align. If a real victim objective includes a supervised classification term, hard-negative contrastive terms, or a text encoder that is not frozen, the poisoned graph-text association can be overwritten or simply never learned. In that case the claimed at-least-92.40% ASR across all datasets and victim GFMs could be a consequence of matching training losses rather than a property of actual GFM training. The central claim therefore rests on the unverified assumption that Eq. (2) is the dominant and faithful description of victim GFM training.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes STAG, a backdoor attack against graph foundation models (GFMs) on text-attributed graphs (TAGs). STAG jointly optimizes a graph-trigger generator and a text-side soft prompt so that triggered graph and text representations are pulled toward the same target-class text region under the graph-language alignment interface. It addresses TAG-specific stealth by realizing trigger-node attributes as readable text via LLM-generated candidate retrieval and by regularizing trigger subgraph degree statistics. Experiments on Cora, CiteSeer, WikiCS, and OGB-arxiv, against GraphGPT, GraphCLIP, and G2P2, and under Prune, Outlier Detection, and DOMINANT defenses, report ASR values above 92.40% with limited clean-accuracy degradation. The paper also includes coordination analysis, stealthiness measurements, ablations, and trigger-size/poison-rate studies. Code is provided.","tokens_in":16881,"tokens_out":7781,"duration_ms":75132,"significance":"If the reported numbers are reproducible and the victim models are trained with their published objectives, the paper makes a useful contribution: it identifies a concrete vulnerability of the graph-language alignment interface, demonstrates why single-modal graph or text triggers fail, and proposes a coordinated cross-modal attack with two concrete stealth mechanisms. The manuscript is technically detailed, ships code, and provides falsifiable empirical claims together with ablation and defense evaluations. The main caveat is that the experimental validation may be confounded by the victim training objective: the attack's surrogate losses are built on the same simplified alignment loss used to abstract GFM training, and the paper does not disclose the actual victim training loss in the experiments. This issue needs to be resolved before the central empirical claim is accepted.","major_comments":[{"comment":"The paper never reports the exact training objective used for the three victim GFMs in the experiments. Eq. (2) is introduced as an abstraction of GFM training, and the attack's surrogate losses in Eqs. (5) and (13) are built from the same sim function. If the victim graph transformers in Table III are trained with exactly Eq. (2) or its poisoned variant, the reported at-least-92.40% ASR may reflect objective matching between the attacker's optimization and the victim's training loss, rather than successful transfer to actual GFM training procedures (e.g., GraphCLIP's contrastive loss with in-batch negatives, GraphGPT's instruction tuning, or G2P2's graph-grounded prompting). Please state precisely the training loss and training pipeline for each victim model used in the evaluation; if the published objectives differ, rerun the attack against victims trained with those objectives or against official pretrained checkpoints and report the resulting ASR and ACC.","section":"Sec. III-A (Eq. (2)) and Sec. V-A"},{"comment":"All results are said to be averaged over five runs, but no standard deviations, confidence intervals, or error bars appear in any table or figure. Point estimates such as 100.00% ASR in Table III and 95.20% ASR under defenses in Table IV are presented without uncertainty, so the central claim that STAG 'reaches at least 92.40% ASR in every setting' is a claim about unreported point estimates. Please add run-level variance (standard deviation or confidence interval) for the main results in Tables III and IV and for the ablation, trigger-size, and poison-rate curves.","section":"Sec. V-A, Tables III and IV"},{"comment":"There is a mismatch between the threat model and the reported implementation. The threat model states that the attacker releases a frozen text encoder wrapped with a learnable soft prompt e_p and that the victim uses e_p as its text-side encoder; the implementation details, however, mention only 'frozen SBERT' for node texts and do not state that the victim uses the soft prompt. The fixed trigger phrase r and its insertion position are also never defined. Because the coordinated cross-modal mechanism in Eq. (7) depends on e_p being applied to all text (clean and triggered) during victim training, the experimental setup must explicitly confirm that e_p is part of the victim's text encoder. Please clarify this and specify r and its insertion rule.","section":"Sec. III-B and Sec. V-A"},{"comment":"The ablation study in Fig. 3 reports only ASR, without numeric values or error bars, and the conclusion that the graph-side objective is 'especially critical' relies on visual inspection. Additionally, Table IV aggregates defense results over the three victim GFMs and only two datasets, so the per-model behavior under defenses is not reported. Please provide a numeric