acceptodds
Under review as a conference paper at ICLR 2027

Propagation Identity Perturbations: A New Adversarial Threat Model for Robust Social Event Forecasting

Abstract

Social event forecasting models depend critically on who participates in the propagation process, which differs from conventional graph learning tasks. Existing robustness research in this area intensively studies perturbations on graph topology, node features, or textual content, but leaves propagation identities underexplored. We propose Propagation Identity Perturbation (PIP), a new adversarial threat model that formalizes attacks on participant identities in temporal diffusion trajectories while preserving realistic propagation behavior. We instantiate PIP as a gradient-guided discrete attack that searches behaviorally plausible identity substitutions within constrained candidate spaces. Moreover, we provide theoretical support for PIP in feasible search space reduction, transferability lower bound, monotonic ascent property, and guaranteed loss disruption bound. Extensive experiments on five representative social event forecasting architectures against five mainstream defense measures further expose the severity of this overlooked vulnerability in diffusion forecasting systems. These findings highlight an urgent need to develop identity-aware defense mechanisms to defend social event forecasting models against PIP.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Propagation identity perturbations, robust social event forecasting, realistic propagation behavior

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_8FUyOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Propagation Identity Perturbation (PIP), an adversarial threat model for social event / information diffusion forecasting, where the attacker replaces up to B participant identities in an observed cascade while keeping timestamps and the social graph fixed. Replacements are restricted to “behaviourally plausible” candidates (active within ρ days, similar relative position in cascades, within r hops of the previous participant). The attack is a HotFlip-style greedy search on a source model (MILD): gradient scores propose candidates, the top 128 are verified by exact forward passes, and the best substitution is accepted each round. The resulting perturbed cascades are transferred, without queries, to five target models (RotDiff, DyHGCN, MS-HGAT, DisenIDP, GODEN) on one Twitter dataset. With B = 6, average Hits@10 drops from 34.02% to 25.29% (relative drop 25.66%), and five adversarial-training-style defenses reduce this only slightly (best: AT, 22.75%). The paper also gives four theorems on search space reduction, transferability, monotonic ascent and a loss bound.

Critical strengths

A new and relevant attack surface. Most graph robustness work perturbs edges or node features, while this paper looks at who participates in a cascade, which is the main input of diffusion forecasting models. The results show that these models are sensitive to a small number of identity changes across five quite different architectures. This matters to my assessment because it identifies a vulnerability that has received little attention in diffusion forecasting.

Critical weaknesses
  1. The defense evaluation does not support the claim that existing defenses fail.

Table 2 only reports defended models on perturbed test data. The clean accuracy of each defended model is not reported, and RHD for the defended rows is still computed against the undefended model’s clean accuracy. Without both clean and robust accuracy after defense, we cannot judge what AT (or any defense) actually costs or gains. Only one budget (B = 6) is reported for the defenses. A defense evaluation should show results under several budgets, and ideally with the defense trained and tested under different budgets. The defenses are trained on a fixed set of perturbations generated once on MILD, not regenerated against the model being trained, so “AT” here is static adversarial data augmentation rather than Madry-style adversarial training. Training is also very light (at most 10 epochs, learning rate 2×10⁻⁵), and epoch 0 is allowed as a no-update fallback. Several defended numbers are identical to the undefended ones to four decimals (MS-HGAT with MART and ATFL: 23.3667; DisenIDP with SemRoDe-MMD: 23.8488), which suggests that no update was selected in those cases. All five defenses belong to the adversarial-training family; no detection-based or graph-specific defenses are evaluated, although the paper claims PIP is “unaddressed by standard graph defenses”. The budget is defined as an absolute number of positions (B = 6), but the cascade length distribution and the fraction of positions actually changed are not reported. Changing 6 users in a short cascade is very different from a long one; a ratio of the cascade length would be a more meaningful budget.

  1. The experiments are limited, and the novelty and theory are overstated.

Only one dataset (Twitter, 346 test cascades) is used. The ablation against random legal replacement is only on MS-HGAT at B = 6, and the budget sensitivity (Table A2) is reported on the source model MILD itself, not on the transfer targets. There are no other attack baselines. The plausibility of the perturbations is never tested. The constraints appear loose (on average about 2,900 of 12,627 users are eligible per position, and results do not change for ρ ≥ 3 days), and the attack deliberately picks the least expected substitutions, so a simple likelihood-based detector might catch them. Substituting elements of an ID sequence with gradient-guided search has been studied in sequential recommendation (e.g., substitution-based profile pollution attacks and their defenses, Yue et al., 2022), which is not discussed. The claim of being the first threat model of this kind needs to be positioned against this work. The theory adds little. Theorems 1 and 3 hold by construction. Theorem 2 assumes the source and target representations are aligned up to an orthogonal matrix, but Appendix A.2.3 states that the target models share no common latent dimensions or coordinate system, and the proof of Theorem 2 uses “W is diagonal” instead of orthogonal.

