C-MCSS-Mamba: Counterfactual Mechanism-Contrastive Selective Scan Within Mamba for Block-Causal Speech Deepfake Detection

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Speech deepfake detection (SDD) is commonly formulated as offline utterance-level binary classification, which limits early decisions in streaming communication and provides little mechanism-level evidence for a spoofing prediction. We propose C-MCSS-Mamba, a block-causal SDD framework that processes fixed-duration audio blocks with a partially fine-tuned XLS-R frontend and carries detection states across blocks through Mamba. Its selective scan is replaced by the proposed Counterfactual Mechanism-Contrastive Selective Scan (C-MCSS), which maintains shared, text-to-speech (TTS), and voice conversion (VC) states. Frame-level TTS/VC hypotheses directly modulate state update, retention, and forgetting. A Causal Counterfactual Evidence Reliability Module (C-CERM) further suppresses transient mechanism evidence through a recurrent reliability state that adaptively regulates these dynamics. Model-internal TTS/VC evidence is obtained by counterfactually disabling the corresponding recurrent state and measuring the resulting spoof-logit decrease. The model produces prefix-level spoof decisions and contribution-based state-dependence measures; these measures only characterize the model’s internal decision process, and do not identify the physical speech generation process of an individual sample. A joint objective combines detection, early-prefix, mechanism, contrastive, and intervention supervision. Experiments demonstrate competitive detection performance, achieving 0.47% EER on the ASVspoof 2019 LA evaluation set while maintaining effectiveness across four cross-dataset benchmarks.