license: cc-by-4.0
language:
- en
pretty_name: Generation-Time Poisoning of LLM-Generated Social Networks
task_categories: - graph-ml
tags: - graph-machine-learning
- social-networks
- synthetic-data
- llm
- graph-poisoning
- backdoor-attacks
- node-classification
- link-prediction
size_categories: - n<1K
configs: - config_name: p50_nodes
data_files: processed/p50/nodes.parquet - config_name: p50_edges
data_files: processed/p50/edges.parquet - config_name: p50_graphs
data_files: processed/p50/graphs.parquet - config_name: p200_nodes
data_files: processed/p200/nodes.parquet - config_name: p200_edges
data_files: processed/p200/edges.parquet - config_name: p200_graphs
data_files: processed/p200/graphs.parquet
Generation-Time Poisoning of LLM-Generated Social Networks
This dataset contains synthetic personas, LLM-generated social graphs, cached
text embeddings, and evaluation metrics for clean generation and three
generation-time attack families. All names and profiles are synthetic and do
not represent real people.
Dataset variants
| Variant | Nodes | Generator | Graph seeds per condition | Attack rates |
|---|---|---|---|---|
| p50 | 50 | Qwen3-Max | 10 | 10%, 20%, 30%, 40%, 50% |
| p200 | 200 | DeepSeek-V4-Flash | 3 | 10%, 20%, 30%, 40%, 50% |
| p200 | 200 | MiMo-v2.5-Pro | 3 | 10%, 20%, 50% |
Conditions use the following identifiers:
c: clean generation.f_rXX: feature poisoning in persona generation.s_rXX: structural poisoning in link generation.b_rXX: backdoor generation with a persona trigger and link bias.
The attack rate defines a global node budget, floor(N * rate), capped by the
size of the eligible candidate set. It is not the realized fraction of changed
graph nodes or edges. Structural poisoning changes link generation only and
therefore reuses clean personas and clean persona embeddings.
Repository layout
raw/
p50/{personas,node_ids,arrays,graphs}/
p200/{personas,node_ids,arrays,graphs}/
processed/
p50/{nodes,edges,graphs}.parquet
p200/{nodes,edges,graphs}.parquet
metrics/
p50/*.csv
p200/*.csv
metadata/
manifest.csv
generation_config.json
checksums.sha256
P200 raw and metric filenames include the generator name so that the DeepSeek
and Mimo artifacts can coexist. Legacy Mimo source files are published with themimo-v2.5-pro namespace added; their contents are unchanged. Graph files are
grouped by condition under graphs/<condition>/.
The files under raw/ preserve the generated content exactly. Some adjacency
file headers contain the historical script names used when the experiments
were run; these provenance headers were intentionally not rewritten after the
codebase was reorganized.
The normalized node tables contain one row per persona and condition. The edge
tables contain directed adjacency-list entries for every graph seed. The graph
tables provide condition, generator, seed, node count, and edge-count metadata.
Cached arrays include embeddings and labels for clean, feature-poisoning, and
backdoor persona sets. Structural conditions reuse the clean arrays by design.
The raw/p50/arrays/embeddings_b_r*_trigger_masked.npy files cache backdoor
persona embeddings after removing the literal trigger phrase. They support
offline recomputation of the topology-mediated probability shift and are the
authoritative fallback when the experiment repository has no local cache.
Metrics
The metrics/ directory contains the final tables used to evaluate structural
statistics, feature separability, GCN/SGC node classification, targeted
backdoor attack success and clean accuracy, and link-prediction AUROC. API cost
logs are intentionally excluded because they are operational records rather
than scientific observations.metrics/p50/backdoor_probability_gap.csv reports the corresponding
topology-mediated probability-gap analysis.
Integrity and reproducibility
metadata/manifest.csv records each artifact's source path, dataset variant,
condition, attack family, attack rate, generator, graph seed, byte size, and
SHA-256 digest. metadata/checksums.sha256 covers every published file except
the checksum file itself.
LLM-backed generation services can change over time, so the raw artifacts are
the authoritative inputs for reproducing the reported metrics. Re-running the
same prompts and seeds with a later service version may not recreate identical
graphs.
Limitations and responsible use
The personas and political affiliations are simulated from demographic
distributions and model-generated text. They must not be interpreted as
measurements of real individuals or as evidence about demographic groups.
Generated names may coincidentally resemble real names. The data is intended
for research on graph generation, robustness, and auditing, not for profiling
or consequential decisions about people.
License
The dataset is released under CC BY 4.0. See LICENSE_DATA.md for scope and
attribution requirements. Source code is licensed separately.