Questions for authors
  1. Please report the clean Hits@10 of every defended model, and which checkpoint (including epoch 0) was selected for each defense. Are the identical defended and undefended results due to the epoch-0 fallback?
  2. How do the defenses perform when adversarial examples are regenerated against the model being trained, with properly tuned training, and under several budgets?
  3. What is the cascade length distribution, and how many positions are actually changed per cascade? Can the budget be expressed as a fraction of cascade length?
  4. How detectable are the perturbed cascades, e.g., using the clean model’s likelihood of each participant given the prefix?
  5. In what realistic scenario can an attacker replace (rather than add) participants in an observed cascade, given that platforms keep logs of who actually reshared?
  6. How does Theorem 2 apply when, as Appendix A.2.3 states, the targets share no latent coordinate system with the source?
6 comments
AuthorsOpenReview ↗

Reply to Reviewer [1/3]

Thank you for identifying the importance of participant identities as an attack surface. The revision addresses the evaluation questions with clean/attacked results for all 25 baseline target-defense combinations, stronger online training, absolute and length-proportional budgets, a second dataset, likelihood detection and filtering, and label-blind cross-architecture experiments. These additions establish both the practical scope of PIP and the accuracy losses it produces under substantially broader evaluation.

1. Clean utility, robust accuracy, and checkpoint selection.

The revised Table A6 reports clean and attacked Hits@10 for all 25 target–defense combinations, with each relative hit decrease (RHD) computed against that checkpoint's own clean accuracy:

RHD=100H@10clean−H@10attackedH@10clean.\mathrm{RHD}=100\frac{\mathrm{H@10}_{\mathrm{clean}}-\mathrm{H@10}_{\mathrm{attacked}}}{\mathrm{H@10}_{\mathrm{clean}}}.

For example, AT on DyHGCN changes clean/attacked accuracy from 33.16%/25.38% to 33.41%/26.21%, yielding 21.56% RHD. On DisenIDP, the corresponding change is 34.52%/23.85% to 35.13%/25.71%, yielding 26.80% RHD. The table therefore separates clean utility from robustness recovery. Across the baseline defended checkpoints, clean accuracy changes by less than one percentage point, while RHD remains 21.56–30.91%.

The identical baseline entries contribute zero measured robustness gain. Checkpoint selection uses validation data and includes epoch 0. Table A10 marks the initial selections for the stronger-training comparisons: MS-HGAT with TRADES and MART, RotDiff with MART, and GODEN with TRADES. These entries represent selection of the initial model and are interpreted accordingly. The table separately presents the retained trained checkpoints and their clean/attacked outcomes.

2. Online adversarial training, stronger optimization, and training–test budget transfer.

The original baseline uses fixed source-generated training pairs. The expanded evaluation includes online white-box AT, which regenerates adversarial training inputs against the model being trained (the revised Tables A8–A9). Table A8 increases the learning rate from 2×10−52\times10^{-5} to 2×10−42\times10^{-4} or 10−310^{-3} and extends training to at most 20 epochs, with patience five. At test budget B=6B=6, tuned AT reduces RHD to 9.20% on DyHGCN, 9.54% on MS-HGAT, 11.97% on DisenIDP, 16.12% on GODEN, and 21.12% on RotDiff. Thus stronger training produces substantial architecture-dependent recovery, and the revised conclusions reflect that result.

Table A10 also strengthens TRADES and MART and reports their clean-accuracy costs. For DyHGCN, tuned AT reaches 30.04% attacked accuracy with 33.08% clean accuracy. Tuned MART reaches a smaller 6.22% RHD, but its clean/attacked accuracies are 27.72%/25.99%. Reporting both quantities makes the trade-off explicit: the smallest relative drop need not provide the best attacked accuracy.

The defense comparison now spans B=1,…,6B=1,\ldots,6 and four length-proportional budgets (Table A7). Online AT is also trained separately at Btrain=2B_{\mathrm{train}}=2 and 66 and evaluated at multiple test budgets (Table A9). at Btest=6B_{\mathrm{test}}=6, their RHD differs by at most 0.54 percentage points under the baseline schedule. The transfer evaluation uses the same frozen, target-independent test perturbations for the compared checkpoints. Training-time online optimization and test-time transfer evaluation are therefore reported as distinct experimental components.