table with ASR and ACC (with run-level variance) for STAG/G, STAG/T, STAG/C, and full STAG on all datasets used in the ablation, and provide per-victim breakdowns for the defense experiments.","section":"Sec. V-D, Fig. 3, and Sec. V-B, Table IV"}],"minor_comments":[{"comment":"The 'Clean Graph' column header is confusing because the column contains clean accuracy only, not the 'ASR|ACC' format used for the other columns; consider renaming it 'Clean ACC'.","section":"Table III"},{"comment":"In the coordination analysis, the definitions of s_t, s_y, and Delta_gap should state explicitly that e(y_i) is the text anchor of the node's true class and that s_t and s_y are computed on the same set of attacked test nodes; the current prose is informal.","section":"Sec. V-C"},{"comment":"Algorithm 1 does not specify the values of the learning rates eta_sigma, eta_s, eta_g, eta_p, the inner-loop count M, the outer-loop count N, or the loss weights lambda_c1 and lambda_c2; even with code available, the paper should report these hyperparameters for reproducibility.","section":"Algorithm 1 and Sec. V-A"},{"comment":"The defense implementations are described in one line each (e.g., 'Prune removes edges with cosine similarity below 0.4'); please specify whether these values were tuned on clean data and how the defenses were applied to triggered subgraphs, since the reported 95.20% ASR after defenses depends on these choices.","section":"Sec. V-B"}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate new attack surface—coordinating graph-side and text-side triggers through the graph-language alignment interface—and the paper demonstrates the mechanics cleanly. But the experiments may be validating the attack against a simplified training loss rather than actual GFM training, and the claims outrun the evidence.\n\nThe core idea is good: instead of poisoning one modality, STAG jointly optimizes a trigger subgraph and a soft prompt so both pull toward the same target-class text region. The equations are clear, and the embedding-closure analysis in Fig. 2 is a convincing demonstration of why single-modal attacks fail. The stealthiness work—realizing trigger nodes as readable text via candidate retrieval and matching degree statistics—is a thoughtful response to the TAG setting. Code is out, which helps.\n\nThe problems are in the experimental validation. Most importantly, the paper never states the victim training objective. The attack is optimized against a lower-level surrogate trained with the same alignment loss L_align in Eq. (2) that is used to abstract GFM training. If the victims in Sec. V were also trained with that exact loss—or a close variant—then the high ASR might just mean the attack matches the training loss, not that it transfers to a real GFM trained with instruction tuning, in-batch contrastive negatives, or an auxiliary classification head. The stress-test note has it right: this is the load-bearing assumption, and it isn't defended. Adding experiments with victims trained under their published objectives (e.g., actual GraphCLIP with hard negatives, or GraphGPT with instruction tuning) would settle it.\n\nSecond, the two closest prior attacks, GFM-BA and DTGBA, are discussed but not included as baselines. Without them, the novelty comparison is incomplete. Third, no error bars are reported despite averaging over five runs. Fourth, the text realization step depends on Claude Haiku 4.5, so the exact trigger texts are not reproducible without specifying the prompt template.\n\nThese are fixable. The attack idea is worth taking seriously, and the paper deserves a real referee. My recommendation: send it out, but the reviewers should require the victim-training-objective experiments, the missing baselines, and error bars before acceptance.","headline":"A genuinely new coordinated attack on graph-language alignment, but the experiments may only show it works when the victim is trained with the same simplified loss the attack is built to exploit.","tokens_in":17393,"tokens_out":3242,"duration_ms":29804},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper is trying to prove that graph foundation models on text-attributed graphs are backdoor-vulnerable at their graph-language alignment interface, not merely on the graph side or text side alone.","keywords":["backdoor attack","graph foundation models","text-attributed graphs","graph-language alignment","soft prompt","trigger stealthiness","trojaning","cross-modal attack"],"falsifier":"Train a victim graph foundation model under the same STAG poisoning but with a training objective that adds a standard supervised classification loss to the graph-text similarity loss, then measure attack success on trigger-attached test nodes; if the alignment-only abstraction of Eq. (2) is load-bearing, the reported 92.40% attack-success floor should drop substantially. The same test can be run with a fine-tuned rather than frozen text encoder.","tokens_in":16400,"feed_emoji":"🎯","tokens_out":7933,"duration_ms":70443,"temperature":0.7,"pith_summary":"The paper is trying to establish that graph foundation models on text-attributed graphs are backdoor-vulnerable at their graph-language alignment interface, rather than only on the graph side or the text side. It argues that single-modality