Let prompts bridge defense knowledge: the transferable graph purification method ProGAP based on vulnerability-aware GPL
Let Prompts Bridge Defense Knowledge: Transferable Graph Purification via Vulnerability-Aware GPL
The authors propose a graph adversarial purification method, ProGAP, which uses vulnerability-aware graph prompt learning to transfer defense knowledge, and the authors claim it outperforms 8 SOTA baselines in 78% of evaluation scenarios.
Two attack scenarios: gray-box global poisoning attack (the attacker can see the graph data and labels, performs structural perturbation by inserting or deleting edges, and maximizes damage to the downstream classifier) and white-box adaptive attack (the attacker has all details of features, graph structure, labels, and the defense model, including the pretrained edge detector, prompts, and downstream classifier).
- Existing graph adversarial purification methods are limited to single-domain defense: single-domain data provides insufficient coverage of vulnerability patterns, and changing domains requires retraining from scratch with high computational cost, making it difficult to transfer to cross-domain multimedia graph scenarios.
- Proposes ProGAP: first pretrain a perturbation-capturing edge detector that jointly uses topological and semantic information on a data-rich upstream graph, then inject vulnerability-aware prompts into vulnerable nodes such as low-degree and low-similarity nodes in the downstream graph, freeze the detector, and optimize only the prompts to complete purification.
- Compared with 8 baselines on 10 graph datasets under multiple attacks: the authors state that ProGAP achieves the best results in 78% of scenarios, and in Table 1, 22 of 28 adversarial settings are best; PubMed runtime drops from 22.85s to 10.11s; it also outperforms the baselines under adaptive attacks.
- The authors state that future work will explore the application of this approach in more complex, dynamic multimedia scenarios (Conclusion). The experiments cover 10 datasets, 4 attacks, and a perturbation rate of up to 50%; the paper does not report hyperparameter details such as the number of iterations for pretraining and prompt optimization.
- Evaluated models
- ProGAP
- Benchmarks
- CoraCiteSeerCoraMLPubMedComputersPhotoChameleonSquirrelCornellWisconsin
- Metrics
- accuracyrunning time
ProGAP is a transferable graph purification scheme that bridges adversarial defense knowledge through vulnerability-aware graph prompt learning (GPL), used to defend against graph adversarial perturbations on GNNs. It first pretrains a perturbation-capturing edge detector on data-rich graphs, then uses vulnerability-aware prompts to inject purification guidance into biased nodes, enabling the detector to adapt to downstream graph distribution shifts without heavy parameter updates. Experiments show that compared with current best-performing baselines, ProGAP improves by 1%-9% and reduces time consumption by up to 2.2 times.
Paper analysis
What problem does this paper attempt to solve?
Existing graph adversarial purification methods are limited to a single domain; the authors propose ProGAP, which can transfer defense knowledge across domains.
Core problem: graph adversarial purification methods can only be used in a single domain and cannot transfer across domains, resulting in insufficient coverage of vulnerability patterns and high retraining costs.
- Problem scenario: GNNs are widely used in multimedia tasks such as cross-platform user interest modeling and cross-modal semantic alignment, but their message propagation mechanism depends on graph structure; the authors state that small structural perturbations can severely damage performance, so purification defense that removes adversarial edges before message propagation is needed.
- Shortcomings of existing methods: The authors point out two: (1) vulnerable structures appear at limited frequency in single-domain data, so purifiers cannot learn robust criteria for distinguishing adversarial perturbations from normal structural changes; (2) changing domains requires retraining from scratch and repeatedly generating adversarial examples, with high computational cost.
- Core observation: The authors believe that adversarial perturbations have node-level heterogeneity and are concentrated on specific vulnerable nodes such as low-degree and low-neighbor-similarity nodes, so defense guidance should target these nodes rather than being applied uniformly.
- Paper's approach: Proposes ProGAP, which pretrains a perturbation-capturing edge detector on a data-rich upstream graph, then uses vulnerability-aware graph prompt learning (vulnerability-aware GPL) to transfer defense knowledge to the downstream graph, freezing the detector and optimizing only the prompts.
- Threat model (Section 3.2):
- Attacker's goal: perform structural perturbation by inserting or deleting edges to maximize damage to the overall performance of the downstream classifier.
- gray-box global poisoning attack: the attacker can see the graph data and labels and generates perturbations based on a surrogate GNN.
- white-box adaptive attack: the attacker has all details of features, graph structure, labels, and the defender's model (pretrained edge detector, prompts, downstream classifier), with the goal of bypassing prompt-guided purification.
- Victim system: the downstream GNN classifier on the graph.
What related research exists?
Related work is divided into three groups: graph adversarial defense, graph knowledge transfer, and cross-domain graph prompt learning; the authors state that none of them can transfer defense knowledge across domains.
