Skip to content
KitploitKITPLOIT
ToolsExploitsBlog
Log in
Submit
ToolsExploitsBlog
Submit

Hacking, PenTest, and Cybersecurity Tools for Your Security Arsenal!

Kitploit is a directory of hacking, cybersecurity, and pentesting tools. Discover the latest project updates to find vulnerabilities, analyze systems, automate testing, and strengthen your security.

FeedsContactPrivacy© 2026 Kitploit

Tool Directory

Categories

View all categories
Loading categories
SwordStamp — Semantic Watermarking with Order-Robust Detection over Sub-sentence Units | Kitploit
Tools/GitHubGitHub/d-diaa/swordstamp
CryptographyPapers & ResearchLearning & EducationCurated ResourcesAI Security
GitHubd-diaa/swordstamp

SwordStamp

Semantic Watermarking with Order-Robust Detection over Sub-sentence Units

View Repository
1342 months agoNot yet reviewed
Share

SwordStamp

SwordStamp adds order-robust detection and semantic-span units to embedding-based semantic watermarks. This artifact contains the exact paper configuration, generation and detection code, the unchanged attack suite, quality evaluation, and deterministic result extraction and plotting.

Paper: Semantic Watermarking with Order-Robust Detection over Sub-sentence Units (arXiv:2608.27666)

Headline results

At 5% false-positive rate and a 90% content-preservation requirement, the scheme-specific embedding displacement attack (EDA-S) achieves:

WatermarkEDA-S attack success rate
SemStamp47.9%
k-SemStamp33.3%
PMark46.1%
SAMark32.6%
SwordStamp18.7%
k-SwordStamp10.8%

Relative to their published baselines, SwordStamp and k-SwordStamp reduce EDA-S success by 29.2 and 22.5 percentage points, at clean-fidelity costs of 1.8 and 4.4 points. Their clean detection rates are 94.9% and 97.0%. Under the stronger detector-access EDA-D stress test, maximum quality-gated success is 39.7% on k-SwordStamp versus 65.5% on k-SemStamp.

Strongest no-box attack success across content-preservation requirements

Clean fidelity versus strongest no-box attack success

These results cover English news continuations, one provider model, two provider encoders, and one EDA paraphraser/surrogate pair.

Plot a compiled bundle

visualization consumes only the compiled CSV/Parquet bundle. It fails on missing paper cells rather than drawing a partial figure.

uv sync --frozen --only-group plot --no-install-project
uv run --no-sync python -m visualization extract \
  --bundle results/paper --output results/paper
uv run --no-sync python -m visualization render \
  --bundle results/paper --output results/paper

This writes auditable tables under results/paper/tables/ and seven figures in PNG and PDF under results/paper/figures/. The anonymous review branch includes the compiled bundle; a full run on the release branch creates the same files with python -m visualization compile.

Embedding displacement attack (EDA)

The embedding displacement attack (EDA) is an adaptive whole-text paraphrasing attack. At each output unit it samples K continuations and keeps the candidate with the largest cosine displacement from its source anchor under a surrogate encoder. Because every candidate continues the full rewrite prefix, one objective can reword content, reorder it, and split or merge detector units.

No-box EDA sees the marked text and the public watermark design. It does not query the detector or generator and receives no secret key, provider encoder, or quality oracle.

VariantAnchor and unitConfiguration
EDA-Ppositional sentenceadaptive, anchor=positional, attacker unit sentence
EDA-Starget's public designpositional sentences for SemStamp, k-SemStamp, and PMark; bag/sentence for SAMark; bag/semspan for both SwordStamp variants
EDA-Dprovider detectororacle; inherits the defender encoder, partition, and segmentation

For bag anchoring, the paper uses attack.bag_agg=min: maximize distance from the nearest source anchor. Reported no-box runs use Qwen/Qwen2.5-3B-Instruct, BAAI/bge-base-en-v1.5, K={1,2,4,8,16,32,64}, temperature 1.0, top-p 0.95, at most 96 tokens per candidate, and at most 512 rewrite tokens.

