Robust Deepfake Audio Detection via Multi-Level Intermediate Feature Fusion

Existing deepfake audio detectors, such as the XLSR-Mamba, primarily rely on final-layer features, a paradigm where discriminative details from intermediate layers can be progressively diluted during feature propagation. To alleviate this limitation, we propose a Multi-level Intermediate Feature Fusion (MIFF) module, which explicitly aggregates representations from all Mamba layers. MIFF employs an adaptive recalibration mechanism to dynamically estimate the contribution of each intermediate feature, producing a context-aware weighted fusion. This design preserves discriminative high-level semantics while enriching them with complementary multi-level details, leading to more comprehensive representations. On ASVspoof 2021 DF and In-The-Wild, our system achieves EER of 1.68% and 5.66%, corresponding to relative error reductions of 10.6% and 15.6% over XLSR-Mamba (1.88% and 6.71%), respectively, and demonstrating strong robustness to diverse spoofing attacks.