3. Realized edits and length-proportional budgets.

Table A11 reports the number and share of positions actually changed. On Twitter, the nominal B=6B=6 attack makes 1,645 substitutions across 346 cascades: 4.75 edits per cascade on average and 20.0% of all recorded positions. At B=1B=1, the realized share is 4.2%, and mean target RHD is 5.87%.

We additionally evaluate the per-cascade budget min⁡(6,⌈ϵL⌉)\min(6,\lceil\epsilon L\rceil) for ϵ∈0.03,0.05,0.10,0.20\epsilon\in\\{0.03,0.05,0.10,0.20\\}. For nominal ϵ=3\epsilon=3\\%, integer rounding produces 479 substitutions, or a realized 5.8% of positions, and 11.07% mean RHD. All five target-specific 95% cascade-bootstrap intervals are positive. Reporting realized shares alongside the nominal ratio makes the effect of short cascades and integer rounding directly visible.

AuthorsOpenReview ↗

Reply to Reviewer [2/3]

4. Behavioral plausibility, likelihood detection, and filtering.

The new experiments directly test the proposed likelihood detector. With genuine participants from the same edited positions as negatives, target-model conditional-likelihood scores achieve AUC 0.72–0.80 on PIP, versus 0.84–0.90 on random feasible substitutions (Table A26). At 5% false positives, they detect 12.2–22.5% of PIP edits. These results quantify detectable structure while showing that likelihood-based rejection leaves most PIP edits undetected at that operating point.

We also evaluate detection as an actual defense: thresholds are calibrated on clean validation data, flagged observations are removed, and attacked accuracy is compared with the corresponding filtered clean accuracy. At nominal 5% false positives, mean RHD decreases from 25.84% to 21.24%; the filter flags 41.1% of edited identities, has 6.7% realized test false positives, and removes 1.46 genuine observations per clean cascade (Table A28). Under independently generated label-blind prefix attacks, filtering leaves 14.59–20.30 percentage points of target Hits@10 loss (Table A29). The evaluation thus extends beyond adversarial-training objectives to an observation-level detection and filtering defense.

Behavioral checks also support the constraints. PIP replacements have a 4.1% observed predecessor-transition rate, compared with 0.4% for random feasible replacements, and a smaller mean distance to recorded activity, 0.62 versus 0.67 days (Table A20). Regenerating attacks with a one-day activity window yields 25.47% mean RHD. Requiring one-hop substitutions under the defined transition-plus-social-neighbor relation yields 20.16% (Table A19). The vulnerability persists under substantially tighter candidate rules.

5. Dataset breadth, competitive searches, and the contribution.

The revised experiments include all 349 Douban test cascades. Mean RHD increases from 12.51% at B=1B=1 to 44.47% at B=6B=6, with a 95% interval of [41.9%, 47.0%] for the latter. The nominal 3% length-proportional budget gives 15.06% mean RHD (Table A12). In label-blind endpoint evaluation, using each of six architectures as the source on both datasets gives positive Hits@10-decrease estimates in all 60 cross-architecture comparisons, with 56 intervals excluding zero (Tables A23–A24).

Search comparisons now cover random proposals with exact verification, gradient-only substitution, fourfold candidate expansion, and a rank-10 margin objective (Tables A13–A17). PIP achieves 25.84% mean RHD using 320 exact source candidate evaluations per cascade, compared with 25.43% and 1,154 evaluations for random proposals with exact verification (Table A14). This is a 72.3% reduction in exact candidate evaluations at comparable aggregate damage. Matched per-step candidate caps further show a consistent gradient advantage on MS-HGAT, with gains of 5.54–8.86 RHD points and positive paired intervals (Table A15).

Yue et al. (2022) study gradient-guided item substitutions in interaction histories. HotFlip is also a relevant optimization precedent. PIP's contribution is the explicitly constrained robustness problem for identity-bearing diffusion records: a participant-position budget, fixed timestamps and background graph, training-derived activity/role/reachability constraints, and transferable attacks evaluated across forecasting architectures. The revised analysis also formalizes the identity-to-incidence representation of this problem (Appendix B.1) and cites such relevant references, making its connection to constrained graph rewiring precise.

6. Attacker control and the unchanged prediction label.

Sections 3.1–3.2 define an input-record integrity threat. A concrete setting is a forecasting service consuming a cascade feed through a compromised collection, identity-mapping, or preprocessing component: the adversary controls selected identity fields in the record delivered to the predictor. The platform's realized reshare history supplies the evaluation label, while the downstream input record is corrupted. Features derived from the edited prefix are recomputed consistently, and timestamps and the background social graph remain fixed.

