PDetective: Beyond Shallow Artifacts in Partial Deepfake Speech Detection
Published in Proceedings of ACM MM, 2026
Recommended citation: Junhao Wang, Kun Wang, Li Lu*, Chong Zhang, Peng Cheng, Zhongjie Ba, Kui Ren. "PDetective: Beyond Shallow Artifacts in Partial Deepfake Speech Detection." Proceedings of ACM MM. Rio de Janeiro, Brazil. pp. 10163-10172. 2026. doi: 10.1145/3767308.3835715.
ACM International Conference on Multimedia is the premier international conference dedicated to advancing research and applications across diverse multimedia fields. ACM MM is a CCF-A conference.
Abstract: With the rapid advancement of fine-grained speech synthesis, partial deepfake speech has emerged as an increasingly concerning threat. Recent end-to-end speech editing systems further exacerbate this challenge by enabling acoustically and perceptually seamless partial forgeries, leading to the systematic failure of existing partial deepfake detectors. Our investigation of existing partial deepfake detection models reveals that these models exhibit a critical learning bias toward model-specific shallow decoding artifacts, which fundamentally undermines their generalization. To address this issue, we propose PDetective, a framework that mitigates artifact-dependent decision-making through targeted training constraints, encouraging the model to learn more intrinsic and stable representations. We further introduce a large-scale partial editing spoofing dataset comprising more than 250k utterances to underpin and validate our framework. Experiments demonstrate that PDetective significantly improves cross-domain generalization, reducing EER by over 11% on unseen editing-based forgeries, and achieving an average EER reduction of over 30% relative to the best-performing baselines in cross-dataset evaluations.
