GraphSeal: Graph-Augmented Speech Deepfake Detection via Temporal-Spectral Artifact Propagation

Authors: Lee Seung-woo, Bowen Qi, Kim Min-jun, Jang Won-young

Published: 2026-11-04 08:52:12+00:00

AI Summary

This paper introduces GraphSeal, a novel speech deepfake detection system that addresses the limitations of existing detectors by modeling long-range spectral inconsistencies. GraphSeal constructs a dynamic temporal-spectral graph over speech segments, propagating artifact cues through graph attention layers to produce a bonafide-vs-spoof decision. This approach allows the network to capture suspicious spectral patterns that may span non-adjacent regions of an audio signal, enhancing detection accuracy across various deepfake benchmarks.

Abstract

Current speech deepfake detectors treat each utterance as a flat sequence of frames, limiting their ability to capture long-range spectral inconsistencies that span non-adjacent regions of an audio signal. We propose \\textbf{GraphSeal}, which constructs a temporal-spectral graph over short speech segments and propagates artifact cues through graph attention layers before a global readout produces the bonafide-vs-spoof decision. Each node in the graph corresponds to a fixed-length segment; edges encode temporal adjacency and spectral similarity, letting the network attend to distant frames that share suspicious spectral patterns. A multi-head graph attention mechanism weights neighboring nodes according to both learned features and edge attributes, and a hierarchical readout aggregates node representations into a single utterance-level embedding for binary classification. GraphSeal is front-end agnostic: it operates on top of any self-supervised or handcrafted feature extractor. On ASVspoof 2019~LA, ASVspoof 2021~LA, ASVspoof~5, and the In-the-Wild dataset, GraphSeal achieves equal error rates of 0.52\\%, 2.41\\%, 8.17\\%, and 5.24\\%, respectively, matching or outperforming prior methods that rely on fixed-topology graph networks or purely sequential architectures.


Key findings
GraphSeal achieves state-of-the-art or competitive EERs on all evaluated benchmarks, notably 0.52% on ASVspoof 2019 LA and 2.41% on ASVspoof 2021 LA. Ablation studies confirm the crucial role of both spectral and temporal edges, edge attributes, paralinguistic features, and the hierarchical readout in achieving high performance. The dynamic graph construction enables better generalization to unseen attack types and yields more discriminative, well-separated embeddings compared to prior methods.
Approach
GraphSeal processes speech waveforms by first extracting frame-level representations from a frozen self-supervised encoder. It then segments the audio into overlapping fixed-length segments, which become nodes in a dynamic graph. Edges are established based on both temporal adjacency and spectral similarity between segments. A multi-head graph attention mechanism propagates artifact cues across these nodes, considering learned features and edge attributes, and a hierarchical readout aggregates node representations into a single utterance-level embedding for binary classification.
Datasets
ASVspoof 2019 LA, ASVspoof 2021 LA, ASVspoof 5, In-the-Wild dataset
Model(s)
WavLM-Large, wav2vec 2.0-Large, Whisper-Medium (as front-ends for feature extraction), Graph Attention Networks (GATv2) for the graph module.
Author countries
South Korea, China