The revision additionally evaluates attacks whose generator sees only the observed prefix. The next participant is available exclusively to the evaluator. Endpoint entropy attacks decrease Hits@10 on all five targets in both datasets, with positive 95% intervals (Table A22). Independent attacks at all 7,883 Twitter prediction prefixes, with a fresh six-edit budget for each prefix, cause 16.44–24.36 percentage points of target accuracy loss across the two label-blind objectives (Table A25). These experiments directly establish transfer damage without future-label access.

AuthorsOpenReview ↗

Reply to Reviewer [3/3]

7. Theorem 2, heterogeneous representations, and prediction damage.

Theorem 2 now uses a general linear map W∈Rdθ×dϑW\in\mathbb R^{d_\theta\times d_\vartheta} and κ=∣W∣2>0\kappa=\\|W\\|_2>0, allowing different representation dimensions. Its endpoint-alignment condition gives

∣Δhϑ∣2≥max⁡0,∣Δhθ∣2−2ϵκ.\\|\Delta h_\vartheta\\|_2\geq\frac{\max\\{0,\\|\Delta h_\theta\\|_2-2\epsilon\\}}{\kappa}.

The revised proof uses the reverse triangle inequality and operator norm. This is an explicit conditional displacement bound. Theorems 1 and 3 characterize the feasible search space and the acceptance rule. The connection to target loss is developed through directional gain, readout behavior, and finite-edit terms in Theorem 4 and Appendix B.2.

Appendix B.2 further works in the shared output coordinates of participant identities, which are available even when hidden-state dimensions differ. Proposition 8 provides an exact transported-loss identity and a sufficient condition for excluding the true participant from the target's top-kk set. The prediction-level conclusion is supported directly by the expanded Hits@10 experiments above.

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

This paper introduces Propagation Identity Perturbation (PIP), an adversarial attack setting for social event forecasting that replaces participant identities in observed propagation sequences. The authors argue that existing graph attacks mainly modify topology or features and overlook the role of participant identities in predicting future propagation. PIP replaces a few users while keeping the graph, timestamps, sequence length, and labels fixed. It restricts replacements using activity, propagation-position, and neighborhood constraints, then uses source-model gradients and loss evaluations to select effective substitutions. The resulting sequences are then evaluated on different target models without querying them during attack generation. The paper provides theoretical analyses of the search space, representation changes across models, and loss behavior. Experiments on a Twitter cascade dataset cover five forecasting models and five adapted defense methods. The reported results show that the attack transfers across the evaluated models and that the tested defenses provide limited recovery.

Critical strengths
  • Studying how changes to propagation participants affect forecasting is practically relevant, and the activity and neighborhood constraints make the attack setting more meaningful than unrestricted identity replacement.

  • The proposed threat model considers real-world constraints, making the studied adversarial robustness practical.

  • The same perturbed sequences are evaluated across diverse target models without target feedback during attack generation, providing useful evidence of transferability across the evaluated architectures.

Critical weaknesses
  • The paper does not sufficiently distinguish PIP from existing temporal graph attacks or explain why participant substitution constitutes a fundamentally new threat model.

  • The theoretical guarantees describe representation changes and source-model loss increases, but do not establish that predictions on other models will become worse.

  • The paper does not adequately discuss Continuous-Time-Dynamic-Graph models and attacks or justify their exclusion, while the limited coverage of temporal graph defenses and anomaly detection methods weakens its conclusions about existing defenses.

Questions for authors
  • How does PIP fundamentally differ from existing temporal graph attacks such as T-SPEAR (Lee et al., 2024) and MemFreezing (Dai et al., 2025)? Since participant substitution can be modeled as changing a temporal interaction’s endpoint, please clarify why this is a new threat model rather than a constrained attack on temporal interactions.

  • Could you clarify what the theoretical results guarantee about prediction performance? Theorem 2 bounds representation changes, while Theorem 3 guarantees source-loss increases through the acceptance rule. What connects these properties to worse predictions on target models, and do the evaluated architectures support the representation-alignment assumption?

  • What makes social event forecasting sufficiently different from the temporal prediction settings studied by TGN (Rossi et al., 2020), APAN (Wang et al., 2021), and JODIE (Kumar et al., 2019)? Please explain the choice of baselines and whether temporal graph defenses such as T-SHIELD (Lee et al., 2024), or suitable anomaly detection and graph-filtering methods, could be adapted to this setting. Such comparisons, or a clear explanation of their limitations, would help support the conclusion that existing defenses provide limited protection.

[1] Dongjin Lee, Juho Lee, and Kijung Shin. Spear and shield: Adversarial attacks and defense methods for model-based link prediction on continuous-time dynamic graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024.

