SensiFake

SensiFake is a non-commercial research release of 4,499 unique images from OpenFake, SID-Set and RRDataset. It adds semantic sensitivity annotations to original source real/fake labels for research on sensitivity-aware authenticity detection. Authenticity and sensitivity are two separate targets.

The mutually exclusive subsets contain 1,101 authoritative human-annotated images and 3,398 automatic-only images with V3 silver labels. Every row retains the V3 prediction. final_sensitivity_* uses human ground truth where available and the V3 prediction otherwise. Human annotations are higher-confidence ground truth; automatic annotations are not equivalent to human labels.

See DATASET_CARD.md for the full dataset card.

Composition

Source	Unique images	Human	Automatic-only
OpenFake	1500	380	1120
SID-Set	1500	270	1230
RRDataset	1499	451	1048

Authenticity: 2,249 real / 2,250 fake. Source authenticity is preserved; the sensitivity model never predicts real/fake.

Final sensitivity	All images	Human	Automatic-only
LOW (0)	1267	354	913
MEDIUM (1)	3008	595	2413
HIGH (2)	224	152	72

There were 4,500 selected physical files. Two RRDataset real files are byte-identical and have one SHA-256 identity, leaving 4,499 unique images. Both original paths are retained in provenance; the image is copied once.

Labels and metadata

• real_fake_label: 0 real, 1 fake; real_fake_label_name retains the text label.
• Sensitivity class indices: 0 low, 1 medium, 2 high. These are not the historical rubric's 0–5 aggregate scores.
• annotation_type: human or automatic; sensitivity_label_quality explicitly marks automatic rows as silver.
• Human fields: human_annotation_available, human_annotation_source, human_sensitivity_score, human_sensitivity_label.
• Automatic fields: predicted_sensitivity_score, predicted_sensitivity_label, p_ge_1, p_ge_2, model_version, checkpoint_sha256.
• Final targets: final_sensitivity_score, final_sensitivity_label select the human label first.
• Identity/provenance: image_id, sha256, release-relative image_path, source_dataset, source_original_path, all paths in source_original_paths_json, source URL/license, original ID/revision where available, and authenticity provenance.
• v3_supervision_split preserves training, validation, frozen_test, or not_used; it does not create a new split. Empty human fields mean no human annotation, never LOW by default.

Every column is defined in label_schema.json. p_ge_1 and p_ge_2 are cumulative ordinal sensitivity probabilities, not three-class softmax scores or authenticity confidence.

Structure

README.md, DATASET_CARD.md, LICENSE, CITATIONS.md
metadata/       master CSV, two subset CSV views, schema, sources, statistics, audit
human_annotated/images/    human-supervised image files
autoannotated/images/    automatic-only image files
model_info/     V3 summary, metrics, model card (no binary checkpoint)

All CSV image paths are relative to this directory. The three CSVs are intentional views: use the master once or combine the two disjoint subsets, never concatenate all three. Source-original paths are provenance references to curator inputs, not paths to files inside this release.

Recommended usage

Use source authenticity for real/fake detection and keep sensitivity as a separate ordinal target. Use the human subset for evaluation; treat the automatic-only subset as noisy training supervision. The 769 training and 192 validation images were used for V3 development; the 140 frozen_test images were used for final comparisons and model selection. Do not use their V3 predictions as independent ground truth or mix these roles when reproducing the model evaluation.

Upload the directory as one Kaggle Dataset without an extra nested archive. For Hugging Face, the README declares three CSV configurations with a corpus split, preventing accidental concatenation of duplicate metadata views. Image paths remain ordinary strings pointing to included files; no custom loading script is needed to interpret the release.

Silver-label model and limitations

V3 uses a trained ResNet50 backbone and a two-output ordinal head, with calibrated thresholds t1=0.50, t2=0.44. Its frozen 140-image test results are accuracy 0.757143, macro-F1 0.708611, balanced accuracy 0.709889, and ordinal MAE 0.242857. See model card.

HIGH test support is only 9 images. LOW recall is 0.576923, and OpenFake performance is weaker than other sources. The reviewed batch was targeted, the source selections are not representative population samples, and the fixed test was reused for model comparisons. Labels are subjective and may miss contextual risk; silver labels may contain errors. This is not a production moderation or safety decision system.

License and attribution

SensiFake's added annotations, curation, metadata, deduplication and organization are licensed CC BY-NC 4.0. This is a non-commercial release. SensiFake does not claim ownership of original images. Original dataset and image terms continue to apply: OpenFake CC BY-NC 4.0; SID-Set and RRDataset CC BY 4.0. The RRDataset code repository's MIT license is distinct from the dataset license. See LICENSE and CITATIONS.md.

Official source references and explicit outstanding verification items are in source_datasets.json. The RRDataset license record and formal bibliography, per-image upstream credit obligations, and the curator-approved SensiFake release citation need manual completion before upload. Nothing has been uploaded.

Citation

Attribute the SensiFake contributions and the original source datasets. The publication URL/DOI and approved SensiFake author list have not been assigned; do not infer them from local paths. See CITATIONS.md for source references and completion notes.