Foundation Graph Anomaly Detection

• Foundation Graph Anomaly Detection is a universal framework that generalizes anomaly detection across diverse graph structures and domains.
• It employs multi-task graph autoencoders, contrastive learning, and prompt-based alignment to achieve robust zero-shot and few-shot inference.
• The paradigm addresses node, edge, subgraph, group, and full-graph anomalies, validated by benchmarks using metrics like AUROC and AUPRC.

Foundation graph anomaly detection is a paradigm that seeks to enable graph-based anomaly detectors to generalize across disparate domains, granularities, and anomaly distributions, using a single model or unified learning pipeline. Unlike conventional graph anomaly detection (GAD) models that are tailored to specific data sets and assume aligned train-test distributions, foundation GAD aims for universal detection capability across unseen domains, unseen anomaly types, changing data distributions, and multiple anomaly granularities (node, edge, subgraph, group, whole-graph) (Pan et al., 8 Sep 2025).

1. Evolution from Task-specific to Foundation Paradigms

Traditional GAD methods are defined by a dataset-specific, single-task objective: given a set of graphs $\mathcal{D} = \{\mathcal{G}_1,...,\mathcal{G}_N\}$, learn an anomaly scoring function $f_\theta: o \mapsto s \in \mathbb{R}$ so that $s=f_\theta(o)$ is high for anomalous samples under the stipulated single-domain assumption $\mathcal{D}^{train} = \mathcal{D}^{test}$ (Pan et al., 8 Sep 2025). In contrast, foundation GAD aims to learn a universal function $f_\Theta: o \mapsto s$ that generalizes to any $o \in \mathrm{Samples}(\mathcal{D})$ for any $\mathcal{D} \in \mathcal{T}^{te}$, where $\mathcal{T}^{te}$ consists of previously unseen test graphs—potentially from heterogeneous feature spaces, domains, or anomaly distributions. This shift from one-model-per-dataset to one-model-for-all enforces new demands for transferability, domain adaptation, and scalability.

Early approaches to GAD generalization leveraged transfer learning (source $\to$ target) but typically with aligned feature spaces and restricted semantic shifts. Foundation models unite multi-domain, multi-task training—often via shared encoders and decoders—and target adaptation via zero-shot or few-shot mechanisms (Pan et al., 8 Sep 2025). This progression is treated formally via the taxonomy of generalization in GAD, differentiating:

• Conventional GAD: model is confined to one dataset.
• Transfer GAD: pretrain on $\mathcal{D}^s$, fine-tune on $f_\theta: o \mapsto s \in \mathbb{R}$0.
• Foundation GAD: pretrain across diverse $f_\theta: o \mapsto s \in \mathbb{R}$1; inference on $f_\theta: o \mapsto s \in \mathbb{R}$2 without retraining.

2. Foundational Model Architectures and Learning Objectives

Foundation graph anomaly detection employs a range of architectural innovations to realize cross-domain and cross-granularity generalization:

• Multi-task graph autoencoders (GAE): A shared encoder $f_\theta: o \mapsto s \in \mathbb{R}$3 feeds into task-specific decoders for nodes, edges, or graphs. The pretraining loss integrates reconstruction of all granularities,

$f_\theta: o \mapsto s \in \mathbb{R}$4

Anomaly scores are based on reconstruction errors, which serve as domain-agnostic anomaly signals (Pan et al., 8 Sep 2025).

• Contrastive learning and cross-view models: Generate two graph views $f_\theta: o \mapsto s \in \mathbb{R}$5 and optimize a contrastive loss to align their embeddings. Embedding inconsistency serves as the anomaly metric:

$f_\theta: o \mapsto s \in \mathbb{R}$6

• Prompt-based and prototype-based alignment: Foundation models such as AnomalyGFM (Qiao et al., 13 Feb 2025), ARC, and UNPrompt use learnable class-independent prototypes or prompt vectors to align node-level representation residuals. For instance, AnomalyGFM computes the residual $f_\theta: o \mapsto s \in \mathbb{R}$7, and anomaly class prototypes are aligned to these residuals with class-wise squared loss. Zero-shot scoring for node $f_\theta: o \mapsto s \in \mathbb{R}$8 is:

$f_\theta: o \mapsto s \in \mathbb{R}$9

(Qiao et al., 13 Feb 2025)

• Foundation group anomaly models: GFM4GA (Chen et al., 15 Jan 2026) extends foundation GAD to group anomaly detection using dual-level (node and subgraph) contrastive pretraining, adaptive group-contextual estimation, and group-proportion weighted losses, enabling detection of anomalous groups that may not be apparent from their individual members.
• Large graph LLMs and random-walk transformers: Models such as CyberGFM (King et al., 9 Jan 2026) treat graph walks as tokenized corpora and train BERT-style transformers on masked-node/edge prediction. After pretraining, fine-tuning and inference are performed via (masked) link prediction objectives or classifier heads, achieving state-of-the-art anomaly detection in unsupervised cyber-security tasks.
• Domain adaptation and disassortativity-aware fusion: Models such as TA-GGAD (Zhang et al., 10 Mar 2026) introduce explicit domain discrepancy quantification—called "Anomaly Disassortativity"—measuring the divergence between source and target node and structure-level anomaly scores. At test time, node- and structure-aware anomaly scores are aggregated with domain-adapted weights, and final outputs are further combined by pseudo-labeled, test-time-adaptive fusion, enabling robust cross-domain detection.