[2] Yue Dai, Liang Liu, Xulong Tang, Youtao Zhang, and Jun Yang. MemFreezing: A novel adversarial attack on temporal graph neural networks under limited future knowledge. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pp. 12121–12160, 2025.

[3] Emanuele Rossi, Ben Chamberlain, Fabrizio Frasca, Davide Eynard, Federico Monti, and Michael Bronstein. Temporal graph networks for deep learning on dynamic graphs. arXiv preprint arXiv:2006.10637, 2020.

[4] Xuhong Wang, Ding Lyu, Mengjian Li, Yang Xia, Qi Yang, Xinwen Wang, Xinguang Wang, Ping Cui, Yupu Yang, Bowen Sun, and Zhenyu Guo. APAN: Asynchronous propagation attention network for real-time temporal graph embedding. In Proceedings of the 2021 International Conference on Management of Data, 2021.

[5] Srijan Kumar, Xikun Zhang, and Jure Leskovec. Predicting dynamic embedding trajectory in temporal interaction networks. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2019.

5 comments
AuthorsOpenReview ↗

Reply to Reviewer [1/2]

Thank you for recognizing the practical relevance of participant identity robustness and the value of evaluating identical perturbations across heterogeneous forecasting architectures. The revision directly addresses the relationship to temporal graph attacks, connects finite identity edits to target loss and ranking, and adds cross-dataset transfer and observation-filtering experiments.

1. Relationship to T-SPEAR and MemFreezing: an explicit threat model and an exact graph correspondence.

PIP studies the integrity of the observed identity record consumed by a cascade forecaster. Section 3.1 specifies the attacker-controlled fields: at most BB participant identities in an observed prefix. Event times, prefix length, the background social graph, and the logged next-event outcome remain fixed. Features derived from the edited prefix are recomputed consistently. This defines both the adversary's intervention and the prediction task against which robustness is measured.

The temporal graph interpretation is now formalized in Appendix B.1. Lemma 5 introduces one event node eie_i per position and the incidence encoding I(T)=(ei,ui)i=1kI(\mathcal T)=\\{(e_i,u_i)\\}_{i=1}^k, and proves

∣I(T′)△I(T)∣=2dH(T′,T).|I(\mathcal T')\triangle I(\mathcal T)|=2d_H(\mathcal T',\mathcal T).

Thus PIP has an equivalent constrained rewiring representation with budget 2B2B when the temporal, role, neighborhood, and candidate constraints are transported together. This equivalence makes comparisons precise: an identity-position budget corresponds to coordinated incidence edits, with one identity assignment retained per event. Arbitrary temporal-edge edits define a different feasible set.

T-SPEAR injects temporal-edge perturbations before victim training for continuous-time link prediction. MemFreezing injects fake nodes or edges to disrupt TGNN memory and sustain that disruption as the graph evolves. PIP targets identity attribution in an already observed forecasting prefix, under a fixed event schedule and fixed background graph. Its contribution is the formulation and systematic evaluation of this constrained input-integrity threat. Graph encodability provides an implementation of that threat in another representation. The explicit budget, feasible interventions, and forecasting objective determine what is being tested. We have added such relevant references in the revised manuscript.

2. What the theory establishes about target loss and prediction.

The revised analysis connects three quantities explicitly: representation displacement, target loss change, and top-kk ranking. Theorem 2 now permits a rectangular map W∈Rdθ×dϑW\in\mathbb R^{d_\theta\times d_\vartheta} and proves

\frac{\max\\{0,\\|\Delta h_\theta\\|_2-2\epsilon\\}}{\\|W\\|_2}.$$ Its proof uses $\\|W\\|_2>0$ and the stated alignment errors at both clean and attacked endpoints. This accommodates different latent dimensions, while retaining alignment as an explicit condition. Appendix A.2.3 describes the architectures' independently learned representations. The theorem states precisely the condition under which displacement transfers between them. Appendix B.2 supplies the connection to target loss on a fixed common eligible output set containing the true label, with finite logits and differentiable readouts. Proposition 6 gives the exact finite-edit decomposition $$\Delta\ell_m =v_m^\top\Delta h_m+C_m+\operatorname{KL}(p_m\Vert p'_m),$$ where $v_m$ is the clean loss gradient and $C_m$ accounts for nonlinear readout changes. Corollary 7 bounds target loss growth using source loss gain, displacement mismatch, gradient compatibility, and curvature. A positive lower bound guarantees target loss growth. These terms identify the conditions needed beyond a change in representation magnitude or monotonic source loss. More directly, Proposition 8 operates in the shared participant-output coordinates, allowing every model to retain its own hidden space. Write $\xi_{m,j}=\Delta z_{m,j}-\Delta z_{m,y}$, $\Gamma_{\theta\to\vartheta}=\log\sum_j p_{\vartheta,j}e^{\xi_{\theta,j}}$, and $r_j=\xi_{\vartheta,j}-\xi_{\theta,j}$. It proves $$\Gamma_{\theta\to\vartheta}-\max_j|r_j| \leq\Delta\ell_\vartheta\leq \Gamma_{\theta\to\vartheta}+\max_j|r_j|.$$
AuthorsOpenReview ↗

