Network-Aware Audio Deepfake Score Calibration Under WebRTC Channel Variability

Abstract

Audio deepfake detectors are commonly evaluated on source audio under controlled data conditions, whereas realtime voice communication can introduce transmission-dependent distortions through browser-based delivery, packet loss, jitter, and packet-loss concealment. This paper investigates networkaware audio deepfake score calibration for WebRTC-transmitted speech. Using CodecFake as the source dataset, we construct a speaker-disjoint audio deepfake subset and train an AASIST-L audio-only detector as the baseline system. We then build a WebRTC-based transmitted subset by replaying bona fide and spoofed utterances through browser-based audio communication paths and collecting both received audio and receiver-side network statistics. Instead of treating network statistics as independent spoofing cues, we use them as channel-quality information for calibrating the reliability of audio detector scores under channel degradation. The proposed protocol evaluates whether receiver-side WebRTC statistics can improve score reliability under mild, moderate, and severe network conditions without introducing label shortcuts.