Detection alone is insufficient to mitigate the harm by deepfake audio

Hongbin Liu, Youzheng Chen, Arun Narayanan, Athula Balachandran, Pedro Jose Moreno Mengibar, Lun Wang

Keywords: Deepfake detection

Abstract: This paper confronts the challenge of detecting increasingly sophisticated deepfake audio from advanced Text-to-Speech (TTS) systems with voice cloning. We posit that achieving high-accuracy, long-term detection of synthetic audio, particularly against motivated adversaries, is likely an unrealistic goal. This stance is supported by two primary observations. Firstly, the ongoing advancements in TTS and Synthetic Speech Detection (SSD) mirror an offline Generative Adversarial Network dynamic, with TTS as the generator and SSD as the discriminator. This suggests an eventual convergence towards synthetic speech that is nearly indistinguishable from human speech, making detection inherently challenging, if not impossible, especially as SSD development inherently lags behind TTS progress because SSD relies on TTS to generate training data. Secondly, current SSDs demonstrate a critical vulnerability to active, malicious evasion attacks, where the audio is carefully edited to bypass the target SSDs. Consequently, addressing deepfake audio demands a more systematic and multifaceted strategy, integrating approaches such as detection, legislative frameworks, watermarking technologies, robust enforcement mechanisms, and fostering cultural awareness.

Submission Number: 106