Auditing Representation-Space Fraud Detection: A Bayesian IRL Framework for Calibrated Uncertainty and Its Limits
Organizations: JAIN University (Deemed-to-be), Bengaluru, India · Zayed University, Dubai, UAE
Abstract
Shilling detectors for recommender systems are almost always evaluated by one number, detection AUROC, with no account of when it can be trusted. We audit representation-space detection with VEIL (Variational Evaluation via Inverse-reinforcement Learning), a pairwise (Bradley-Terry) Bayesian reward-inference framework over frozen encoder embeddings, following the Bayesian auditing paradigm for LLM alignment [1]: (A) a base posterior; (B) a hierarchical ARD extension with learned confidence; and (C) a label-routed organic/attacker mixture used as an evidence-accumulation check.We prove (Proposition 1) that no detector built on a representation invariant to a transformation of user profiles can outperform chance against an attack that is indistinguishable from organic profiles under that invariance.Detection is near-perfect against synthetic-structure attacks (AUROC 0.936-1.000) but collapses under donor-mimicry, which clones a real user's item set and splices in 1-3 target items. Performance falls to chance on MovieLens-1M (0.48-0.51), partially degrades on Amazon Musical Instruments, and degrades least on CiaoDVD.Donor-mimicry is not undetectable: a non-learned baseline, based on each profile's maximum raw item-set Jaccard similarity to any other profile, reaches AUROC 0.92-0.97 (average precision 0.50-0.57) on all four donor-mimicry configurations, including ML-1M, where every embedding-based score is at chance.The signal is present in the raw profiles, but LightGCN's representation discards it, placing injected profiles closer to their true donor than its zero-propagation MF reduction does (distance 1.98 vs. 3.09). Severity varies by dataset, encoder, and detection method, with no single axis predicting the others. The limit lies in the representation, not the attack, and raw-similarity checks are a necessary complement to any learned detector.
Figures & tables
| config | attack | A AUROC | A AP | B AUROC | B | C AUROC | C |
|---|---|---|---|---|---|---|---|
| supgnn_base | random+avg+bw | 0.998 | 0.961 | 0.998 | 3.22 | 0.998 | 3.24 / 3.00 |
| lightgcn_base | random+avg+bw | 0.999 | 0.995 | 1.000 | 0.42 | 1.000 | 0.47 / 0.30 |
| lightgcn_aush | AUSH | 1.000 | 1.000 | 1.000 | 0.40 | 1.000 | 0.69 / 0.31 |
| lightgcn_aush_het | AUSH+Het. | 0.942 | 0.423 | 0.944 | 0.70 | 0.936 | 0.83 / 0.49 |
| lightgcn_aush_het_donor | +Donor-mimicry | 0.500 | 0.051 | 0.513 | 0.54 | 0.507 | 0.54 / 0.39 |
| config | AUROC (95% CI) | ECE (95% CI) |