Reply to Reviewer [2/2]

It also gives a ranking condition: if ten eligible competitors have transported gaps zϑ,j−zϑ,y+ξθ,jz_{\vartheta,j}-z_{\vartheta,y}+\xi_{\theta,j} larger than their maximum residual magnitude, the attacked target excludes yy from its top ten, provided these competitors remain natively eligible after perturbation. For a clean hit, this certifies a hit-to-miss transition. Proposition 9 extends the residual bound to simultaneous coverage over a new cascade under exchangeability, with the models, attack protocol, and evaluation rule fixed. These are conditional performance statements. Target quantities used to assess the conditions are measured after the source-only attack has been fixed.

The experiments directly assess the corresponding predictive effects. The new six-source evaluation covers all ordered cross-architecture transfers on Twitter and Douban: all 60 transfer comparisons have positive H@10-decrease estimates, and 56 confidence intervals exclude zero (Tables A23–A24). The endpoint label is withheld from attack generation. On all 7,883 independently attacked Twitter prefixes, each with its own six-edit budget, the five targets lose 16.44–24.36 H@10 percentage points across two label-blind objectives (Table A25). Mechanistically, actual displacements produce 1.44–10.35 times the mean loss increase of norm-matched random directions (Table A35), while the exact decomposition also captures finite-edit effects beyond the first-order term. A separate experiment uses legal identity edits matched by cascade, edited position, one-edit budget, and source displacement magnitude. Higher source direction increases mean target loss in all ten dataset-target pairs, with nine paired intervals excluding zero (Table A36). This provides a complementary test in the original discrete attack space.

3. Continuous-time models, baseline selection, and temporal defenses.

TGN, APAN, and JODIE are relevant temporal representation frameworks: TGN combines memory and graph operators for timed event streams, APAN separates asynchronous propagation from inference, and JODIE learns evolving entity embeddings from temporal interactions. Applying them here requires an explicit cascade encoding, cascade-conditioned next-participant readout, and a consistent evaluation and eligibility protocol. Such an adaptation is feasible, and the event–identity construction in Appendix B.1 specifies how to preserve the PIP intervention when using graph coordinates. We have added such relevant references in the revised manuscript.

Our benchmark selects models that directly implement the cascade-prefix-to-next-participant task. RotDiff, DyHGCN, MS-HGAT, DisenIDP, and GODEN cover hyperbolic geometry, dynamic heterogeneous graphs, sequential hypergraphs, disentangled influences, and continuous-time ODE dynamics, respectively (Appendix A.2). GODEN directly includes continuous-time modeling in the evaluated architectures. The expanded six-source experiments additionally test transfer with each architecture serving as the source.

T-SHIELD combines edge filtering with temporal embedding regularization [1]. Its filtering component can be adapted to the observed identity sequence by detecting suspicious participant records and recomputing predictions after removal. For the PIP interface, Appendix A.10 evaluates a likelihood-based observation filter with thresholds calibrated on clean validation data. At a nominal 5% false-positive threshold, mean RHD decreases from 25.84% to 21.24%, measured against each filtered model's own clean accuracy (Table A28). On independently attacked prefixes, all five targets retain 14.59–20.30 percentage points of H@10 loss after filtering, again relative to filtered clean predictions (Table A29). These experiments quantify both the protection delivered by a concrete anomaly filter and the substantial identity-attack damage that remains. The broader defense evaluation also adds online adversarial training and stronger TRADES/MART settings (Appendix A.6), supporting conclusions about the evaluated defenses under explicit training and attack protocols.

AuthorsOpenReview ↗

Deadline for Discussion is Approaching in Less than 2 Days

Dear reviewer PGZr,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 2 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

Reviewer_Kq1wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Propagation Identity Perturbation (PIP), a new adversarial threat model for social event forecasting that investigates the vulnerability of forecasting models to changes in propagation participant identities. Unlike conventional graph attacks targeting topology or node features, PIP replaces a limited number of participant identities while preserving propagation timestamps, sequence length, and the original graph structure. To maintain behavioral plausibility, the authors introduce temporal consistency, propagation position consistency, and neighborhood consistency constraints, combined with a gradient-guided discrete search strategy to generate transferable adversarial examples. The paper further provides four theoretical results and evaluates the proposed attack across five forecasting architectures and five defense strategies, demonstrating its effectiveness and highlighting limitations of existing defenses.

