🤖 AI Summary
This work addresses the limited generalization of existing audio deepfake detection methods in cross-dataset scenarios, where performance degrades significantly on unseen synthetic speech. To tackle this challenge, the authors propose FlowFake, the first approach to integrate Liquid Time-constant (LTC) neural networks into deepfake detection. By modeling hidden states through learnable ordinary differential equations, FlowFake adaptively captures spectrotemporal and prosodic artifacts across multiple time scales. Remarkably, with only 34K parameters, FlowFake achieves up to 79.97% accuracy across four cross-domain benchmarks—matching the performance of models 300 times larger—and substantially enhances cross-dataset generalization under stringent parameter constraints.
📝 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