Graph adversarial defense
- GCN-SVD (2020): Performs low-rank approximation on the graph to purify the structure.
- GPR-GAE (2025): Separates robustness from the classifier and uses multiple Generalized PageRank filters for graph purification.
- ProGNN (2020), GADC (2024): Abandon the purification route and instead design robust GNN architectures; the former jointly learns the low-rank, sparse, and feature-smoothness properties of the graph, and the latter proposes min-max optimization based on Laplacian distance perturbation. The paper points out that the limitation of these methods is that they can only defend in a single domain and cannot transfer across domains.
Graph knowledge transfer
- Pretraining + fine-tuning paradigm: Pretraining a GNN on a source graph and then fine-tuning it on a target graph is the mainstream paradigm; the paper points out that full fine-tuning is inefficient and prone to overfitting and catastrophic forgetting.
- Graph prompt learning: Increasingly adopted as a lightweight transfer mechanism to reduce resource requirements.
Cross-domain graph prompt learning
- GPPT (2022): Converts independent nodes into token pairs.
- GPF / GPF-Plus (2023): Introduces learnable parameters in the input graph feature space as generic prompts.
- All-in-One (2023), MultiGprompt (2024): The former extends independent tokens into a prompt graph, and the latter introduces multi-task pretraining and a dual-prompt mechanism. The paper points out that these methods all focus on the compatibility between pretraining and downstream tasks, whereas this paper uses vulnerability-aware prompts to achieve cross-domain transfer of defense knowledge.
Baselines and benchmarks
- Defense baselines: GNNGuard, GCN-SVD, JaccardGCN, HANG-quad, GADC, GPR-GAE, GRAND, Soft-Median-GDC, 8 in total; the GPL comparison baselines are GPF and GPF-Plus (selected based on the benchmark work ProG (2024)).
- Datasets: Six homophilic graphs: Cora, CiteSeer, CoraML, PubMed, Computers, Photo; four heterophilic graphs: Chameleon, Squirrel, Cornell, Wisconsin.
The authors position this paper as: the first approach to use graph prompt learning for cross-domain transfer of adversarial defense knowledge—the original source states that existing prompt methods focus on task compatibility, whereas this paper uses vulnerability-aware prompts for defense knowledge transfer.
How does the paper address this problem?
ProGAP has three steps: pretrain a perturbation-capturing edge detector upstream, generate vulnerability-aware prompts downstream, and then purify the graph under prompt guidance.
Overall approach: first pretrain a perturbation-capturing edge detector that jointly models topology and semantic information on a data-rich upstream graph, then identify vulnerable nodes on the downstream graph and inject vulnerability-aware prompts, freeze the detector, and optimize only the prompts and the classification head, achieving cross-domain transfer of defense knowledge (Figure 2).
General adversarial knowledge awareness (Section 4.2)
- Perturbation construction: use a PGD attack to perturb the clean graph to obtain Ĝ, and assign binary labels to edges: edges inserted by the attacker are labeled 0, retained normal edges are labeled 1. The authors state that PGD is a representative gradient attack that can expose general adversarial patterns.
- Dual evaluators: the topology evaluator D_t uses a GCN to learn structural representations, concatenates the original node features, and outputs through a linear layer the probability t_uv that an edge is an adversarial edge; the semantic evaluator D_s scores s_uv using the cosine similarity of node features.
- Attention fusion: the final edge score is duv = α1 · tuv + (1−α1) · suv, where α₁ is a learnable weight; the detector is trained with a cross-entropy loss between predicted scores and edge labels.
Vulnerability-aware prompt generation (Section 4.3)
- Vulnerable node identification: first estimate the homophily of the whole graph as H=1/(|E|)∑(u,v)∈ E sim(Xu,Xv) (sim is cosine similarity), then adaptively set a threshold based on it to find the low-degree node set S_d and the low-similarity node set S_s.
- Biased node expansion: compute the ratio of a node's connections to normal nodes as the connection score; nodes below the thresholds τ_ld and τ_ls are recorded as biased node sets S_d′ and S_s′, i.e., nodes that mainly connect to vulnerable nodes are also included.
- Prompt injection: the degree-vulnerability prompt token p_d is added to nodes in S_d′, and the similarity-vulnerability prompt token p_s is added to nodes in S_s′; nodes belonging to both sets weight multiple prompts with a learnable fusion weight.
Prompt-guided graph purification (Section 4.4)
- Purification: the prompt-enhanced graph is fed into two pretrained evaluation modules; edges with scores below the threshold τ_edge are judged as adversarial edges and removed, yielding the purified graph G_p.
- Pseudo-labels: a structure-agnostic MLP is introduced to generate high-confidence pseudo-labels for a small number of unlabeled nodes, which are merged with labeled nodes into the augmented training set V_aug; the authors' stated reason is that adversarial perturbations disrupt local structure but the global feature distribution is relatively stable.
