Social-RDH-Bench

Social-RDH-Bench: Auditable Benchmark Dataset for RDH Steganalysis Social-RDH-Bench accompanies the paper: “RDH Steganalysis under Social Platform Re-encoding: A Unified Benchmark with Mechanism Analysis and Trustworthy Deployment”Forensic Science International: Digital Investigation, 2026. [Article DOI: fill when available] Dataset overviewRole: Audit-ready benchmark for forensic RDH steganalysis under social-style re-encoding (QQ / WeChat–like tracks), with mechanism- and deployment-oriented reporting.Platform tracks: m01 (QQ Shuoshuo–style), m02 (WeChat Moments–style), plus an original-domain track without social-network simulation.RDH mechanisms: LSB, DCT, and DWT embeddings at relative payloads {0.1, 0.2, 0.4} (protocol-defined).Sources: 16 anonymous source IDs (source_001–source_016) with device- and synthesis-aware coverage; identities are not disclosed in public metadata.Splits: Source-aware, device-aware, and protocol-frozen CSV splits for Exp1–Exp4 and bidirectional cross-platform settings, as described in the paper and frozen protocol file.Randomness: Five fixed seeds {42, 43, 44, 45, 46} for multiseed training/evaluation; aggregates report means (and related tests where applicable).File layout (this deposit) Typical contents of this record: Logical role Typical path in the archiveImage-level metadata (sanitized paths) 03_manifest/manifest.csvProtocol splits 02_splits/Frozen protocol & seed documentation 01_protocol/ (e.g. protocol_freeze.md)Pre-computed tables & checkpoints (where released) 04_results/ (often symlinked as results in full-tree workflows)Aggregation & figure scripts (subset) 05_scripts/Integrity & policy notes 00_docs/ Exact filenames may follow the repository’s Zenodo assembler; the semantic roles above are stable. Data collection and methodology (summary)Image pool: Built from curated, anonymized sources spanning device captures, desktop/UI captures, and controlled synthetic content, indexed under source_001–source_016 with no public mapping to raw device or vendor labels in released tables.RDH generation: Payload-limited LSB / DCT / DWT reversible embeddings were produced under the frozen experimental protocol (implementation details and hyperparameters are fixed in protocol_freeze.md, the paper).Social re-encoding: m01 and m02 tracks approximate QQ- and WeChat-style client/server re-encoding behaviour used in the benchmark; the exact pipeline (capture vs. simulation, client versions) is documented in the paper Methods and tied to the frozen split files. Note: This deposit may omit full-resolution pixel archives (large datasets/core/) and provide tabular splits, manifests, code, and aggregated results for auditing; pixel data may be distributed as a separate Zenodo record or controlled access per institutional policy. [Add second DOI or URL here if applicable.] Reproducibility and auditingSplits are source-aware with stable source IDs in CSVs.Fixed seeds support bit-level reproducibility of split assignments and training runs given the same code revision and pixel inputs.Multiseed aggregation supports uncertainty-aware reporting in summary tables.Independent groups can re-run the pipeline using splits + protocol + code from the project repository together with authorized pixel data.Citation When using Social-RDH-Bench, please cite the associated article: Zhu, N. (2026). RDH Steganalysis under Social Platform Re-encoding: A Unified Benchmark with Mechanism Analysis and Trustworthy Deployment. License Data and tabular results are released under CC BY 4.0 as stated in LICENSE. Software is under MIT or other licenses noted in the same file. Checkpoint (.pt) redistribution follows the terms in LICENSE.