triggers fail because clean text pulls poisoned graph representations back toward the benign semantic region, while text-only triggers never move the graph embedding that is actually scored. STAG is a trojan framework that jointly optimizes a graph-trigger generator and a text-side soft prompt so trigger-attached graphs and triggered text move toward the same target-class text region. It also makes triggers hard to spot by rendering trigger-node attributes as human-readable candidate text and by regularizing the trigger-attached subgraph's degree statistics. If the paper is right, poisoning an upstream text encoder or public text-attributed graph can implant a persistent, transferable misclassification in downstream graph foundation models without degrading clean accuracy.","feed_headline":"Reach 92-100% attack success by hijacking graph-language alignment","feed_subtitle":"STAG coordinates graph and text triggers so poisoned graphs land in the target class while clean accuracy survives.","key_machinery":"The load-bearing object is the shared graph-language alignment interface, abstracted as an objective that maximizes similarity between a graph encoder's embedding of a node's K-hop subgraph and a text encoder's embedding of the node's text. STAG's machinery consists of a graph-trigger generator, an MLP that maps the node's text embedding into the feature matrix and adjacency matrix of a trigger subgraph; a text-side soft prompt prepended to the frozen text encoder; a surrogate graph encoder trained in place of the inaccessible victim graph encoder; and a bi-level optimization that alternates projected-gradient ascent on an adversarial perturbation, descent on the surrogate, descent on the trigger generator, and descent on the soft prompt. The trigger feature is kept near the target-class text centroid under a bounded perturbation, then replaced by the closest human-readable candidate text retrieved from an LLM-generated candidate pool, and the trigger subgraph is regularized to keep the mean and variance of its degree distribution close to the original subgraph.","core_discovery":"The central claim is that a backdoor against graph foundation models on text-attributed graphs should be planted in the shared graph-language alignment objective rather than in either modality alone. The paper abstracts victim training as maximizing similarity between graph embeddings and text embeddings, and designs STAG to solve a bi-level min-max problem in which a learned trigger generator produces a trigger subgraph for each poisoned node while a learnable soft prompt, prepended to the frozen text encoder, shifts triggered text embeddings in the same direction. The two attack signals are coordinated through a target-class text centroid: the graph-side objective pulls trigger-attached subgraphs toward that centroid, the text-side objective aligns triggered text with those subgraphs, and a similarity function that combines cosine similarity with a Euclidean penalty makes the objective more robust to an unknown downstream prediction head. Empirically, STAG reports attack success rates of at least 92.40% across all tested datasets and victim models, with clean accuracy close to the clean baseline, while single-side baselines stay below roughly 42% attack success and often damage clean accuracy.","pith_inferences":["Implicit consequence: the same upstream compromise that enables STAG could propagate to many downstream graph foundation models that reuse the poisoned graph or text encoder, so a single poisoned encoder release may backdoor models the attacker never directly trained against.","A natural extension the paper does not test is whether an alignment-level defense that measures how close trigger-attached graph embeddings are to target-class text centroids would detect STAG, since the attack's own closure-rate diagnostic shows an unusually large fraction of poisoned embeddings entering that region.","The mechanism should carry over to graph retrieval-augmented generation pipelines, where poisoned alignment could steer which subgraphs get retrieved and fed to an LLM; the authors point to this direction as open, and it would be testable by attaching the same trigger subgraph and prompt to retrieval queries.","An untested robustness question is whether STAG survives a victim that fine-tunes the text encoder or adds classification losses, since the paper's abstraction of victim training as pure graph-text alignment is what makes the coordinated poisoning transfer."],"forward_implications":["Single-modality backdoor attacks adapted from the graph side or text side are not sufficient against graph foundation models on text-attributed graphs, so future attack and defense work should treat the alignment interface as the target.","Because STAG reports at least 92.40% attack success while keeping clean accuracy near the clean baseline, a backdoor can survive validation based on clean accuracy alone.","Readable trigger text and degree-matched trigger subgraphs keep the attack effective under feature-based, outlier-based, and structural defenses, meaning text-attributed-graph-specific stealth is achievable without sacrificing attack strength.","The attack works under a gray-box supply-chain assumption where the attacker controls