FlowFake: Liquid Networks for Audio Deepfake Detection

Authors: Shivaay Dhondiyal, Divyansh Sharma, Dinesh Kumar Vishwakarma

Published: 2026-06-17 20:32:32+00:00

Comment: Accepted at the Workshop on Learning to Listen: Machine Learning for Audio at ICML 2026

AI Summary

FlowFake introduces a Liquid Time-Constant (LTC) neural architecture for audio deepfake detection, specifically addressing the challenge of cross-dataset generalization. The model's hidden state evolves via a learned Ordinary Differential Equation (ODE) with per-neuron adaptive time constants, enabling it to capture multi-timescale speech artifacts from spectral to prosodic cues. This approach leads to robust detection and strong generalization across unseen forgery types and acoustic conditions, significantly outperforming existing methods in parameter efficiency.

Abstract

Audio deepfakes generated by neural text-to-speech and voice-cloning systems threaten speaker verification and public discourse at scale. The core challenge is cross-dataset generalization: detectors trained on one synthesis pipeline collapse on unseen forgeries. We argue that this failure is primarily because of structural synthetic speech artifacts which are multi-timescale trajectory anomalies. Though every existing detector aggregates a fixed-window frame statistics, this misaligns the architecture with the signal. We propose FlowFake, a Liquid Time-Constant (LTC) architecture whose hidden state evolves via a learned ODE, with per-neuron adaptive time constants simultaneously resolving spectral (10ms) and prosodic (2s) cues. At only 34K parameters FlowFake achieves formal BIBO stability and O(dt^4) integration error. On a four-dataset cross domain benchmark (ASVspoof2019-LA, FakeOrReal, InTheWild, MLAAD), FlowFake reaches 75.29% on ASVspoof2019 trained only on FakeOrReal and 79.97% trained only on MLAAD. It outperforms RawGAT-ST and Whisper-DF on every evaluated pair and matching SSL Wav2vec2 (300x larger) at 0.01% of its parameter count. The source code is available on : https://github.com/GhostRider2023/FlowFake

Key findings

FlowFake achieves 75.29% accuracy on ASVspoof2019 when trained on FakeOrReal and 79.97% when trained on MLAAD. It consistently outperforms baselines like RawGAT-ST and Whisper-DF on cross-domain pairs and matches self-supervised learning (SSL) Wav2vec2, which is 300 times larger, using only 0.01% of its parameters. The model exhibits stable performance with dramatically lower cross-seed variance and strong zero-shot transfer capabilities, such as 90.41% accuracy on WaveFake.

Approach

FlowFake processes log-Mel spectrograms through a convolutional encoder, then feeds these embeddings into a Liquid Time-Constant (LTC) cell. The LTC cell models the hidden state's evolution using a learned ODE, integrating it with a 4th-order Runge-Kutta method, and employs adaptive time constants to resolve both fast spectral and slow prosodic cues for deepfake artifact detection. A final FC layer classifies the audio.

Datasets

ASVspoof2019-LA, FakeOrReal (FoR), InTheWild (ITW), MLAAD v1, WaveFake, LJSpeech

Model(s)

FlowFake (a Liquid Time-Constant (LTC) architecture), which incorporates a convolutional encoder, an LTC cell (solving a learned ODE), and a two-layer fully connected head.

Author countries

India

← Previous