|---|---|---|
| lightgcn_base (easy) | 1.000 [0.999, 1.000] | 0.006 [0.004, 0.008] |
| lightgcn_aush_het | 0.945 [0.932, 0.958] | 0.090 [0.082, 0.099] |
| CiaoDVD (easy) | 0.877 [0.866, 0.888] | 0.177 [0.171, 0.184] |
| Musical Instruments (hard) | 0.805 [0.797, 0.814] | 0.263 [0.259, 0.268] |
| lightgcn_aush_het_donor (hard) | 0.494 [0.440, 0.549] | 0.374 [0.360, 0.388] |
| Dataset | Point | A AUROC | B AUROC | C AUROC |
|---|---|---|---|---|
| ML-1M | easy | 0.999 | 1.000 | 1.000 |
| ML-1M | hard (matched, no AUSH) † | |||
| Musical Instruments | easy | 0.991 | 0.992 | 0.991 |
| Musical Instruments | hard † | |||
| CiaoDVD | easy † | |||
| CiaoDVD | hard † |
| Architecture | A AUROC | B AUROC | C AUROC |
|---|---|---|---|
| LightGCN | |||
| mean-pool | |||
| SASRec | |||
| MF |
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| dataset | config | VEIL A/B/C | kNN | LOF | -means | radial | Jaccard |
|---|---|---|---|---|---|---|---|
| ML-1M | classic mix | 1.000/1.000/1.000 | 0.613 | 0.344 | 0.618 | 0.670 | 0.340 |
| AUSH | 1.000/1.000/1.000 | 0.000 | 0.350 | 0.971 | 0.937 | 1.000 | |
| donor, full ladder | 0.494/0.488/0.481 | 0.478 | 0.547 | 0.603 | 0.577 | 0.974 | |
| donor, matched (no AUSH) | 0.533/0.536/0.530 | 0.444 | 0.530 | 0.558 | 0.557 | 0.924 | |
| CiaoDVD | classic mix | 0.877/0.878/0.879 | 0.299 | 0.102 | 0.503 | 0.365 | 0.389 |
| donor (no AUSH) | 0.846/0.848/0.846 | 0.477 | 0.472 | 0.582 | 0.631 | 0.942 |
| dataset | config | kNN | LOF | -means | radial |
|---|---|---|---|---|---|
| ML-1M | classic mix | 0.620 [0.569, 0.672] | 0.335 [0.304, 0.364] | 0.617 [0.580, 0.653] | 0.668 [0.626, 0.706] |
| AUSH | 0.000 [0.000, 0.000] | 0.355 [0.302, 0.411] | 0.972 [0.966, 0.979] | 0.939 [0.929, 0.949] | |
| donor, full ladder | 0.471 [0.412, 0.540] | 0.536 [0.470, 0.604] | 0.597 [0.544, 0.652] | 0.580 [0.521, 0.636] | |
| donor, matched (no AUSH) | 0.438 [0.400, 0.476] | 0.525 [0.491, 0.562] | 0.538 [0.504, 0.576] | 0.541 [0.503, 0.577] | |
| CiaoDVD | classic mix | 0.304 [0.275, 0.334] | 0.105 [0.090, 0.121] | 0.509 [0.487, 0.532] | 0.362 [0.323, 0.400] |
| donor (no AUSH) | 0.482 [0.447, 0.516] | 0.473 [0.436, 0.509] | 0.568 [0.534, 0.600] | 0.621 [0.590, 0.649] |
| dataset | config | attack | VEIL A † | kNN | LOF | -means | radial |
|---|---|---|---|---|---|---|---|
| ML-1M | classic mix | random | 0.999 | 0.928 | 0.364 | 0.807 | 0.896 |
| average | 0.999 | 0.930 | 0.392 | 0.828 | 0.906 | ||
| bandwagon | 1.000 | 0.000 | 0.279 | 0.232 | 0.221 | ||
| AUSH | aush | 1.000 ‡ | 0.000 | 0.350 | 0.971 | 0.937 | |
| CiaoDVD | classic mix | random | 0.825 | 0.447 | 0.098 | 0.538 | 0.141 |
| average | 0.825 | 0.456 | 0.114 | 0.505 | 0.207 |