only the public graph and a frozen text encoder, so upstream data or encoder compromise is the threat scenario that needs monitoring.","The graph-side objective is the more critical component: ablations show removing it causes the largest drop in attack success, while removing either concealment loss mainly hurts robustness against defenses."],"supporting_citations":[{"why":"supplies the LLM-as-predictor victim model used in the attack evaluation and defines one GFM paradigm being attacked.","marker":"[17]"},{"why":"supplies the graph-grounded pretraining and prompting method used as a second victim model and the alignment-style objective STAG abstracts as Eq. (2).","marker":"[18]"},{"why":"supplies the contrastive graph-language alignment victim model used in the main experiments and coordination analysis.","marker":"[19]"},{"why":"is the text-side backdoor baseline whose low success motivates adding a coordinated graph-side trigger.","marker":"[24]"},{"why":"is the graph-side backdoor baseline whose pull-back toward clean text motivates the text-side soft prompt.","marker":"[29]"},{"why":"supplies the parameter-efficient prompt-tuning method that makes the soft-prompt capability a realistic deployment assumption.","marker":"[38]"},{"why":"provides the frozen sentence encoder whose embedding space defines the target-class text region and node-text features.","marker":"[40]"},{"why":"is the multimodal backdoor baseline adapted to the text-attributed-graph setting and included in the comparison tables.","marker":"[45]"}],"fun_headline_variants":["Stealthy backdoor attacks via coordinated graph-text triggers","Hijack GFM alignment for stealthy, high-success backdoors","STAG: poisoning the alignment interface, not just one side","Graph-language alignment is the new backdoor surface","Coordinated triggers beat single-side attacks on GFMs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The attack assumes that the victim's graph foundation model is trained essentially by maximizing similarity between graph embeddings and text embeddings, so the alignment objective in Eq. (2) is the whole training signal that imprints the trigger-target association.","fun_headline_variants_meta":{"raw":{"variants":["Stealthy backdoor attacks via coordinated graph-text triggers","Hijack GFM alignment for stealthy, high-success backdoors","STAG: poisoning the alignment interface, not just one side","Graph-language alignment is the new backdoor surface","Coordinated triggers beat single-side attacks on GFMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1707,"prompt_tokens":1041,"completion_tokens":666,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":584}},"tokens_in":657,"tokens_out":666,"duration_ms":6526,"temperature":1.0,"reasoning_tokens":584,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-27T19:08:19.520505+00:00","model_set":{"reader":"deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.0-alltime-deepseek-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Train a victim graph foundation model under the same STAG poisoning but with a training objective that adds a standard supervised classification loss to the graph-text similarity loss, then measure attack success on trigger-attached test nodes; if the alignment-only abstraction of Eq. (2) is load-bearing, the reported 92.40% attack-success floor should drop substantially. The same test can be run with a fine-tuned rather than frozen text encoder.","supporting_citations":[{"cited_title":"Graphgpt: Graph instruction tuning for large language models,","cited_arxiv_id":null,"evidence_quote":"supplies the LLM-as-predictor victim model used in the attack evaluation and defines one GFM paradigm being attacked."},{"cited_title":"Augmenting low-resource text classification with graph-grounded pre-training and prompting,","cited_arxiv_id":null,"evidence_quote":"supplies the graph-grounded pretraining and prompting method used as a second victim model and the alignment-style objective STAG abstracts as Eq. (2)."},{"cited_title":"Graphclip: Enhancing transferability in graph foundation models for text-attributed graphs,","cited_arxiv_id":null,"evidence_quote":"supplies the contrastive graph-language alignment victim model used in the main experiments and coordination analysis."},{"cited_title":"Poisonprompt: Backdoor attack on prompt- based large language models,","cited_arxiv_id":null,"evidence_quote":"is the text-side backdoor baseline whose low success motivates adding a coordinated graph-side trigger."},{"cited_title":"Cross- context backdoor attacks against graph prompt learning,","cited_arxiv_id":null,"evidence_quote":"is the graph-side backdoor baseline whose pull-back toward clean text motivates the text-side soft prompt."},{"cited_title":"The power of scale for parameter-efficient prompt tuning,","cited_arxiv_id":null,"evidence_quote":"supplies the parameter-efficient prompt-tuning method that makes the soft-prompt capability a realistic deployment assumption."},{"cited_title":"Sentence-bert: Sentence embeddings using siamese bert-networks,","cited_arxiv_id":null,"evidence_quote":"provides the frozen sentence encoder whose embedding space defines the target-class text region and node-text features."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}