Enhancing Graph Neural Network-based Fraud Detectors against Camouflaged Fraudsters

Yingtong Dou; Zhiwei Liu; Li Sun; Yutong Deng; Hao Peng; Philip S. Yu

Source-linked AI summary

Enhancing Graph Neural Network-based Fraud Detectors against Camouflaged Fraudsters

Yingtong Dou, Zhiwei Liu, Li Sun, Yutong Deng, Hao Peng, Philip S. Yu

arXiv:2008.08692v1Aug 19, 2020cs.SIcs.CRcs.LG

TL;DR

GNN-based fraud detectors often overlook feature and relation camouflage, which can hamper neighborhood aggregation. CARE-GNN counters these behaviors with label-aware neighbor selection, reinforcement learning, and relation-aware aggregation, outperforming state-of-the-art baselines on two real-world fraud datasets.

Problem
Existing GNN-based fraud detectors do not adequately address feature and relation camouflage, while noisy nodes and edges expose limitations in their aggregation process.
Method
CARE-GNN uses a label-aware similarity measure, a reinforcement-learning neighbor selector, and a relation-aware aggregator to filter and combine informative neighbors.
Results
CARE-GNN boosts GNN performance on two real-world fraud datasets and outperforms state-of-the-art baselines under various settings.
Takeaways & Limitations
The experiments provide evidence that fraudster camouflage affects GNN-based fraud detection and that the proposed enhancement modules, especially reinforcement learning, are effective and efficient.
Takeaways & Limitations
The parameterized similarity measures discussed have time complexity O(|V| ¯Dd), which is costly when real-world graphs have very high average degree.

Abstract

from arXiv · show

Graph Neural Networks (GNNs) have been widely applied to fraud detection problems in recent years, revealing the suspiciousness of nodes by aggregating their neighborhood information via different relations. However, few prior works have noticed the camouflage behavior of fraudsters, which could hamper the performance of GNN-based fraud detectors during the aggregation process. In this paper, we introduce two types of camouflages based on recent empirical studies, i.e., the feature camouflage and the relation camouflage. Existing GNNs have not addressed these two camouflages, which results in their poor performance in fraud detection problems. Alternatively, we propose a new model named CAmouflage-REsistant GNN (CARE-GNN), to enhance the GNN aggregation process with three unique modules against camouflages. Concretely, we first devise a label-aware similarity measure to find informative neighboring nodes. Then, we leverage reinforcement learning (RL) to find the optimal amounts of neighbors to be selected. Finally, the selected neighbors across different relations are aggregated together. Comprehensive experiments on two real-world fraud datasets demonstrate the effectiveness of the RL algorithm. The proposed CARE-GNN also outperforms state-of-the-art GNNs and GNN-based fraud detectors. We integrate all GNN-based fraud detectors as an opensource toolbox: https://github.com/safe-graph/DGFraud. The CARE-GNN code and datasets are available at https://github.com/YingtongDou/CARE-GNN.

1 INTRODUCTION

GNN-based fraud detectors can be undermined by feature and relation camouflage during neighborhood aggregation. CARE-GNN addresses these behaviors by selecting informative neighbors, learning selection thresholds with reinforcement learning, and aggregating across relations.

Existing GNN-based fraud detectors overlook fraudster camouflage, despite known vulnerabilities to noisy nodes and edges.
Feature camouflage disguises fraud through altered node features, while relation camouflage connects fraudsters to many benign entities.Examples include special characters in reviews and links to reputable users or regular reviews.
Aggregating camouflaged neighbors can smooth out or eliminate the suspiciousness of a center fraudster.
CARE-GNN uses a label-aware similarity measure, a reinforcement-learning neighbor selector, and a relation-aware aggregator.The selector learns filtering thresholds during GNN training and the aggregator combines information across relations.
Experiments on two real-world fraud datasets show that CARE-GNN boosts GNN performance and outperforms state-of-the-art baselines under various settings.
CARE-GNN adaptively selects neighbors, maintains high computational efficiency, and supports plugging in other neural modules or external knowledge.

2 PROBLEM DEFINITION

The paper formulates fraud detection on multi-relation graphs, where node features, labels, and relation-specific edges support learning suspiciousness. GNN-based fraud detection separates neighborhood aggregation within relations from aggregation across relations.

A multi-relation graph contains nodes with d-dimensional features, relation-specific edges, and labels for each node.
In graph-based fraud detection, the target node may be a review or transaction, with label 0 for benign and 1 for suspicious.
GNNs embed graph-structured data by aggregating information from neighboring nodes.
Fraud-detection GNNs train on partially labeled nodes using binary classification loss functions.
Neighbor aggregation is separated into intra-relation aggregation and inter-relation aggregation across different relations.Neighbor embeddings are aggregated within each relation before information from different relations is combined.

