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

2026-08-03

AI Summary

This work addresses the challenges of poor generalization and performance asymmetry between generation and detection in audio deepfake detection. We propose a multi-backbone self-supervised ensemble framework that integrates representations from WavLM-Large, Wav2Vec2-XLS-R-300M, ECAPA-TDNN, and x-vector, and conducts adversarial exploration using state-of-the-art TTS systems such as F5-TTS and GLM-TTTS. For the first time, we systematically reveal a pronounced asymmetry between generation and detection capabilities, formulate the “architectural insurance” hypothesis, and validate its causal mechanism through a pre-registered falsification experiment. On the ImageCLEF 2026 task, our detector achieves an overall score of 0.9522 (100% accuracy on fake audio, 88.75% on real), while generated speech attains remarkably low WER/CER of 4.99%/2.07%, evading 56.2%–61.4% of detectors. We further report an 11.25% false positive rate on organizer-provided real data, highlighting a critical open issue.

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.

Problem

• Audio Deepfake Detection
• Generation-Detection Asymmetry
• Self-Supervised Learning
• False Positive Rate
• Deepfake Generation

Innovation

• multi-backbone ensemble
• self-supervised learning
• audio deepfake detection
• generation-detection asymmetry
• falsification-based ablation