What do deepfake speech detectors actually learn?

This repository contains the source code for the Interspeech 2026 submission: "What do deepfake speech detectors actually learn?"

It provides a framework for training, evaluating, and interpreting speech deepfake detectors using Integrated Gradients (IG) to understand what features (and artefacts) self-supervised learning (SSL) models rely on to make their decisions.

Project Structure

• models/: Contains the model architectures (aasist.py, mhfa.py, sls.py, wavlm_aasist.py, wavlm_camhfa.py, wavlm_sls.py).
• utils/: Utility scripts for data loading (asvspoof5_dataset.py), processing (audio_utils.py), metrics (metrics.py), and visualization/Integrated Gradients support (ig_utils.py, ig_visualization.py).
• augmentation/: Code for augmenting speech data (RawBoost, Codec, NoiseFilter, etc.).
• artefacts_check/: Tools to compute and analyze correlations and Equal Error Rates (EER) of specific audio artefacts (compute_artefact_correlations.py, compute_artefact_stats.py).
• labelling_app/: A PHP-based web application for manually annotating/labeling audio features and artefacts.
• scores/: Scripts for error rate calculation and model fusion scoring.
• Root Scripts:
  • train.py & eval.py: Main scripts for training and evaluating the deepfake detectors.
  • compute_ssl_means.py: Script to compute mean SSL representations.
  • generate_ig_plots.py & plot_combined_ig.py: Scripts for generating Integrated Gradients visualizations.
  • download_model.py: Script to optionally pull base models/weights.

Usage

1. Environment Setup: Initialize the required dependencies using the provided environment scripts:

   # Linux/macOS
   source env/setup_env.sh
   # Windows
   env\setup_env.bat

   (Alternatively, install packages directly from env/requirements.txt)

2. Training: Configure your paths in config.py and run:

   python train.py --model camhfa

   Arguments for train.py:

   • --model: Model architecture to train (sls, camhfa, or aasist). [Required]
   • --data_dir: Path to ASVspoof5 root directory (defaults to config.DATA_DIR).
   • --train_protocol: Train protocol filename (defaults to config.TRAIN_PROTOCOL).
   • --dev_protocol: Dev protocol filename (defaults to config.DEV_PROTOCOL).
   • --output_dir: Directory to save checkpoints (defaults to config.OUTPUT_DIR).
   • --epochs: Number of training epochs (default: 10).
   • --batch_size: Batch size per GPU (default: 4).
   • --device: Computation device (cuda or cpu).
   • --augment: Flag to apply data augmentation.
   • --freeze_wavlm: Flag to freeze the WavLM backbone entirely.
   • --warmup_epochs: Number of epochs to freeze WavLM before switching to end-to-end training (default: 0).

3. Generating Integrated Gradients (IG) Explanations: Visualize the learned representations using IG plots:

   python generate_ig_plots.py --input_csv selections.csv --audio_dir /path/to/flac/

   Arguments for generate_ig_plots.py:

   • --input_csv: Path to a CSV or text list containing FileID strings to process. [Required]
   • --audio_dir: Root directory containing the audio .flac files. [Required]
   • --output_dir: Directory to save the output .png plots and interactive .json data (default: outputs/plots).
   • --models_dir: Directory containing the .pt model checkpoints (default: models).

Paper Context

This repository provides the experimental framework addressing the core question of our IS26 submission: establishing exactly what cues state-of-the-art self-supervised and end-to-end deepfake systems utilize when distinguishing bona fide speech from spoofed audio.