3. Enabling Universal Generalization: Mechanisms and Adaptation

Foundation GAD pipelines employ a spectrum of mechanisms for robust transfer:

• Feature space unification: Standardization via SVD/PCA, or residualization relative to neighborhoods, producing embeddings that are comparable across diverse graphs (Qiao et al., 13 Feb 2025).
• Zero-shot inference: After pretraining, the entire parameter set is kept frozen, and only model outputs—residuals, prompt alignments, or prediction scores—are used for anomaly scoring.
• Few-shot prompt tuning: On encountering new graphs, only small prompt/prototype parameters are adapted using a handful of labeled normal (or, occasionally, anomalous) samples; backbone parameters remain fixed (Qiao et al., 13 Feb 2025).
• Test-time adaptation via disassortativity-aware adapters (ADAs) and score adapaters (TSAs): Weights for different anomaly score sources are dynamically assigned based on distributions of scores in source vs. target, enabling real-time adaptation without retraining (Zhang et al., 10 Mar 2026).

A non-exhaustive comparative table of foundation GAD adaptation mechanisms:

Model	Pretrain/Finetune	Zero-shot	Few-shot prompt	Test-time adaptation
AnomalyGFM	Full param, one aux graph	Yes	Prompt tuning	Subgraph sampling
GFM4GA	Dual-contrastive, unlabeled	Yes	Constrained group	Group context
TA-GGAD	Multi-graph, contrastive	Yes	—	ADA/TSA fusion
CyberGFM	MLM on walks, no graphs	Yes	—	Masked fine-tune

4. Granularities: Node, Edge, Subgraph, Group, Full-Graph

Foundation GAD targets all major anomaly granularities:

• Node-level and edge-level: Most foundation models (AnomalyGFM, TA-GGAD, ARC, UNPrompt) focus on node and edge anomaly scores, leveraging representation residuals, neighborhood inconsistency, and contrastive signals.
• Group-level: GFM4GA (Chen et al., 15 Jan 2026) is specifically designed for group anomaly detection, employing contrastive learning at subgraph and node levels, followed by contextual adaptation via neighbor selection and group-contextual estimators.
• Full-graph: Approaches such as HimNet (Niu et al., 2023) and multi-level models (Bridges et al., 2014) support anomaly detection at both local (node, subgraph) and global (graph) scales, using hierarchical memory and multi-scale likelihoods, respectively.

5. Key Datasets, Evaluation Protocols, and Benchmarking

Benchmarks for foundation GAD research span:

• Node-level: Cora, Citeseer, Pubmed, BlogCatalog, Flickr, Facebook, Reddit, Amazon, YelpChi, T-Social, T-Finance, Elliptic (Qiao et al., 13 Feb 2025, Pan et al., 8 Sep 2025, Zhang et al., 10 Mar 2026).
• Edge-level: LANL, eBay, OpTC, UNSW (Zhou et al., 2024, King et al., 9 Jan 2026).
• Graph-level: MUTAG, PROTEINS, real and synthetic community-structured graphs (Niu et al., 2023, Bridges et al., 2014).
• Group-level: Weixin, Weibo, Facebook (group subgraphs synthesized or labeled) (Chen et al., 15 Jan 2026).

Protocols include:

• Zero-shot: Train on source graphs, entirely hold out target domains for evaluation.
• Few-shot/few-labeled: Allow $s=f_\theta(o)$0 labeled (usually normal) samples for prompt adaptation.
• Metrics: AUROC, AUPRC, precision@k, recall@k, F1, and sometimes interpretable visualizations at community or group level (Niu et al., 2023, Bridges et al., 2014).

Evaluation tables consistently demonstrate that properly engineered foundation models not only outperform prior supervised or unsupervised baselines but also retain their edge in severe domain shift and class imbalance scenarios (Qiao et al., 13 Feb 2025, Zhang et al., 10 Mar 2026, Chen et al., 15 Jan 2026, King et al., 9 Jan 2026).

6. Theoretical Advances and Emerging Research Directions

Open challenges include:

• Theoretical guarantees: There is a lack of principled theoretical characterization of when cross-domain pretraining, contrastive objectives, or prompting mechanisms succeed in GAD (Pan et al., 8 Sep 2025).
• Scalability and universality: Developing architectures and scaling laws that ensure monotonic improvements as data/model sizes grow, akin to other modalities' foundation models.
• Interactive/online adaptation: Human-in-the-loop elements—rapid updating with expert/analyst feedback, dynamic anomaly prior adjustment—are underexplored (Pan et al., 8 Sep 2025).
• Benchmarking: No standardized, unified evaluation suite matching the breadth of real-world cross-domain, cross-granularity, and low-shot GAD challenges is currently universally accepted.

Notably, emerging work—including the spectral theory underpinning BWGNN (Tang et al., 2022), Markov logic reasoning in KnowGraph (Zhou et al., 2024), and robust multi-level modeling (Bridges et al., 2014)—continues to shift the design of foundation GAD models toward spectrum-first, knowledge-integrated, and hierarchical probabilistic methods.

7. Impact and Outlook

Foundation graph anomaly detection has redefined the scope of GAD from narrow task-specific deployment to broadly generalizable, efficient, and scalable universal models. Advances in architectural design (shared encoders, prototype alignment, dual-level contrastive learning), adaptation mechanisms (prompt tuning, disassortativity-aware fusion, test-time adaptation), and rigorous benchmarking have collectively resulted in significant AUROC and AUPRC performance gains in both zero- and few-shot scenarios across highly heterogeneous domains (Qiao et al., 13 Feb 2025, Zhang et al., 10 Mar 2026, Chen et al., 15 Jan 2026, King et al., 9 Jan 2026). However, realization of truly universal foundation GAD models remains contingent on advances in theory, scalable pretraining datasets, principled evaluation frameworks, and adaptive methodologies for dealing with evolving, adversarial, or highly imbalanced anomaly regimes (Pan et al., 8 Sep 2025).