| config | organic flagged/total | raw AUROC / AP | donor-excl. AUROC / AP |
|---|---|---|---|
| ML-1M, full ladder | 90 / 1,812 | 0.974 / 0.515 | 1.0000 / 1.000 |
| ML-1M, matched (no AUSH) | 257 / 1,812 | 0.924 / 0.503 | 1.0000 / 1.000 |
| CiaoDVD, donor | 227 / 1,822 | 0.942 / 0.569 | 0.9999 / 0.999 |
| Musical Instruments, donor | 1,773 / 15,973 | 0.940 / 0.549 | 0.9997 / 0.994 |
| config | A AUROC | A ECE | A Brier | A NLL | B AUROC | B ECE | C AUROC | C ECE |
|---|---|---|---|---|---|---|---|---|
| supgnn_base | 0.997 | 0.006 | 0.006 | 0.056 | 0.999 | 0.002 | 0.999 | 0.002 |
| lightgcn_base | 1.000 | 0.006 | 0.005 | 0.047 | 1.000 | 0.006 | 1.000 | 0.005 |
| lightgcn_aush | 1.000 | 0.000 | 0.000 | 0.000 | 1.000 | 0.000 | 1.000 | 0.000 |
| lightgcn_aush_het | 0.945 | 0.090 | 0.071 | 0.287 | 0.946 | 0.087 | 0.945 | 0.077 |
| lightgcn_aush_het_donor | 0.494 | 0.374 | 0.232 | 0.679 | 0.488 | 0.375 | 0.481 | 0.366 |
| Musical Instruments (easy) | 0.990 | 0.050 | 0.038 | 0.139 | 0.991 | 0.049 | 0.991 | 0.048 |
| test | train | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| config | AUROC | AP | Acc | F1 | pairs-Brier/ECE | AUROC | AP | Acc | F1 |
| supgnn_base | 0.9982 | 0.9614 | 0.9984 | 0.9939 | 0.0029 / 0.0176 | 0.9987 | 0.9872 | 0.9959 | 0.9843 |
| lightgcn_base | 0.9993 | 0.9948 | 0.9920 | 0.9697 | 0.0001 / 0.0004 | 1.0000 | 1.0000 | 0.9986 | 0.9948 |
| lightgcn_aush | 1.0000 | 1.0000 | 1.0000 | 1.0000 | 0.0000 / 0.0000 | 1.0000 | 1.0000 | 1.0000 | 1.0000 |
| lightgcn_aush_het | 0.9424 | 0.4234 | 0.9492 | 0.5246 | 0.0428 / 0.0318 | 0.9781 | 0.6188 | 0.9655 | 0.6406 |
| lightgcn_aush_het_donor | 0.4996 | 0.0511 | 0.9142 | 0.0755 | 0.3554 / 0.2697 | 0.7433 | 0.1067 | 0.9189 | 0.1563 |
| config | test AUROC | test AP | test Acc | test F1 | train AUROC | |||
|---|---|---|---|---|---|---|---|---|
| supgnn_base | 3.2168 | 0.0268 | 0.4751 | 0.9982 | 0.9512 | 0.9984 | 0.9939 | 1.0000 |
| lightgcn_base | 0.4209 | 0.0729 | 0.3904 | 0.9996 | 0.9973 | 0.9920 | 0.9697 | 1.0000 |
| lightgcn_aush | 0.4025 | 0.1317 | 0.2386 | 1.0000 | 1.0000 | 1.0000 | 1.0000 | 1.0000 |
| lightgcn_aush_het | 0.7044 | 0.0512 | 0.4866 | 0.9442 | 0.4391 | 0.9527 | 0.5574 | 0.9801 |
| lightgcn_aush_het_donor | 0.5412 | 0.1076 | 0.4984 | 0.5134 | 0.0523 | 0.9177 | 0.0784 | 0.7382 |
| config | mean | std | SE | 95% CI ( , df ) |
|---|---|---|---|---|
| lightgcn_base | 0.3861 | 0.0174 | 0.0062 | [0.3715, 0.4006] |
| lightgcn_aush | 0.3286 | 0.0604 | 0.0213 | [0.2781, 0.3791] |
| lightgcn_aush_het | 0.6877 | 0.0610 | 0.0216 | [0.6367, 0.7387] |
| lightgcn_aush_het_donor | 0.5514 | 0.0850 | 0.0301 | [0.4803, 0.6225] |
| config | test AUROC | test AP | test Brier | test ECE | train AUROC | ||
