Two-Box Audio Deepfake Detection
A small anti-spoofing model that says whether speech is synthetic, and a glass-box layer that says why, in plain language.
Sole author. M.Tech minor project, guided by Prof. Srividhya Laxmi.
- Equal error rate, ASVspoof 2019 LA: 0.84% (source: Minor project report, Nirma University, Apr 2026)
- In-The-Wild, after fine-tuning: 0.85% (source: Minor project report, Nirma University, Apr 2026)
- Model size: 297K params (source: Minor project report, Nirma University, Apr 2026)
- In-The-Wild, before fine-tuning: 38.47% (source: Minor project report, Nirma University, Apr 2026)
Train the black box
AASIST on ASVspoof 2019 LA for 30 epochs on a single A100 GPU. Best development error 0.55% at epoch 23.
Measure the real-world gap
The trained model is tested, with no adaptation, on In-The-Wild: 31,779 clips of 58 public figures taken from social platforms, with unknown deepfake methods and real codecs.
Close the gap
Fine-tuning on In-The-Wild for 20 epochs, with the split made by speaker so no voice is in both training and test.
Open the glass box
Attention is read out of the heterogeneous graph-attention layers: 23 spectral nodes (which frequencies) and 29 temporal nodes (which moments).
Say it in words
A rule-based template states the peak band, the temporal pattern and the verdict, so the facts are fixed. A language model then rephrases it for a non-technical reader.