Contextual Consistency Analysis: A Longitudinal Framework for Detecting Deepfake Audio

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The existence of high-quality audio deepfakes is a major threat to the integrity of digital communication, especially with respect to voice messages. Current detection mechanisms mainly rely on the analysis of an audio file independently. This makes the system vulnerable to attacks where high-quality deepfakes are synthesized. This paper proposes a novel approach to deepfake detection by scrutinizing the contextual consistency of vocal and environmental features over a series of voice messages sent by a single speaker. This proposed approach, named Contextual Consistency Deepfake Detector (CCDD), establishes a baseline profile based on voiceprint, prosody, and environmental features from a series of trusted voice messages. Each voice message is then scrutinized based on this established contextual baseline. Inconsistencies are identified by computing similarity scores. A high degree of inconsistency is an indication of a deepfake. Experiments were conducted with a simulated series of voice messages based on existing datasets, including the ASVspoof 2021 dataset. Results showed high efficacy with an accuracy rate of 96%, thus supporting the hypothesis that contextual inconsistencies are a robust feature to detect deepfakes.