How Meta-Learning Shapes LoRA Adapter Geometry in Speech Deepfake Detection

Authors: Ivan Kukanov, Janne Laakkonen, Ville Hautamäki

Published: 2026-07-24 06:15:39+00:00

Comment: 7 pages, 5 figures, 3 tables. Submitted to SLT 2026 IEEE

AI Summary

This paper investigates how meta-learning for domain generalization (MLDG) shapes the geometry of low-rank adapters (LoRA) in speech deepfake detection compared to empirical risk minimization (ERM). By analyzing the empirical Fisher on the trained adapters, the study reveals that MLDG reorganizes the loss-relevant capacity within the adapter, concentrating it in query/key projections while distributing it in the output projection.

Abstract

Meta-learning for domain generalization (MLDG) improves out-of-distribution speech deepfake detection over empirical risk minimization (ERM) when both objectives train low-rank adapters on the same frozen self-supervised speech model. Because the architecture and adapter capacity are held fixed, this gap points to differences in how the training objective shapes the adapter, yet the field characterizes objectives through error rates rather than through the geometry of the solution they reach. We introduce a descriptive diagnostic for this question: holding architecture, rank, data, and seeds fixed and varying only the objective, we use the empirical Fisher on the finished adapter to compare the geometry that ERM and MLDG leave behind. We characterize each adapter with effective-rank diagnostics that separate where the adapter changes from where those changes matter to the loss, resolved by projection and by depth. Applied to ERM and MLDG, the diagnostic shows that the objective does not reshape all adapter projections alike: the loss-relevant update concentrates in the query and key projections while becoming more distributed in the output projection, consistently across six corpora and most strongly in the upper layers. The same contrast appears in the merged update independently of the low-rank factorization, indicating that it reflects the geometry of the effective update rather than the parameterization. These results show that the gap between ERM and MLDG is not only a difference in error rate, but a difference in how loss-relevant capacity is organized inside the adapter, and that loss-aware adapter geometry is a way to see it.

Key findings

MLDG concentrates the loss-relevant update in the query and key projections (up to -33% RankMEF decrease) while distributing it in the output projection (up to +42.2% RankMEF increase), especially in the upper layers. This geometric difference, consistently observed across datasets and seeds, demonstrates that the improved out-of-distribution generalization of MLDG stems from a distinct organization of loss-relevant capacity within the LoRA adapters, rather than just error rate differences.

Approach

The authors analyze the geometry of LoRA adapters trained with either ERM or MLDG on a frozen self-supervised speech model (Wav2Vec 2.0–AASIST). They use a descriptive diagnostic combining the empirical Fisher and effective rank to quantify where the loss-relevant adapter updates are concentrated across different projections and layers, and how this differs between the two training objectives.

Datasets

ASVspoof 2019 LA Eval, ASVspoof 2021 LA, ASVspoof 2021 DF, ASVspoof 5, InTheWild, FakeAVCeleb

Model(s)

Wav2Vec 2.0 XLSR-53 (front-end) with AASIST (back-end) and LoRA adapters

Author countries

Singapore, Finland

← Previous