Other watermark designers can call EDA directly:

from attacks.paraphrasing.adaptive import adaptive_attack_paraphrase

attacked = adaptive_attack_paraphrase(
    texts,
    base_model="Qwen/Qwen2.5-3B-Instruct",
    surrogate_model="BAAI/bge-base-en-v1.5",
    num_candidates=32,
    anchor="positional",       # EDA-P; use "bag" for an order-robust design
    bag_agg="min",
    segmentation_type="sentence",
    segmentation_backend="nltk",
)

For the paper's SwordStamp-aware EDA-S, set anchor="bag", segmentation_type="semspan", semcut_max_words=15, and semcut_window=5. Within this repository the equivalent CLI is:

uv run python -m attacks DATA_PATH --config PRESET \
  --set attack.paraphraser=adaptive \
  --set attack.custom_model=Qwen/Qwen2.5-3B-Instruct \
  --set attack.surrogate_model=BAAI/bge-base-en-v1.5 \
  --set attack.num_candidates=32 \
  --set attack.anchor=bag --set attack.bag_agg=min \
  --set segmentation.attacker_type=semspan \
  --set segmentation.attacker_backend=nltk

DATA_PATH is always the base corpus and PRESET identifies the watermarked cell. See attacks/README.md for all retained attacks and configuration fields.

Surrogate displacement versus provider-encoder displacement

Exact paper matrix

config/paper.py is the single source of truth. For each of LSH and k-means, scripts/experiments/swordstamp.sh runs this five-rung additive ladder with 64 provider candidates:

  1. context-dependent rejection over sentences;
  2. best-of-64 selection;
  3. a fixed valid set;
  4. diversity-aware ranking; and
  5. semantic spans with maximum 15 words and comparison window 5.

The grid contains only these ten cells and one no-watermark baseline. PMark is online-only; SAMark uses one flag pattern per run. EDA-D budgets are K={4,8,16,32,64}.

The comparison submodules adapt the public PMark and SAMark evaluation sources only at their dataset/result boundaries and shared false-positive calibration; the watermark algorithms are unchanged.

Complete experiment flow

Prerequisites are Git, uv, Python 3.11, CUDA for model runs, access to the pinned model revisions, and scratch storage.

git clone --recurse-submodules [email protected]:D-Diaa/SwordStamp.git
cd SwordStamp
bash scripts/setup.sh
uv run --frozen python scripts/check_artifact.py
  1. Prepare the deterministic, disjoint C4 partitions:

    uv run --frozen python scripts/prepare_c4.py --output-dir data
    
  2. Install and configure the pinned scheduler. The installer prints its user-level side effects before changing anything:

    bash scripts/install_gpu_scheduler.sh --print-plan
    bash scripts/install_gpu_scheduler.sh --yes
    gpu-scheduler init --gpus 0,1
    gpu-scheduler start
    
  3. Preview, then submit every physical shard:

    for shard in c4-val-def-256 c4-val-def-256b c4-val-def-512; do
      BASE="data/$shard" DRY_RUN=1 bash scripts/experiments/swordstamp.sh
      BASE="data/$shard" DRY_RUN=1 bash scripts/experiments/pmark.sh
      BASE="data/$shard" DRY_RUN=1 bash scripts/experiments/samark.sh
    done
    # Remove DRY_RUN=1 after inspecting the exact DAGs.
    
  4. Compile, extract, and render:

    uv run --frozen python -m visualization all \
      --bundle results/paper --output results/paper
    

The scheduler supplies CUDA_VISIBLE_DEVICES; do not select GPUs by probing nvidia-smi. Full reproduction is a multi-day GPU workload. ARTIFACT.md lists hardware, model revisions, expected outputs, and the claim-to-command map. scripts/experiments/README.md documents resume and force behavior.

Layout and license

Download Tool

Most Popular

View all →

Discover the most used tools by our community.

  • 1

    OsintHunter

    letox4/osinthunter

    🎯 The ultimate OSINT toolkit: investigation graph, confidence scoring, alias families, map, timeline and reports. 49 modules, 400+ sites, no API keys required.

    1.1k
  • 2

    ghostlock-app

    yukonga/ghostlock-app

    GhostLock One-Tap Execution App (CVE-2026-43499)

    628
  • 3

    looking-glass

    anonghosty/looking-glass

    CSIRT platform that ingests Shadowserver reports into MongoDB and replays them as interactive maps, analysis, findings, advisories and executive dashboards.

    413
  • 4

    jumpserver

    jumpserver/jumpserver

    JumpServer is an open-source Privileged Access Management (PAM) platform that provides DevOps and IT teams with on-demand and secure access to SSH, RDP, Kubernetes, Database and RemoteApp endpoints through a web browser.

    366
  • 5

    ARTEX

    baarai/artex

    LLM multi-agent autonomous penetration testing platform with planner/worker agents, asset and exploration graphs, MITM traffic recording, and vulnerability retesting.

    329
  • 6

    Aliens_eye

    arxhr007/aliens_eye

    AI powered OSINT engine to discover and map 7,700+ social media identities across the web

    287

Explore all tools

Browse our collection of tools

View all tools →