Multi-Branch Temporal-Spectral LSTM-CNN in Deepfake Audio Detection

Overview

Malicious deepfake media has been increasing steadily, with audio deepfakes posing an intensive cybersecurity and impersonation threat; multi -branch network machine -learning models have been overlooked in audio deepfake detection. Combined with multi -branch CNN, this project investigates perturbations to train a deep-learning algorithm to differentiate between bonafide and spoofed audio samples, aiming for the public standard of 82.5% accuracy. Spectral and temporal features were extracted through mathemati cal formulas from audio signals and resized to predefined fixed shapes using image -based extrapolation and compression, avoiding the traditional padding/truncating method which degrades data integrity. All features were normalized through amplitude normalization and aggregated into a list of arrays for the 11-branch CNN. Beta 1 used 75 bonafide and 75 spoofed files, while Beta 2 used 250 bonafide and 250 spoofed files; both used an 80:20 training-validation split. Beta 1 reached a peak validation accuracy of 74.13%, with p-value of 2.512e-35 for chi-square test, indicating statistical significance. Beta 2 improved upon Beta 1, receiving a validation accuracy of 95.78% with p -value of 1.221e-32, reaching the engineering goal. These findings indicate that the novel approach of multi-branch CNN combining both handcrafted and automatic feature extraction as input shows potential in improving audio deepfake detection. Future steps include implementing LSTM and applying audio transformations to mimic real-world conditions and distortions.

Competition history

• JSHS 2025 Category not listed

Resources

• National JSHS abstract book ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.