- Optimization: minimize the classification loss Lcls=−∑ui∈ Vaug yi log fθ1(Gp); gradients are backpropagated to both the prompts {p_d, p_s} and the classification head θ₁, while the pretrained detector remains frozen.
Success criteria and evaluation method
- Criteria: defense effectiveness is measured by downstream node classification accuracy (% ± σ), and efficiency by running time (seconds).
- Attack settings: non-adaptive attacks use MetaAttack, GraD, and Heuristic attack with perturbation rates of 25% and 50%; the adaptive attack uses Greedy-Grad-Descent, in evasion and poisoned variants.
What experiments did the paper conduct?
Evaluated on 10 datasets, 8 defense baselines, and 4 attacks; the authors state that ProGAP is best in 78% of scenarios and also takes less time.
Experimental setup
- Datasets: 10 graph benchmarks—homophilous graphs Cora, CiteSeer, CoraML, PubMed, Computers, Photo, and heterophilous graphs Chameleon, Squirrel, Cornell, Wisconsin; the paper does not report the size of each dataset.
- Upstream pretraining settings: ProGAP-Pu, ProGAP-C, and ProGAP-Ph use PubMed, Computers, and Photo respectively as upstream pretraining datasets, corresponding to the scenario where the model provider pretrains on a large graph and the user performs GPL on a cross-domain graph.
- Baselines: 8 defense methods GCN-SVD, JaccardGCN, GNNGuard, HANG-quad, GADC, GPR-GAE, GRAND, Soft-Median-GDC, plus GCN without defense; GPL comparisons are GPF, GPF-Plus, and Without Prompt.
- Attacks: non-adaptive are MetaAttack (M), GraD (G), Heuristic (H), with perturbation rates 25%/50%; adaptive are Greedy-Grad-Descent evasion and poisoned attack.
- Metrics: node classification accuracy (% ± σ, the paper does not specify the number of repetitions) and total running time (seconds).
Main results
Table 1 (accuracy %, 25%/50% perturbation, the three ProGAP variants take the ProGAP-Pu column with PubMed as upstream):
Swipe horizontally to read the table
Dataset Attack GCN GCN-SVD GADC GPR-GAE ProGAP-Pu Cora M-50 35.23 44.13 68.15 65.20 77.31 Cora G-50 30.61 52.91 68.94 65.59 74.96 CiteSeer M-50 41.15 52.03 61.01 64.99 69.70 CoraML H-50 66.84 64.88 73.47 72.18 77.51 PubMed H-50 63.53 81.62 52.59 82.22 84.08 PubMed M-25 70.53 82.62 75.40 83.97 83.64 (According to Table 1, only some settings are listed; the JaccardGCN, GNNGuard, and HANG-quad columns are omitted.)
- Authors' interpretation: ProGAP achieves the best in 22 of 28 adversarial settings, with a more pronounced advantage at higher perturbation rates; the abstract and introduction state that it outperforms 8 SOTA baselines in 78% of evaluation scenarios, with improvements of 1%-9%.
- Exceptions: on clean graphs, GCN-SVD and GNNGuard are slightly better than ProGAP; GADC and GPR-GAE exceed ProGAP-Pu in individual settings (e.g., PubMed M-25), which the authors describe as unstable across datasets.
- Cross-domain: the authors state that even when the upstream is a product-sales dataset (Computers, Photo), downstream still achieves SOTA performance, and argue that what the upstream learns is domain-agnostic representations.
GPL transfer effect (RQ2, Figures 3, 6)
- What was measured: on Cora, CiteSeer, CoraML, PubMed at a 25% perturbation rate, ProGAP is compared with GPF-Plus, GPF, and Without Prompt (Figure 3 is a bar chart; the converted text has no numerical values).
- Results: the authors state that pretrained models with GPL transfer generally outperform directly applied pretrained models, and vulnerability-aware prompts outperform generic prompts; in the embedding visualization of Figure 6, the authors state that GPL brings larger inter-class separation S and smaller intra-class variance CV on CiteSeer and Cora.
- Authors' interpretation: generic prompts only perform feature-space adaptation, whereas the prompts in this paper target vulnerable biased nodes.
Efficiency (RQ3, Table 2)
- What was measured: total running time of each defense method on four datasets. ProGAP is 4.09s on Cora, 5.15s on CiteSeer, 5.93s on CoraML, and 10.11s on PubMed; on PubMed, HANG-quad is 22.85s, GPR-GAE 53.46s, GCN-SVD 169.53s, and GADC 13.10s.
- Authors' interpretation: the efficiency comes from prompt learning reducing the parameters to be updated to a single token, and the transferable detector eliminating repeated perturbation generation per graph; the abstract states a reduction in time of up to 2.2×.