3 PROPOSED MODEL

CARE-GNN resists feature and relation camouflage by learning label-aware neighbor similarity, adaptively selecting neighbors with reinforcement learning, and aggregating the retained neighbors across relations.

3.2 Label-aware Similarity Measure: CARE-GNN computes label-aware similarity between a center node and neighbors, using a layer-specific MLP and prediction-distance measure to identify informative neighbors.The proposed measure uses each node’s embedding as input and reduces first-layer complexity to O(|V|d).
3.2 Label-aware Similarity Measure: The similarity measure is directly trained with a supervised loss so similar neighbors can be selected early and help regularize GNN training.The paper argues that direct parameter updates avoid relying only on later back-propagation through aggregation.
3.3.1 Top-p Sampling: CARE-GNN ranks neighbors by similarity and applies top-p sampling independently within each relation, discarding unselected nodes from aggregation.The threshold p_r^(l) determines the fraction retained at layer l for relation r.
3.3.2 Finding the Optimal Thresholds with RL: Reinforcement learning searches for relation- and layer-specific thresholds by increasing or decreasing them according to rewards based on changes in average neighbor distance.The selector is formulated as a Bernoulli multi-armed bandit with threshold adjustments as actions.
3.3.2 Finding the Optimal Thresholds with RL: The reward is positive when newly selected neighbors have lower average distance than in the previous epoch, and the threshold is adjusted in the corresponding direction.The process uses immediate greedy updates rather than cumulative-reward estimation.
3.3.2 Finding the Optimal Thresholds with RL: After reinforcement learning converges over recent epochs, the discovered filtering thresholds remain fixed while the GNN continues training.This terminal condition identifies an optimal threshold p_r^(l) for the subsequent convergence phase.

4 EXPERIMENTS

The experiments evaluate CARE-GNN on Yelp and Amazon fraud datasets using multi-relation graphs, semi-supervised training, and AUC and Recall metrics. They compare CARE-GNN with general GNNs and GNN-based fraud detectors under controlled implementation settings.

Datasets and Graph Construction: Experiments use Yelp and Amazon review datasets for binary fraud detection with handcrafted node features and multi-relation graphs.Yelp uses reviews as nodes and Amazon uses users as nodes, with three relations designed for each dataset.
Compared Methods: The evaluation compares CARE-GNN against GCN, GAT, RGCN, GraphSAGE, GeniePath, Player2Vec, SemiGNN, and GraphConsis.The comparison uses general GNN models and state-of-the-art GNN-based fraud detectors under semi-supervised learning.
Experimental Setup: GCN, GAT, GraphSAGE, and GeniePath use homogeneous graphs, whereas the other compared models use multi-relation graphs.Homogeneous graphs merge all relations, while multi-relation models handle relations separately.
Training Procedure: The experiments use mini-batch training and under-sampling, matching negative and positive instances within each mini-batch.These techniques are used to improve training efficiency and avoid overfitting in imbalanced fraud datasets.
Evaluation Metrics: ROC-AUC and Recall evaluate classifier performance, with ROC-AUC addressing class imbalance through prediction-probability rankings.The evaluation focuses on fraudsters as positive instances.

4.2 Camouflage Evidence

The datasets exhibit both feature and relation camouflage. Neighboring nodes often have similar features, while label similarity is high only for one relation and low for the others.

Feature Camouflage: Average feature similarity is high across all relations, indicating that fraudsters camouflage their features to resemble neighboring entities.Feature similarity is computed from neighboring-node feature vectors using Euclidean distance and normalized by the total number of edges.
Relation Camouflage: Only the R-U-R relation has high label similarity, while the other relations show lower label similarity consistent with relation camouflage.High label similarity indicates weaker camouflage, whereas low label similarity indicates more successful camouflage.

4.3 Overall Evaluation

CARE-GNN outperforms other baselines under most training proportions and metrics on the two fraud datasets. The results also indicate that filtering dissimilar neighbors is important for noisy multi-relation graphs.

Overall Performance: CARE-GNN outperforms other baselines under most training proportions and metrics across the two datasets.Table 3 reports the best testing results after thirty epochs.
Single-relation vs. Multi-relation: Single-relation GNNs outperform Player2Vec and SemiGNN, while GraphConsis performs best among the other multi-relation GNNs.The comparison distinguishes models operating on merged homogeneous graphs from those using multi-relation graphs.
Single-relation vs. Multi-relation: CARE-GNN and its variants better identify and filter camouflaged fraudsters through parameterized similarity measures and adaptive sampling thresholds.The results characterize neighbor filtering as critical when graphs contain dissimilar or camouflaged neighbors.
Training Percentage: Increasing the training percentage yields little performance gain for GNNs, indicating that small amounts of supervised signal can suffice in the evaluated setting.The datasets provide informative handcrafted features for learning high-quality embeddings.
CARE-GNN Variants: CARE-GNN performs better than CARE-Mean, suggesting that using filtering thresholds as inter-relation aggregation weights can improve performance while reducing model complexity.CARE-GNN variants have similar performance under most training percentages and metrics, but CARE-Att is weaker for Yelp at smaller training percentages.
GNN vs. Similarity Measure: On Yelp, the GNN module has better AUC and Recall than the similarity measure, whereas on Amazon their performances are comparable.The paper attributes the Yelp difference to structural information and the Amazon comparability to sufficiently discriminative input features.