Critical strengths

Strength 1: Meaningful problem formulation and practical relevance.

The paper shifts the focus of adversarial attacks from conventional graph topology and node feature perturbations to propagation participant identities, highlighting a relatively underexplored vulnerability in social event forecasting. This perspective is well aligned with the characteristics of temporal diffusion tasks and provides useful insights into the security risks of propagation forecasting models.

Strength 2: Clear methodology and promising empirical results.

The proposed method constructs a constrained candidate space based on temporal, propagation position, and neighborhood consistency, and combines gradient-guided search with exact loss verification. Experiments across five different forecasting architectures demonstrate the transferability of PIP and its effectiveness against several existing defense strategies, providing empirical support for the proposed threat model.

Critical weaknesses

Weakness 1: Limited methodological novelty.

Although the paper introduces a new threat model based on propagation identity perturbations, the core attack algorithm largely follows the existing HotFlip framework for gradient-guided discrete substitutions. The proposed behavioral plausibility constraints primarily rely on temporal, positional, and neighborhood statistics for candidate filtering. Consequently, the methodological contribution appears to lie mainly in adapting and combining existing techniques for a new attack setting, rather than introducing a substantially novel optimization strategy.

Weakness 2: Insufficient experimental validation and competitive baselines.

The experiments are conducted on a single Twitter dataset, which limits the evidence for generalizability across different diffusion environments. Furthermore, while the paper demonstrates attack effectiveness across multiple forecasting models and defense strategies, it lacks systematic comparisons with stronger discrete substitution attacks. Although random legal replacement is included in the ablation study, the current evaluation does not sufficiently establish the advantages of the proposed optimization strategy over existing alternatives.

Questions for authors
  1. Novelty: The core search strategy of PIP appears closely related to HotFlip. Could the authors clarify the main algorithmic innovations beyond constructing a behaviorally constrained candidate space?
  2. Experimental: Could the authors include comparisons with other gradient-guided discrete attacks or greedy search strategies? It would also be helpful to disentangle the contributions of behavioral constraints, gradient-based candidate selection, and exact loss verification.
  3. Practical attack feasibility: The proposed candidate constraints rely on historical activity, propagation position, and neighborhood statistics. Are these conditions sufficient to ensure that the resulting propagation trajectories are feasible in real-world social networks? Additionally, what level of adversarial control is assumed for modifying participant identities?
5 comments
AuthorsOpenReview ↗

Reply to Reviewer [1/2]

Thank you for recognizing the importance of propagation identities and the evidence of transfer across forecasting architectures. The revision directly addresses your three questions through controlled search comparisons, a second diffusion dataset, and an explicit input-record threat model. These additions show which components improve attack effectiveness, how efficiently they operate, and what adversarial access the setting requires.

1. Contribution and relationship to HotFlip.

PIP formalizes robustness of the ordered participant record consumed by a social event forecaster. Its feasible set combines an identity-position budget with temporal activity, propagation-role, neighborhood, and identity-uniqueness constraints, while preserving event times and the background social graph. Changing a participant record also changes the prefix-derived structures and states used by different forecasting architectures. This provides a common attack interface for evaluating transfer without target feedback.

We explicitly use and cite HotFlip's first-order substitution score, gi⊤(E(v)−E(ui))g_i^\top(E(v)-E(u_i)). The PIP optimizer combines that score with a budgeted proposal-and-verification procedure: it selects four legal users per mutable position, compares up to 128 position–identity pairs by exact source-model loss, accepts only a positive exact improvement, and permits further refinement of already modified positions after the identity budget is filled (Section 3.4 and Algorithm 1). Thus gradient ranking allocates the candidate-evaluation budget, while exact verification handles the finite and discrete changes produced by complete identity substitutions. Every search decision uses the source alone, and the resulting identity record is shared across all targets.

The contribution is the formulation of this constrained forecasting vulnerability, its effective source-only instantiation, and the systematic evidence of transfer. The new component comparisons quantify the benefits of the optimizer within that formulation.

2. Stronger comparisons and separation of constraints, proposals, and verification.

Appendices A.7–A.8 add the requested controls across all five transfer targets.

Position and identity selection. Table A13 holds the total number of edited identities fixed at 1,645. With the same random feasible identity sampler, using PIP's positions increases RHD by 3.1–8.4 points over random positions. The complete attack achieves 1.27–2.29 times the RHD of random feasible position–identity replacement across the five targets. This isolates the contribution of selecting influential positions. Optimizing the replacement identities then improves four of the five target results.

