ALLA1N/rawnet2-la-clean-specialist

RawNet2 — Clean-Domain Anti-Spoofing Specialist (LA)

This repository contains a fine-tuned :contentReference[oaicite:0]{index=0} checkpoint optimized for clean-domain anti-spoofing environments.

The model was trained primarily on the :contentReference[oaicite:1]{index=1} benchmark and serves as one of the parent specialists used in the :contentReference[oaicite:2]{index=2} project.

---

Model Details

Property	Value
Architecture	RawNet2
Domain Specialization	Clean-domain synthetic speech detection
Training Dataset	ASVspoof 2019 Logical Access
Input	Raw mono waveform
Target Sample Rate	16 kHz
Framework	PyTorch

---

Intended Use

This checkpoint is intended for:

• audio anti-spoofing research
• clean-domain deepfake detection
• genetic weight merging research
• transfer learning
• robustness comparison studies

It is NOT optimized for:

• noisy environments
• social-media-distributed audio
• codec-heavy in-the-wild conditions

For robustness-oriented conditions, see:

:contentReference[oaicite:3]{index=3}

---

Training Configuration

Parameter	Value
Optimizer	Adam
Learning Rate	Default RawNet2 training schedule
Loss Function	Weighted Cross Entropy
Batch Strategy	Standard balanced anti-spoofing training
Hardware	NVIDIA RTX Ada Generation GPU

---

Repository Contents

File	Description
best_model.pth	Best validation checkpoint
model.py	RawNet2 architecture definition

---

Usage

Load Checkpoint

import torch
from model import RawNet

model = RawNet()

checkpoint = torch.load(
    "best_model.pth",
    map_location="cpu"
)

model.load_state_dict(checkpoint)
model.eval()

---

Relationship to MeGA-IA

This model serves as the:

Clean-domain specialist parent model

inside the MeGA-IA genetic weight merging framework.

Its role is to preserve:

• clean acoustic priors
• studio-quality spoof discrimination
• low-noise spectral sensitivity

while complementary ITW-specialized models contribute robustness features.

---

Limitations

• Limited robustness to social media artifacts
• Performance degradation under heavy compression
• Reduced generalization to unseen in-the-wild distributions

---

Citation

If you use these weights, please cite:

@inproceedings{ahmad2026megaia,
  title     = {MeGA-IA: Genetic Algorithm-Driven Weight Merging for In-the-Wild Deepfake Detection},
  author    = {Ahmad, Awwab Ext and Ahmed, Rayan and Munir, Uwaid},
  booktitle = {Proceedings of the 23rd International Bhurban Conference on Applied Sciences and Technology (IBCAST)},
  year      = {2026},
  note      = {Under Review}
}

Please also cite the original RawNet2 framework:

@inproceedings{jung2020rawnet2,
  title     = {RawNet2: End-to-End Anti-Spoofing with Raw Waveform Modeling},
  author    = {Jung, Jee-weon and others},
  booktitle = {Proceedings of Interspeech},
  year      = {2020}
}

---

License

This repository is released under the :contentReference[oaicite:4]{index=4}.

Model weights are provided for research and benchmarking purposes.