|---|---|---|---|---|---|---|---|
| supgnn_base | 3.2373 | 3.0016 | 0.9982 | 0.9512 | 0.0016 | 0.0017 | 1.0000 |
| lightgcn_base | 0.4670 | 0.2973 | 0.9997 | 0.9978 | 0.0063 | 0.0062 | 1.0000 |
| lightgcn_aush | 0.6916 | 0.3056 | 1.0000 | 1.0000 | 0.0000 | 0.0000 | 1.0000 |
| lightgcn_aush_het | 0.8285 | 0.4898 | 0.9364 | 0.4642 | 0.0707 | 0.0837 | 0.9848 |
| lightgcn_aush_het_donor | 0.5398 | 0.3934 | 0.5075 | 0.0631 | 0.2384 | 0.3889 | 0.7465 |
| branch AUROC (test) | |||||
|---|---|---|---|---|---|
| 20 | 300 | 78.4868 | 1.8578 | 0.1831 | 0.9990 |
| 50 | 750 | 59.0742 | 1.5937 | 0.2108 | 0.9992 |
| 100 | 1500 | 37.6224 | 1.3496 | 0.2527 | 0.9998 |
| 150 | 2250 | 20.2082 | 1.1777 | 0.2697 | 0.9997 |
| 191 | 2865 | 15.9541 | 1.1388 | 0.2973 | 0.9997 |
| config | Stage 0 (s) | A AUROC | B AUROC | C AUROC | LightGCN comparison |
|---|---|---|---|---|---|
| ml1m_mf_base | 191.0 | 0.9941 | 0.9944 | 0.9956 | indistinguishable (both near-ceiling) |
| ml1m_mf_aush | 221.2 | 1.0000 | 1.0000 | 1.0000 | both saturate |
| ml1m_mf_aush_het | 220.4 | 0.9451 | 0.9465 | 0.9547 | nearly identical |
| ml1m_mf_aush_het_donor ‡ | 1100.6 | 0.6320 | 0.6424 | 0.6305 | MF substantially more robust (LightGCN: 0.481–0.494) |
| ciaodvd_mf_classic | 23.6 | 0.8137 | 0.8119 | 0.8148 | LightGCN notably better (0.877–0.879) |
| ciaodvd_mf_het_donor | 19.3 | 0.7644 | 0.7675 | 0.7614 | LightGCN notably better (0.846–0.848) |
| metric (mean SEM, ) | LightGCN | MF |
|---|---|---|
| injected donor Euclidean dist. | ||
| injected donor cosine similarity | ||
| injected random-organic Euclidean dist. (control) | ||
| injected random-organic cosine similarity (control) |
| dataset | point | A2 AUROC | A2 AP | VEIL AUROC range | note |
|---|---|---|---|---|---|
| CiaoDVD | easy (classic) | 0.8397 | 0.3451 | 0.877–0.879 | comparable, VEIL slightly ahead |
| CiaoDVD | hard (het+donor) | 0.6171 | 0.1630 | 0.846–0.848 | A2 collapses far more; VEIL more robust here |
| ML-1M | easy (base) | 0.9822 | 0.8085 | 0.999–1.000 | comparable, both near-ceiling |
| ML-1M | hard (matched, no AUSH) | 0.5838 | 0.1669 | 0.530–0.536 | both collapse substantially |
| config | Stage 0 | Comp. A | Comp. B | Comp. C | total |
|---|---|---|---|---|---|
| supgnn_base | 2697.5s | 8.4s | 4.2s | 8.2s | 2719.6s (Stage0 log) / 20.9s (ABC log) |
| lightgcn_base | 2661.6s | 7.7s | 4.7s | – | 2679.3s (Stage0) / 12.4s (AB log) |
| lightgcn_aush | 2570.9s | 5.4s | 4.7s | 10.5s | 2591.6s |
| lightgcn_aush_het | 2272.8s | 4.8s | 4.2s | 8.4s | 2290.2s |
| lightgcn_aush_het_donor | 5239.2s | 4.6s | 4.4s | 9.5s | 5257.7s |
Recommender Systems
Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.
See all 573 papers