Audio Deepfake Detection via a Fuzzy Dual-Path Time-Frequency Attention Network

Area of Science:

• Artificial Intelligence
• Information Security
• Signal Processing

Background:

• Audio deepfake technology poses significant threats to information security due to advancements in speech synthesis.
• Existing detection methods struggle with robustness against environmental noise, signal compression, and subtle fake audio features.
• Highly concealed fake audio remains difficult to identify effectively with current techniques.

Purpose of the Study:

• To develop a robust and accurate method for detecting audio deepfakes.
• To address the limitations of existing methods in handling noisy and compressed audio.
• To improve the identification of highly concealed fake audio through advanced feature characterization and enhancement.

Main Methods:

• Proposes a Dual-Path Time-Frequency Attention Network (DPTFAN) incorporating Pythagorean Hesitant Fuzzy Sets (PHFS) for uncertainty modeling.
• Employs a dual-path attention mechanism in time and frequency domains to improve feature representation and discriminative power.
• Introduces a Lightweight Fuzzy Branch Network (LFBN) for explicit enhancement of ambiguous audio features, balancing performance and efficiency.

Main Results:

• Achieved 98.94% accuracy on the ASVspoof 2019 LA dataset.
• Reached 99.40% accuracy on the FoR (Fake or Real) dataset.
• Demonstrated superior performance and robustness compared to existing mainstream audio deepfake detection methods.

Conclusions:

• The proposed DPTFAN method offers excellent detection performance and robustness for audio deepfakes.
• Uncertainty modeling with PHFS and dual-path attention effectively characterizes and enhances fake audio features.
• The LFBN contributes to improved performance while maintaining computational efficiency in deepfake detection.