SONAR: Spectral-Contrastive Audio Residuals for Generalizable Deepfake Detection

2025-11-26

AI Summary

Deepfake audio detectors suffer from poor out-of-distribution generalization, primarily due to spectral bias—models over-rely on low-frequency semantic cues while neglecting high-frequency artifacts. To address this, we propose SONAR, the first unsupervised, architecture-agnostic frequency-domain disentanglement framework that explicitly treats high-frequency residuals as the core learning signal. SONAR employs a learnable high-pass filter (SRM), frequency-cross attention, and a frequency-aware Jensen–Shannon contrastive loss to jointly model and disentangle low-frequency semantic manifolds from high-frequency distortion manifolds. Evaluated on ASVspoof 2021 and real-world datasets, SONAR achieves state-of-the-art detection performance, converges four times faster than baseline methods, and significantly improves cross-distribution robustness and generalization.

Technology Category

• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Machine Learning: Unsupervised & Self-Supervised Learning
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

Deepfake (DF) audio detectors still struggle to generalize to out of distribution inputs. A central reason is spectral bias, the tendency of neural networks to learn low-frequency structure before high-frequency (HF) details, which both causes DF generators to leave HF artifacts and leaves those same artifacts under-exploited by common detectors. To address this gap, we propose Spectral-cONtrastive Audio Residuals (SONAR), a frequency-guided framework that explicitly disentangles an audio signal into complementary representations. An XLSR encoder captures the dominant low-frequency content, while the same cloned path, preceded by learnable SRM, value-constrained high-pass filters, distills faint HF residuals. Frequency cross-attention reunites the two views for long- and short-range frequency dependencies, and a frequency-aware Jensen-Shannon contrastive loss pulls real content-noise pairs together while pushing fake embeddings apart, accelerating optimization and sharpening decision boundaries. Evaluated on the ASVspoof 2021 and in-the-wild benchmarks, SONAR attains state-of-the-art performance and converges four times faster than strong baselines. By elevating faint high-frequency residuals to first-class learning signals, SONAR unveils a fully data-driven, frequency-guided contrastive framework that splits the latent space into two disjoint manifolds: natural-HF for genuine audio and distorted-HF for synthetic audio, thereby sharpening decision boundaries. Because the scheme operates purely at the representation level, it is architecture-agnostic and, in future work, can be seamlessly integrated into any model or modality where subtle high-frequency cues are decisive.

Problem

• Detecting deepfake audio that struggles with out-of-distribution generalization
• Addressing spectral bias where networks miss high-frequency artifacts
• Creating generalizable detection by separating natural and synthetic audio signals

Innovation

• Disentangles audio into complementary frequency representations
• Uses cross-attention to reunite long- and short-range dependencies
• Employs frequency-aware contrastive loss to sharpen decision boundaries