Know Your Enemy, Know Yourself

Rethinking SASV under Realistic Multi-Attack Scenarios

Dual-CM SASV architecture + Attack-aware Detection Cost Function (Aa-DCF)

---

📌 Overview

Recent Spoofing-Aware Speaker Verification (SASV) research has been largely LA-centric (Logical Access), overlooking vulnerabilities under PA (Physical Access, replay) attacks.

This repository accompanies our work:

“Know Your Enemy, Know Yourself: Rethinking SASV under Realistic Multi-Attack Scenarios”
(Interspeech 2026 submission)

We identify two fundamental limitations in current SASV systems:

1. Architectural limitation: A single unified CM struggles to model heterogeneous attack types (LA + PA).
2. Metric limitation: Conventional DCF-based metrics aggregate spoofing attacks into a single class, masking attack-specific vulnerabilities.

---

🔍 Problem Illustration

(Top) Architecture Limitation

• ❌ Single CM → Limited capacity for heterogeneous attacks
• ✅ Dual-CM → Specialized LA CM + PA CM

(Bottom) Evaluation Limitation

• ❌ Conventional metrics → Aggregated spoofing cost
• ✅ Proposed Aa-DCF → Attack-wise cost separation

---

🧠 Proposed Framework

1️⃣ Dual-CM SASV Architecture

We propose three fusion strategies:

• Cascading
• Score Fusion
• DNN Fusion

Each system uses:

• LA-specialized CM
• PA-specialized CM
• ASV

---

2️⃣ Attack-aware Detection Cost Function (Aa-DCF)

🔢 Cost & Prior Configurations

Setting	π_tar	π_non	π_LA	π_PA	C_miss	C_fa,non	C_fa,LA	C_fa,PA
Aa-DCF1	0.97	0.01	0.01	0.01	1	10	10	10
Aa-DCF2	0.93	0.01	0.05	0.01	1	10	10	10
Aa-DCF3	0.93	0.01	0.01	0.05	1	10	10	10

📌 Scenario Interpretation

• Aa-DCF1 → Balanced attack environment
• Aa-DCF2 → LA-dominant deployment
• Aa-DCF3 → PA-dominant deployment

---

Aa-DCF formulation:

Aa-DCF(t) =
C_miss * π_tar * P_miss(t)
+ C_fa,non * π_non * P_fa,non(t)
+ C_fa,LA * π_LA * P_fa,LA(t)
+ C_fa,PA * π_PA * P_fa,PA(t)

---

📂 Repository Structure

.
├── models/
│   ├── CM/
│   ├── ASV/
│   ├── SASV/
├── scores/
├── eval/
├── protocol/
│   └── protocol.txt
├── notebooks/
│   └── experiments.ipynb
├── assets/
│   └── overview.png
└── README.md

---

🧩 Models Used

🔹 Countermeasure (CM)

LA-trained AASIST

https://github.com/clovaai/aasist.git

PA-trained AASIST

Pretrained models: https://drive.google.com/drive/folders/1NIOs21SOxLSFuO5gvVN7twWDghh357fS

LA+PA-trained AASIST

Pretrained models: https://drive.google.com/drive/folders/1NIOs21SOxLSFuO5gvVN7twWDghh357fS

🔹 ASV Models

ECAPA-TDNN

https://github.com/TaoRuijie/ECAPA-TDNN.git

ResNet34

https://github.com/eurecom-asp/sasv-joint-optimisation.git

🔹 SASV Baselines

MFA-Conformer

SKA-TDNN

https://github.com/sasv-challenge/SASV2_Baseline.git

🔹 Proposed SASV

Dual-CM DNN Fusion

Pretrained weights:
https://drive.google.com/drive/folders/1NIOs21SOxLSFuO5gvVN7twWDghh357fS

Dual-CM Score Fusion & Dual-CM Cascading

Score Fusion does not require additional fusion training.
Simply use:

• LA-trained CM (AASIST)
• PA-trained CM (AASIST)
• ASV backbone (ECAPA-TDNN or ResNet34)

---

📊 Evaluation

Dataset:

• ASVspoof 2019 LA + PA combined evaluation set
• The corresponding trial protocol files (including target, nontarget, LA spoof, and PA spoof labels) are provided in:

protocols/protocol.txt

---

🧮 Computing Aa-DCF

Notebook:

notebooks/experiments.ipynb