Does Cognitive Load Affect Human Accuracy in Detecting Voice-Based Deepfakes?

TL;DR

The paper investigates whether cognitive load affects human accuracy in detecting voice-based deepfakes. It uses two experiments with 30 participants: Experiment 1 imposes a 1-back task to create a dual-task condition on audio stimuli, while Experiment 2 adds B-roll video to test audiovisual context. The findings show that a light cognitive load does not consistently reduce detection accuracy, and that watching related video content can significantly improve performance, likely due to an accessory-stimulus effect. These results inform the design of media literacy interventions and platform-level defenses by highlighting the potential benefits of audiovisual cues in deepfake detection and the resilience of human perception under modest cognitive load.

Abstract

Deepfake technologies are powerful tools that can be misused for malicious purposes such as spreading disinformation on social media. The effectiveness of such malicious applications depends on the ability of deepfakes to deceive their audience. Therefore, researchers have investigated human abilities to detect deepfakes in various studies. However, most of these studies were conducted with participants who focused exclusively on the detection task; hence the studies may not provide a complete picture of human abilities to detect deepfakes under realistic conditions: Social media users are exposed to cognitive load on the platform, which can impair their detection abilities. In this paper, we investigate the influence of cognitive load on human detection abilities of voice-based deepfakes in an empirical study with 30 participants. Our results suggest that low cognitive load does not generally impair detection abilities, and that the simultaneous exposure to a secondary stimulus can actually benefit people in the detection task.

Figures (8)

• Figure 1: Our taxonomy of indicative characteristics of stimuli that participants in various studies reported using to distinguish between bona fide and spoofed voices. The categories are detailed in the last part of Section \ref{['related-work']}
• Figure 2: Real-world attack model for spreading fake news videos with voice-based deepfakes. The audience is tricked into believing to hear some well-known newsreader---and thus that the news is spread by the reader's news outlet. Our study is designed to mirror corresponding attacks.
• Figure 3: Flow chart of the conducted study consisting of single-task and dual-task conditions of Experiment 1 and the video condition in Experiment 2. The decisions a) and c) are mechanisms to randomize the order of conditions and ensuring that both conditions have been performed once. The decision b) and d) repeat the listening task for 24 and 8 different stimuli in Experiment 1 and 2, respectively.
• Figure 4: Accuracy in detecting voice clones in the single- and dual-task conditions for spoofed and bona fide stimuli averaged by participant.
• Figure 5: Analysis of correlation between primary (voice-clone detection) task accuracy and secondary task (1-back task) $F_1$ of the participants. There is a weak negative correlation (Pearson $r=-0.19$) indicating that a trade-off of cognitive capacities exists.