4.4 RL Process Analysis

The RL-based neighbor selector adapts relation-specific filtering thresholds during training. As similar neighbors are selected, relation weights converge and neighbor-distance patterns become more comparable.

Relation Weights: Randomly initialized relation aggregation weights converge toward similar values as the selector chooses more similar neighbors under each relation.The paper relates similar weights to relations providing similar information after filtering.
Relation Distance: Neighbor distances across relations become more separated and comparable as training progresses and the GNN makes embeddings more distinguishable.Filtering noisy neighbors brings average distances across relations closer together.
Neighbor Filtering Threshold: All filtering thresholds converge to different values under the RL process, with oscillations preceding the terminal condition.The action step size is τ = 0.02.
Neighbor Selector Comparison: Adaptive filtering reaches faster optimization and smoother, better training performance than selecting the top 50% or all neighbors.The comparison uses Adaptive, Fixed-Half, and Fixed-All neighbor-selection criteria on both datasets.

4.5 Hyper-parameter Sensitivity

CARE-GNN is relatively insensitive to several hyper-parameters, with one layer often preferred on the dense Yelp graph. Under-sampling and embedding-size choices also affect performance, while excessive depth can overfit.

Number of layers: Increasing CARE-GNN’s number of layers barely improves performance, while the three-layer model overfits with Recall = 0.5.The one-layer model saves computational cost and achieves better classification results.
Under-sampling ratio: An equal under-sampling ratio provides good and fair performance, whereas ratios under 1:0.2 and 1:0.5 overfit with Recall = 0.5.CARE-GNN is tested on an imbalanced test set.

4.6 Discussion

CARE-GNN is efficient because one-hop aggregation already captures abundant information in dense multi-relation graphs, while lightweight design choices reduce training overhead.

Discussion: On dense graphs with average node degree > 150, one-layer CARE-GNN aggregates abundant one-hop information and achieves excellent performance.More layers are described as suitable for sparse graphs.
Discussion: CARE-GNN takes 17 seconds per epoch on Yelp and 3 seconds on Amazon while gaining performance over other baselines.Efficiency comes from lightweight similarity, classic RL, positive-node neighbor selection, no attention, and mini-batch under-sampling.

5 RELATED WORK

Related work addresses noisy or irrelevant neighbors through graph structure learning, metric learning, and sampling, but CARE-GNN targets camouflaged fraudsters with adaptive, label-aware filtering.

GNN and Its Enhancement: Graph structure learning methods learn new graph structures, whereas CARE-GNN filters dissimilar neighbors instead of learning new structures.The cited structure-learning works investigate single-relation benchmark datasets without camouflaged fraudsters.
GNN and Its Enhancement: Metric-learning methods measure neighbor similarity, but prior approaches use weak metrics or fixed thresholds that require empirical calibration.CARE-GNN learns similarity from domain knowledge and optimizes relation thresholds during end-to-end GNN training.
GNN-based Fraud Detection: GNN sampling methods select representative nodes to accelerate training, while CARE-GNN filters dissimilar neighbors using domain knowledge and relational information before aggregation.Its goal is to alleviate the negative effect of camouflaged fraudsters.
GNN-based Fraud Detection: CARE-GNN constructs multiple homo-graphs with one node type, like GEM and ASA.This design is presented within the paper’s comparison of GNN-based fraud detectors.
GNN-based Fraud Detection: Only two prior works noticed fraudster camouflage; CARE-GNN addresses their inflexible features, unsupervised similarity, and fixed thresholds with label-aware adaptive filtering.The comparison identifies the shortcomings of the cited approaches and CARE-GNN’s corresponding remedy.

6 CONCLUSION

The paper studies how feature and relation camouflage affect GNN-based fraud detection and proposes CARE-GNN to counter these effects. Its experiments support the effectiveness and efficiency of the enhancement modules, especially reinforcement learning.

Conclusion: The paper proposes a label-aware similarity measure, a reinforcement-learning neighbor selector, and a relation-aware aggregator against feature and relation camouflage.These modules enhance GNN-based fraud detectors against the identified camouflage behaviors.
Conclusion: Experiments on real-world fraud datasets provide evidence of fraudster camouflage and demonstrate the effectiveness and efficiency of the proposed enhancement modules.The conclusion emphasizes the reinforcement learning module in particular.