Alternative search strategies. Table A14 compares the full procedure with gradient-only substitutions, random proposals followed by exact verification, fourfold candidate expansion, and a rank-10 margin objective:

Search procedureMean target RHDExact source candidate evaluations per cascade
PIP25.84%320
Random proposals + exact verification25.43%1,154
Gradient only23.19%0
Fourfold exact-candidate expansion26.08%1,262
Rank-10 margin objective25.97%340

These results establish a useful effectiveness–evaluation-cost tradeoff. Gradient proposals reach 25.84% mean RHD using 72.3% fewer exact candidate evaluations than random proposals. Exact verification raises mean RHD from 23.19% to 25.84%, while fourfold candidate expansion adds 0.24 points. The margin objective supplies a strong alternative and shows that the vulnerability persists under a different attack objective. These evaluation counts concern source-model candidate forwards. The transfer attacks use zero target queries.

Matched proposal budgets. Table A15 separately compares gradient and random proposals under identical per-step exact-candidate caps K∈8,16,64K\in\\{8,16,64\\}. Gradient proposals improve MS-HGAT RHD by 8.74, 8.86, and 5.54 points, respectively, with positive paired 95% intervals. GODEN improves by 2.90 points at K=8K=8. The architecture dependence is visible in the complete table. Tables A16–A17 extend the comparison to absolute and length-proportional edit budgets using saved attack trajectories, showing how the competing search procedures behave as the allowed edits change.

Behavioral constraints. Table A18 regenerates attacks after removing each constraint, applying each individually, and removing all three. The complete constraint set retains 92.4% of unconstrained mean RHD: 25.84% versus 27.97%. Table A19 further tightens the feasible set: a one-day temporal window gives 25.47% mean RHD, and one-hop substitutions under the defined transition-plus-social-neighbor relation give 20.16%. These results quantify the cost of restricting the attack to behaviorally supported candidates while preserving substantial transfer damage.

AuthorsOpenReview ↗

Reply to Reviewer [2/2]

3. Generalization beyond one dataset and one source.

The revision evaluates all 349 Douban test cascades, comprising 5,627 prediction events. In Table A12, mean RHD increases from 12.51% at B=1B=1 to 44.47% at B=6B=6, with a paired 95% interval of [41.9%, 47.0%] for the latter. A length-proportional budget min⁡(6,⌈0.03L⌉)\min(6,\lceil0.03L\rceil) gives 15.06% mean RHD. Its realized edit share is 6.8%, reported to account for rounding on short cascades.

We also add a separate label-blind evaluation (Appendix A.9). The attack generator receives only the observed history, with the held-out endpoint identity and timestamp removed. The evaluator alone receives the endpoint label. Under entropy maximization, H@10 decreases by 4.62–11.56 percentage points across the five Twitter targets and 3.15–4.87 points across the five Douban targets, with every corresponding 95% interval above zero (Table A22). Tables A23–A24 then use each of the six architectures as source: all 60 cross-architecture decrease estimates are positive, and 56 intervals exclude zero. This extends the evidence across datasets, source choices, and a protocol that requires no unknown next-event label during attack generation.

4. Practical feasibility and required adversarial control.

Section 3.1 now states the access model explicitly: the adversary can corrupt at most BB participant-identity fields in the observed record before it reaches the forecaster. Event times, the background graph, and the logged next-event outcome remain fixed. Features derived from the edited prefix are recomputed consistently. The security boundary is the integrity of the predictor's input record. A compromised identity-mapping or preprocessing component in the forecasting input pipeline is one example consistent with this access model. An adversary's ability to alter that record is the required capability. Access to target parameters or prediction queries is unnecessary for generating the transferred attacks.

The candidate constraints provide specific and checkable behavioral requirements: training-observed activity near the event time, participation in the relevant propagation-position bin, supported neighborhood reachability, and exclusion of repeated identities. Their empirical implications are measured in Table A20. PIP replacements have mean nearest-activity distance 0.62 days, compared with 0.77 days for the genuine identities at those positions. Their observed predecessor-transition rate is 4.1%, versus 0.4% for random feasible replacements and 8.8% for genuine participants. These measurements establish behavioral support under the stated candidate rules. Appendix A.9 distinguishes observation corruption, which holds the latent history fixed, from physical participant interventions, which recompute the future distribution. PIP evaluates the former: the realized cascade provides the outcome, and the attack changes the identity record presented to the predictor.

AuthorsOpenReview ↗

Deadline for Discussion is Approaching in Less than 2 Days

Dear reviewer Kq1w,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 2 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.