Speech deepfake detection stops at adjacent frames: GraphSeal links spectrally similar distant segments and cuts ASVspoof 5 error to 8.17%
Synopsis
Speech deepfake detectors usually treat an utterance as a flat sequence of frames, which makes long-range spectral inconsistencies hard to capture. GraphSeal turns short speech segments into graph nodes, connects them with both temporal-adjacency and spectral-similarity edges, and propagates artifact cues through graph attention before a global readout decides bonafide versus spoof. On top of a frozen self-supervised speech encoder it adds only 2.1M trainable parameters, reaching equal error rates of 0.52%, 2.41%, 8.17% and 5.24% on ASVspoof 2019 LA, ASVspoof 2021 LA, ASVspoof 5 and In-the-Wild.
Interpretation
The modeling unit for speech spoofing detection moves from a flat utterance sequence to a dynamic graph built over speech segments, with edges drawn from both temporal adjacency and spectral similarity, so the model can directly compare regions that are far apart in time but close in spectrum. Earlier detectors either processed the utterance as a monolithic sequence or used a fixed graph topology determined by the feature-extraction pipeline, so the structure could not adapt to each input or exploit data-dependent similarity. On ASVspoof 2021 LA the full graph reaches 2.41% equal error rate, rising to 3.12% without spectral edges and 3.97% without temporal edges, indicating the two edge types supply complementary information.
Artifact cues propagate across the graph through edge-aware graph attention, so after a few hops each node reflects not only its own spectral content but also anomaly patterns seen in its temporal and spectral neighborhood. Prior self-supervised detection back ends were mostly linear probes, attentive pooling or shallow MLPs that process frames independently and do not model relations between different temporal positions. With a frozen WavLM-Large front end, GraphSeal reaches 2.41% and 5.24% equal error rate on ASVspoof 2021 LA and In-the-Wild, where WavLM-AASIST with the same front end gives 2.85% and 6.31%.
The utterance decision comes from a hierarchical readout that combines mean pooling with learned attention weights to compress a variable-size graph into a fixed utterance-level vector, which a two-layer MLP maps to a bonafide-versus-spoof logit. Replacing the hierarchical readout with plain mean pooling raises the ASVspoof 2021 LA equal error rate from 2.41% to 2.71%, showing that attention weighting lets the model focus on artifact-bearing nodes. The whole graph module adds only 2.1M parameters on top of the frozen encoder and 12 ms of latency per utterance on a single A100, of which graph construction takes 3 ms and the three graph attention layers take 9 ms.
Perspective
The demonstrated setting is offline utterance-level detection trained on the ASVspoof 2019 LA training partition and evaluated on ASVspoof 2019 LA, ASVspoof 2021 LA, ASVspoof 5 and In-the-Wild, with waveforms truncated or padded to 4 seconds and the front-end encoder weights kept frozen.
The authors report two failure modes: irregular prosody with hesitations and non-verbal vocalizations can be misread as synthesis artifacts, while spectrally homogeneous neural-codec attacks leave spectral edges uninformative and degenerate the graph toward near-complete connectivity. Streaming graph construction for unbounded audio remains an engineering challenge, and the authors list recurrent latent reasoning, stronger front-end representations and privacy-preserving detection as open directions.
