发表机构
Zhejiang University; Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security; Shanghai Institute for Advanced Study, Zhejiang University; Qingdao Institute of Software College of Computer Science and Technology, China University of Petroleum (East China)（浙江大学; 杭州高新区（滨江）区块链与数据安全研究院; 浙江大学上海高等研究院; 中国石油大学（华东）计算机科学与技术学院青岛软件学院）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
针对深度伪造检测模型参数过多难以部署的问题，提出任务感知的联合剪枝与蒸馏框架，结合跨域知识蒸馏与运动引导结构化剪枝，将模型压缩至31.9M参数，FLOPs降低6.3倍，性能仅下降1.30%。
AI 中文摘要
语音合成技术的进步使得深度伪造语音越来越逼真，对安全构成日益严重的威胁。虽然基于自监督学习（SSL）的检测器达到了最先进的性能，但其计算需求（通常超过3亿参数）阻碍了在资源受限设备上的部署。现有的压缩方法主要针对以内容为中心的任务设计，在直接应用于深度伪造检测时难以保持有竞争力的性能。我们提出了一种任务感知的联合剪枝与蒸馏框架，将跨领域知识蒸馏与运动引导的结构化剪枝相结合，以迁移伪造判别性知识并在激进压缩下保留关键结构。我们的框架将模型缩减至3190万参数，计算量降低6.3倍，与未压缩的基线相比，在多个数据集上的平均性能下降仅为1.30%，展示了在设备端部署的巨大潜力。
英文摘要
Advances in speech synthesis have made deepfake speeches increasingly convincing, posing growing threats to security. While self-supervised learning (SSL) based detectors achieve state-of-the-art performance, their computational demands (typically 300M+ parameters) prevent deployment on resource-constrained devices. Existing compression methods, designed mainly for content-centric tasks, struggle to maintain competitive performance when directly adapted to deepfake detection. We propose a Task-Aware Joint Pruning and Distillation framework that combines cross-domain knowledge distillation with movement-guided structured pruning to transfer forgery-discriminative knowledge and preserve critical structures under aggressive compression. Our framework reduces the model to 31.9M parameters with 6.3$\times$ FLOPs reduction, with an average performance drop of only 1.30\% across multiple datasets compared to the uncompressed baseline, demonstrating strong potential for on-device deployment.
Comments6 pages, 4 figures, accepted to Interspeech 2026
DOI：10.21437/Interspeech.2026-1766
