Systems that flag citation gaming increasingly use graph neural networks (GNNs), models that build each paper's representation by repeatedly averaging information from the papers it cites and is cited by. These detectors are trained on labels such as retraction lists and manual annotations. Those labels are incomplete in a predictable way: most cartel papers were never caught, so they enter training labeled "normal." Work on graph label noise shows that such errors do not stay put. Message passing spreads them into the representations of neighboring papers. Robust-training methods have been built to contain this spread, but every one of them has been tested only against random label flips.
A citation cartel is not random. It controls which citations its members add. It can make its own cluster look internally consistent, which is the property these denoisers rely on to spot mislabeled nodes. This proposal treats the cartel as a strategic source of label noise. It measures detector robustness the way computational social choice measures voting rules: by the minimum manipulation a coalition needs to escape detection, and by how well cheap heuristic attackers do compared with the exact optimum. The result is a price tag, counted in added citations, for evading each detector. Alongside it comes a defender's price, counted in audited labels, for restoring detection.
Core Claim
Benchmarks with random label noise overstate how robust graph-based cartel detectors are. Under noise shaped by a coalition, where unlabeled cartel members coincide with citations the cartel itself chose, we expect two failures. Dual-view and influence-based denoisers should lose most of their advantage over a plain GNN. Greedy attackers should evade them with a small number of added citations, close to the exact minimum on instances small enough to solve. If this holds, the worst-case hardness results that are taken as a shield for bibliometric scores give little practical protection on realistic citation graphs.
The new object is a "manipulation cost" curve per detector, which can be compared directly across methods. It replaces post-hoc accuracy drops at a fixed noise rate.
Why This Gap Is Real
The label-noise literature supplies both the motivation and the methods to be tested. Learning on Graphs under Label Noise shows that a memorized wrong label contaminates adjacent clean nodes. EDN shows that noise correlated with graph edges degrades GNNs by 10–15 points more than independent noise at 45% corruption. EDN's noise is still sampled at random, not optimized by an adversary.
The defenses all assume that mislabeled nodes reveal themselves through disagreement. GD² contrasts an ego-only predictor with a neighborhood predictor. ICGNN scores mismatch between a node's label and the labels diffused to it from its neighbors. PI-GNN replaces node labels with "same class or not" pairs. TSS trains first on nodes far from class boundaries. A cartel that cites densely within one topic removes exactly the disagreement these methods look for. Nobody has tested this.
On the citation side, GLAD classifies citation edges as anomalous using graph, metadata and citation-context features. It treats its training labels as clean and never considers an adaptive adversary. COIRank down-weights conflict-of-interest edges with fixed heuristics. The Pagerank-Index argues that a closed clique gains no authority. That argument assumes the clique stays closed.
The complexity work constrains the design. How Hard Is It to Impact the Impact of Your Paper? proves that manipulating the consolidation–disruption (CD) index is W[1]- or W[2]-hard, meaning hard even when the attacker's budget is small. It does not test heuristics. The Shield that Never Was shows that such hardness collapses on structured preference profiles. The Complexity of Bribery in Network-Based Rating Systems shows that greedy bribery of high-degree nodes is provably suboptimal under neighborhood averaging. This is a warning that naive attackers will understate vulnerability.
From Bribery as a Measure of Candidate Success and How Many Votes is a Lie Worth? we borrow two ideas. First, minimum manipulation distance is a graded robustness measure. Second, the defender's corrective resources can be counted in the same units as the attack.
Research Program
Threat model. We fix a feasible attack model. A cartel is a set $C$ of papers. It may add citations only from its own papers, only to papers published earlier, which keeps the citation graph acyclic, and only to targets within a topical radius, so the citations stay plausible. It may also cite external papers, which buys apparent openness.
Training labels. These are generated the way real ones are. A small audited fraction of cartel papers and edges is labeled anomalous, and the rest of $C$ is labeled normal. The noise is therefore concentrated on the cartel by construction.
Manipulation cost. For a detector $f$ and target recall $r$, the cost $M_f(r)$ is the minimum number of added citations after which, once $f$ is retrained on the poisoned graph and labels, fewer than a fraction $r$ of cartel members are flagged. We estimate it with several attackers: random addition, degree-greedy addition, influence-greedy addition (scored by the change in the trained detector's output), and simulated-annealing local search. On small instances we compute the exact value with a mixed-integer program. To keep that program solvable, it attacks a linearized GNN (simplified graph convolution, no nonlinearities). It also attacks the CD index directly, using Yang's CD-MANI-ADD formulation.
Shield gap. The ratio of the heuristic attack cost to the exact cost is the empirical gap in the complexity shield. A ratio near one means hardness offers no practical protection.
Defender's cost. Following the resource-augmentation idea, the defender's cost $D_f(r)$ is the number of additional audited labels needed to restore recall $r$ after the attack. These labels are chosen by an active-selection rule in the spirit of Algorithms and Hardness for Active Learning on Graphs. The ratio $M_f/D_f$ summarizes who wins the cost race.
Baselines
Plain GNN. A two-layer graph convolutional network trained directly on the noisy labels, with no denoising. Every method must beat it, and $M_f$ for this model sets the floor.
