A Training-Free Proactive Defense Against Partial Speech Manipulation via Self-Embedding Steganography

Authors: Yigitcan Özer, Zhe Zhang, Wanying Ge, Xin Wang, Junichi Yamagishi

Published: 2026-08-26 01:45:35+00:00

Comment: 6 pages; 4 figures; 1 tables; accepted at Interspeech 2026

AI Summary

This paper proposes a training-free proactive defense against partial deepfake speech manipulation using self-embedding steganography. A clean speech signal embeds a compressed representation of itself, enabling post-hoc extraction and restoration. This method effectively detects partial deepfakes by comparing the received signal with its self-reconstruction, complementing existing passive defenses.

Abstract

Partial deepfake speech, where only limited segments of an utterance are synthesized or manipulated, poses a significant challenge to existing deepfake detection systems. As the proportion of spoofed regions decreases, passive detectors become increasingly unreliable, and accurate detection and restoration remain challenging. In this paper, we revisit audio steganography from a new perspective and propose its use as a proactive defense against partially deepfaked audio. In particular, we consider a self-embedding strategy in which a clean speech signal embeds a compressed representation of itself, enabling post-hoc extraction of reference content. We demonstrate how existing audio steganography methods can be repurposed to support detection of partial deepfakes through codec-based restoration. Experiments on a benchmark dataset show that the proposed approach complements passive defenses. Remarkably, the proposed method operates without any training, providing a robust and data-efficient alternative for partial deepfake detection.

Key findings

The proposed training-free method significantly outperforms existing passive deepfake detectors (LAV-DF, LAV-DF+, ResNet) in detecting partial speech manipulations, achieving EERs as low as 4.4% compared to ~50% for baselines. It demonstrates robust detection even when manipulated segments are sparse, highlighting its effectiveness in scenarios where passive detectors struggle.

Approach

The proposed method employs a self-embedding steganography strategy where a clean speech signal embeds a compressed representation of itself using a neural speech codec. Detection of manipulation is achieved by comparing the received (potentially manipulated) signal with its self-reconstruction via Dynamic Time Warping (DTW), which identifies inconsistencies without requiring training.

Datasets

AV-Deepfake1M Dataset (validation split, audio modality only)

Model(s)

SNAC (for neural speech codec representation), LSB embedding (temporally repetitive adaptation), Dynamic Time Warping (DTW)

Author countries

Japan

← Previous