The A-Files

Research toolkit for evaluating audio steganography and watermarking: payload recovery, signal quality, robustness, capacity and empirical detectability under a shared experimental protocol. Includes a Python API, declarative experiments and a browser-based research interface.

Documentation · UI guide · References

Quick start

Python 3.10–3.12:

pip install the-a-files
taf-eval direct-no-metrics

Bundled VCTK and LibriSpeech subsets support initial experiments. FFmpeg is needed for codec attacks; PESQ may require C++ build tools. See installation and optional models.

Methods · 30

• Sample and adaptive embedding: LSB, prime-factor interpolation, FBS-LSB, AAC-STC.
• Transforms: DCT-Delta-LSB, DWT-LSB, DCT-b1, norm-space, FSVC, blind SVD, LWT, wireless DWT-LSB, sync-code DWT-DCT, EMD.
• Phase: phase coding, improved phase coding.
• Spread spectrum and quantisation: DSSS, ISS, QIM / ST-DM.
• Echo: single echo, backward–forward echo, time-spread echo.
• Statistical: patchwork, histogram, low-frequency amplitude modification.
• Reversible: prediction-error expansion (PEE), which also restores the exact cover.
• Neural and approximated learned schemes: FGAS, LE-GA, AudioSeal, WavMark.

Mechanisms, implementation limits and paper references.

Attacks

• Noise: awgn, pink_noise, impulse_noise.
• Coding: codec, with mp3, aac, opus, vorbis shortcuts.
• Filtering: low_pass, high_pass, band_pass, notch, smoothing.
• Sampling and quantisation: resample, clock_drift, bit_depth.
• Amplitude: gain, clipping, compression_dynamic.
• Time and pitch: time_shift, crop, zero_padding, sample_jitter, dropout, time_stretch, speed, pitch_shift.
• Acoustics: echo, reverb, acoustic_channel.
• Pipelines: streaming_upload, voice_call, broadcast, over_the_air, desync_attack.

Individual descriptions, parameters and scientific context.

Metrics · 25

• Signal and spectral distortion: SNR, SNRseg, fwSNRseg, WSS, LLR, CD, MCD, BSD, LSD, MRSC, SI-SDR, BSSEval v4.
• Perceptual quality: PESQ, ViSQOL Audio, Csig, Cbak, Covl, MOSNet.
• Intelligibility: CSII, NCM, STOI, eSTOI, STGI, wSTMI.
• Reverberation-related: SRMR.

BER, exact-message recovery, payload rate and processing time are reported separately. Definitions, score directions and references.

Research UI

The whole platform (PostgreSQL, API and web client) runs in Docker from the published images; only docker-compose.yml is needed:

curl -O https://raw.githubusercontent.com/pawel-kaczmarek/The-A-Files/master/docker-compose.yml
docker compose up -d --no-build

The web client is at http://localhost:3000 and the API at http://localhost:8000 (OpenAPI at /docs). Prepared corpora and uploads live in the taf-data volume. TAF_IMAGE_TAG selects a release (default latest; edge follows master). In a repository checkout, docker compose up -d --build builds the images locally; TAF_EXTRAS=platform,neural (or platform,ai) then adds the optional neural baselines or TensorFlow to the API.

For development, start only the database and run the API and web client locally:

docker compose up -d db
pip install -e ".[platform]"
taf-api

In a second terminal:

cd web
npm ci
npm run dev

Open http://localhost:3000. Create an experiment, select data, methods, payloads, attacks and metrics, review the execution plan, then start a run. Results include statistics, individual trials with audio playback, provenance and CSV/Markdown/LaTeX exports. English and Polish are available. Step-by-step UI guide.

Quality is measured separately for embedding and attack damage. Published results and local measurements remain distinct; implementation adaptations are documented in the method catalogue. GitHub Pages hosts the documentation; the research UI runs with the Python API and PostgreSQL.

Licence and authors

GPL-3.0-or-later. Paweł Kaczmarek and Zbigniew Piotrowski — Military University of Technology, Faculty of Electronics. Scientific background and bibliography.