Multi-Backbone Self-Supervised Ensembles for Audio Deepfake Detection and a Cross-Track Analysis of Generation-Detection Asymmetry

Authors: Seunghyun Kim, Junghyun Kim, Jiyoung Woo

Published: 2026-08-03 07:04:51+00:00

Comment: Accepted at CLEF 2026, ImageCLEF-Deepfake task. Published in CEUR-WS CLEF 2026 Working Notes

AI Summary

This paper presents team Go-To-Germany's participation in the ImageCLEF 2026 Audio Deepfake Detection and Generation task. Their detection system, a four-backbone self-supervised learning ensemble, achieved a high score of 0.9522, perfectly identifying participant-generated deepfakes. A cross-track analysis revealed a significant asymmetry: while the detection system was highly effective, their first-ranked generation entry still evaded a substantial percentage of detectors, highlighting an inherent challenge in deepfake detection.

Abstract

This paper describes the participation of team Go-To-Germany in the ImageCLEF 2026 Audio Deepfake Detection and Generation task. Our detection system, built on a four-backbone self-supervised learning (SSL) ensemble combining WavLM-Large, Wav2Vec2-XLS-R-300M, ECAPA-TDNN, and x-vector representations, achieved a final score of 0.9522 on the official ImageCLEF 2026 evaluation, with perfect accuracy (1.0000) on participant-generated deepfakes and 0.8875 on the held-out organizer ground-truth real data. For the Generation sub-task, our official team submission, an F5-TTS v1 baseline processed with a uniform reverberation pass and submitted as a deliberate anti-forensic probe, ranked first with a final score of 0.4304 (word error rate (WER) 4.99%, character error rate (CER) 2.07%); details of our four-model program (GLM-TTS, F5-TTS, XTTS v2, CosyVoice3), from which the official entry was drawn, appear in the paper. We present a cross-track analysis revealing a pronounced asymmetry: our detection system identifies 100% of participant-generated deepfakes, while our official generation entry, despite ranking first in the Audio Generation sub-task and evading 61.4% and 56.2% of participant and organizer detectors, attains a Final Score of 0.4304 against 0.9522 on the Detection side. We further report falsification-based ablation experiments (LOSO 56-speaker cross-validation, three-region backbone geometry, bootstrap confidence intervals, and PCA analysis) that motivate our architectural-insurance hypothesis for multi-backbone SSL ensembling. We complement these results with five cross-track insights and five pre-registered falsification experiments connecting generation-side evasion to detection-side design decisions, and we openly report an 11.25% false-positive gap on held-out organizer real recordings as the principal open challenge for deployment.

Key findings

The detection system achieved a final score of 0.9522, with perfect accuracy (1.0000) on participant-generated deepfakes and 0.8875 on held-out organizer ground-truth real data. A significant cross-track asymmetry was observed, where the detection system identified 100% of participant-generated deepfakes, but their top-ranked generation entry still evaded 61.4% and 56.2% of participant and organizer detectors, respectively. This highlights an 'architectural-insurance' hypothesis, suggesting that ensemble robustness comes from representational diversity across different backbone types rather than just a high backbone count.

Approach

The detection system employs a four-backbone self-supervised learning ensemble using WavLM-Large, Wav2Vec2-XLS-R-300M, ECAPA-TDNN, and x-vector representations. Each backbone's embeddings are fed into independent logistic regression classifiers, and their outputs are combined via a weighted ensemble. A top-960 conservative threshold strategy is then used for the final binary classification.

Datasets

ImageCLEF 2026 Deepfake Task (test set), organizer-provided Real recordings, team-synthesized fake utterances, LibriSpeech dev-clean (for 56-speaker extension).

Model(s)

WavLM-Large, Wav2Vec2-XLS-R-300M, ECAPA-TDNN, x-vector

Author countries

Republic of Korea

← Previous