Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Zero-Shot to Zero-Lies: Detecting Bengali Deepfake Audio through Transfer Learning

Most. Sharmin Sultana Samu, Md. Rakibul Islam, Md. Zahid Hossain, Md. Kamrozzaman Bhuiyan, Farhad Uz Zaman

TL;DR

The paper benchmarks Bengali deepfake audio detection using BanglaFake, comparing zero-shot transfer from multiple pretrained models with fine-tuned detectors across six architectures. Zero-shot results show limited effectiveness (best: 53.8% accuracy), while fine-tuned models achieve strong performance, notably ResNet18 with 79.17% accuracy and 84.37% AUC. This work provides the first systematic Bengali deepfake audio benchmark and demonstrates the substantial gains from task-specific fine-tuning in a low-resource language. It lays the groundwork for robust, language-aware deepfake detection and guides future research on cross-lingual transfer and efficient deployment.

Abstract

The rapid growth of speech synthesis and voice conversion systems has made deepfake audio a major security concern. Bengali deepfake detection remains largely unexplored. In this work, we study automatic detection of Bengali audio deepfakes using the BanglaFake dataset. We evaluate zeroshot inference with several pretrained models. These include Wav2Vec2-XLSR-53, Whisper, PANNsCNN14, WavLM and Audio Spectrogram Transformer. Zero-shot results show limited detection ability. The best model, Wav2Vec2-XLSR-53, achieves 53.80% accuracy, 56.60% AUC and 46.20% EER. We then f ine-tune multiple architectures for Bengali deepfake detection. These include Wav2Vec2-Base, LCNN, LCNN-Attention, ResNet18, ViT-B16 and CNN-BiLSTM. Fine-tuned models show strong performance gains. ResNet18 achieves the highest accuracy of 79.17%, F1 score of 79.12%, AUC of 84.37% and EER of 24.35%. Experimental results confirm that fine-tuning significantly improves performance over zero-shot inference. This study provides the first systematic benchmark of Bengali deepfake audio detection. It highlights the effectiveness of f ine-tuned deep learning models for this low-resource language.

Zero-Shot to Zero-Lies: Detecting Bengali Deepfake Audio through Transfer Learning

TL;DR

The paper benchmarks Bengali deepfake audio detection using BanglaFake, comparing zero-shot transfer from multiple pretrained models with fine-tuned detectors across six architectures. Zero-shot results show limited effectiveness (best: 53.8% accuracy), while fine-tuned models achieve strong performance, notably ResNet18 with 79.17% accuracy and 84.37% AUC. This work provides the first systematic Bengali deepfake audio benchmark and demonstrates the substantial gains from task-specific fine-tuning in a low-resource language. It lays the groundwork for robust, language-aware deepfake detection and guides future research on cross-lingual transfer and efficient deployment.

Abstract

The rapid growth of speech synthesis and voice conversion systems has made deepfake audio a major security concern. Bengali deepfake detection remains largely unexplored. In this work, we study automatic detection of Bengali audio deepfakes using the BanglaFake dataset. We evaluate zeroshot inference with several pretrained models. These include Wav2Vec2-XLSR-53, Whisper, PANNsCNN14, WavLM and Audio Spectrogram Transformer. Zero-shot results show limited detection ability. The best model, Wav2Vec2-XLSR-53, achieves 53.80% accuracy, 56.60% AUC and 46.20% EER. We then f ine-tune multiple architectures for Bengali deepfake detection. These include Wav2Vec2-Base, LCNN, LCNN-Attention, ResNet18, ViT-B16 and CNN-BiLSTM. Fine-tuned models show strong performance gains. ResNet18 achieves the highest accuracy of 79.17%, F1 score of 79.12%, AUC of 84.37% and EER of 24.35%. Experimental results confirm that fine-tuning significantly improves performance over zero-shot inference. This study provides the first systematic benchmark of Bengali deepfake audio detection. It highlights the effectiveness of f ine-tuned deep learning models for this low-resource language.
Paper Structure (15 sections, 4 figures, 4 tables)

This paper contains 15 sections, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Background Study
  4. Pretrained Models for Zero-Shot Audio Deepfake Detection
  5. Deep Learning Models for Fine-Tuned Bengali Audio Deepfake Detection
  6. Evaluation Metrics
  7. Methodology
  8. Dataset and Preprocessing
  9. Zero-Shot Inference Models
  10. Fine-Tuning Models
  11. Experimental Setup
  12. Result Analysis
  13. Zero-Shot Classification
  14. Fine-Tuned Classification
  15. Conclusion and Future Works

Figures (4)

  • Figure 1: Proposed Methodology
  • Figure 2: Confusion matrix of the fine-tuned classification models.
  • Figure 3: ROC curve of the fine-tuned classification models.
  • Figure 4: DET curve of the fine-tuned classification models.