Strongest existing detector: GLAD. GLAD combines graph-infomax node embeddings, a metadata edge autoencoder and a SciBERT citation-purpose classifier. It reports precision 0.9683 and F1 0.9075 on its 4,718-edge Microsoft Academic Graph dataset. We run it at edge level on that dataset and on planted-cartel graphs, using the same splits, label budget and hyperparameter search as every other method. Beating GLAD means a higher $M_f(0.5)$ at matched clean-data F1, within two points.
Strongest denoisers. GD², ICGNN, PI-GNN, CGNN and TSS, run through the NoisyGL library on the node-level task "is this paper a cartel member." Each gets equal tuning trials and a fixed GPU budget. ICGNN's cubic-cost diffusion step is reported as out-of-memory rather than tuned around. The closest reported anchor is TSS at 86.69±0.32% accuracy on Cora with 40% symmetric noise. For denoisers on citation-cartel detection, no reported result was found on Lacuna.
Ranking scores. The PageRank-index, COIRank and the CD index are attacked under the identical attack model. This tests the claim that rankings resist cartels without any learning.
Ablation. Each detector also faces EDN noise at the same per-node corruption rate, and random-flip noise at that rate. This isolates whether strategic placement, rather than mere correlation with edges, causes the damage.
Why not just…?
Why not just use EDN-style correlated noise instead of a strategic attacker? EDN already shows that correlated noise is worse than independent noise. If EDN at the same corruption rate hurts the denoisers as much as our attackers do, the strategic framing adds nothing, and we will report that outcome. The matched-rate ablation settles it. Our prediction rests on the observation that the dual-view and diffusion denoisers need ego–structure disagreement, which EDN's random sampling leaves in place but a cartel can remove.
Why not just reuse generic gradient-based graph poisoning attacks? Unconstrained attackers may edit any edge anywhere. That overstates what a cartel can do, and it says nothing about the gap between exact and heuristic attacks that the complexity question asks about. We include an unconstrained gradient attacker as an upper bound on vulnerability. The constrained, time-respecting attacker is the realistic estimate.
Why not just rely on PageRank-style scores, which supposedly ignore closed cliques? This is a real threat to the motivation. If the PageRank-index or COIRank holds high $M_f$ under our attack model, the right conclusion is that learned detectors should borrow their authority weighting, not that they need new defenses. The outward-citation option in our threat model tests the "closed clique" assumption directly.
Why not just filter co-author and same-affiliation citations? COIRank finds that 80.88% of APS articles receive co-author citations. A hard filter would therefore flag most legitimate science. We include it only as a reference row reporting its false-positive rate.
Proposed Experiments
Data. We plant cartels in Cora, CiteSeer, PubMed and OGBN-arxiv, following the synthetic-injection practice of Jolly et al. (100% precision and 86.01% F1 on their synthetic journal-citation anomaly dataset). Cartel sizes range from 5 to 200 papers. Planted cartels are calibrated so that their co-authorship density matches the APS PRC conflict-of-interest statistics, and so that they are not trivially separable. Realism checks use the GLAD Microsoft Academic Graph subset and APS PRC.
Metrics. We report $M_f(r)$ at $r = 0.5$ and $r = 0.8$, the heuristic-to-exact ratio on instances with at most 300 nodes, $D_f(r)$, clean-data F1, and wall-clock cost.
First decisive check. On the node-level task, do the best denoisers reach $M_f$ within a factor of two of the plain GNN, while beating it clearly under random noise? That pattern would confirm that their robustness does not carry over to strategic noise.
Second decisive check. Is the heuristic-to-exact ratio for CD-index manipulation below about 1.2 on real citation subgraphs? That would show Yang's hardness result gives little practical protection.
Third decisive check. Does the $M_f/D_f$ ratio favor the attacker? If tens of added citations require hundreds of audits to undo, defenders must prioritize audit selection over model architecture.
A negative result on any of these checks is still informative. It would identify which detector family, or which ranking score, survives strategic noise and is worth building on.
References
- Learning on Graphs under Label Noise
- EDN: A Novel Edge-Dependent Noise Model for Graph Data
- GD$^2$: Robust Graph Learning under Label Noise via Dual-View Prediction Discrepancy
- Identifying and Correcting Label Noise for Robust GNNs via Influence Contradiction
- Noise-robust Graph Learning by Estimating and Leveraging Pairwise Interactions
- Mitigating Label Noise on Graphs via Topological Sample Selection
- Deep Graph Learning for Anomalous Citation Detection
- Identifying Anomalous Citations for Objective Evaluation of Scholarly Article Impact
- The Pagerank-Index: Going beyond Citation Counts in Quantifying Scientific Impact of Researchers
- How Hard Is It to Impact the Impact of Your Paper?
- The shield that never was: societies with single-peaked preferences are more open to manipulation and control
- The Complexity of Bribery in Network-Based Rating Systems
- Bribery as a Measure of Candidate Success: Complexity Results for Approval-Based Multiwinner Rules
- How Many Votes is a Lie Worth? Measuring Strategyproofness through Resource Augmentation
- Algorithms and Hardness for Active Learning on Graphs
- NoisyGL: A Comprehensive Benchmark for Graph Neural Networks under Label Noise
- Unsupervised Anomaly Detection in Journal-Level Citation Networks