Adaptive attacks (RQ4, Table 3)
- What was measured: accuracy under white-box adaptive evasion and poisoned attacks, where the attacker has full knowledge of the defense details.
- Results: ProGAP is highest in all 8 settings, e.g., under evasion, Cora 71.49%, CiteSeer 65.97%, PubMed 84.00%, CoraML 73.85%; under poisoned, Cora 72.14%, CiteSeer 67.69%, PubMed 84.64%, CoraML 75.60%. GNNGuard and Soft-Median-GDC are OOM on PubMed.
- Authors' interpretation: even when the attacker has complete information about the defense mechanism, the approach can still effectively reconstruct the adjacency matrix.
Other ablations and analyses
- Component ablation (Table 4, G-25/M-25): removing D_t or D_s causes the largest drop (e.g., Cora G-25 drops from 76.94% to 60.50% and 65.66% respectively); removing attention causes a moderate drop; removing pseudo-labels causes only a slight drop; removing p_s or p_d both cause drops.
- Homophily recovery (Figures 4, 5): the authors state that after purification, the homophily of both node features and embeddings improves.
- Heterophilous transfer (RQ6, Figure 7): transferring from a homophilous reference network to Chameleon and Squirrel, under Metattack 10%/25% perturbation, the authors state that ProGAP outperforms baselines (the figure is a bar chart; the converted text has no numerical values).
What are some points that can be further explored?
The authors only propose future exploration of more complex and dynamic multimedia scenarios; the experiments cover 10 datasets and 4 types of attacks.
Limitations and future directions pointed out by the authors
- Future direction: The authors plan to explore the application of this approach in more complex, dynamic multimedia scenarios in the future (Conclusion). The paper does not specifically discuss other limitations.
Scope of experimental coverage
- Subjects tested: 10 graph datasets (6 homophilic, 4 heterophilic), 8 defense baselines, 2 GPL baselines; the downstream task is node classification.
- Attack coverage: 3 non-adaptive attacks (MetaAttack, GraD, Heuristic) with perturbation rates of 25% and 50%; 1 adaptive attack (Greedy-Grad-Descent), divided into evasion and poisoned; heterophilic transfer experiments use Metattack with perturbation rates of 10% and 25%.
- Main experiment scope: The main results in Table 1 cover four datasets: Cora, CiteSeer, CoraML, and PubMed; Computers and Photo appear only as upstream pretraining datasets in variant names, and Cornell and Wisconsin are listed in the dataset list but do not appear in the results tables.
- Adaptive attack scope: Table 3 evaluates only on Cora, CiteSeer, PubMed, and CoraML, and the baselines include only GCN, GCN-SVD, GNNGuard, GRAND, and Soft-Median-GDC.
- Not reported in the paper: hyperparameters such as the number of iterations and learning rate for pretraining and prompt optimization, the number of experimental repetitions (only ± σ is given), and the specific values in Figures 3 and 7.
Summarize the main content of the paper
ProGAP uses vulnerability-aware graph prompt learning to transfer graph adversarial defense knowledge across domains; the authors state it outperforms SOTA in 78% of scenarios.
A graph adversarial defense paper that proposes a transferable graph purification approach, ProGAP: existing purification methods can only be used in a single domain, have insufficient coverage of vulnerability patterns, and require retraining from scratch when changing domains.
- Method: Upstream, PGD perturbation is used to construct edge labels, and a perturbation-capturing edge detector that jointly uses a topological evaluator and a semantic evaluator (attention fusion) is pretrained; downstream, low-degree and low-similarity nodes and their connected biased nodes are adaptively identified according to graph homophily, and degree vulnerability prompts p_d and similarity vulnerability prompts p_s are injected; the detector is frozen, only the prompts and classification head are optimized, and the training set is expanded with pseudo-labels from a structure-agnostic MLP.
- Threat model: gray-box global poisoning attack and white-box adaptive attack that has all details of the defense; the attack method is inserting or deleting edges.
- Main results: The authors state that it outperforms 8 SOTA baselines in 78% of evaluation scenarios, with an improvement of 1%-9%; in Table 1, 22 of 28 adversarial settings are best, e.g., under Cora M-50, ProGAP-Pu 77.31% vs. GADC 68.15%.
- Efficiency: Total runtime on PubMed is 10.11s, lower than HANG-quad's 22.85s and GPR-GAE's 53.46s (Table 2); the abstract states that time is reduced by up to 2.2 times.
- Adaptive attack: In Table 3, all 8 settings are highest, e.g., PubMed 84.64% under poisoned attack.
- Authors' conclusion: Lightweight prompts can serve as a cross-domain transfer carrier for adversarial topological knowledge, and vulnerability-aware prompts achieve more robust transfer than generic prompts; the authors plan to explore more complex, dynamic multimedia scenarios in the future.
Related papers
- RGDM: Robust Graph Diffusion ModelDefensesarXiv