DeFake: Data-Efficient Adaptation for Generalized Deepfake Detection

Venue: iclr2026 (Withdraw) Authors: Debarshi Brahma, Suvam Dey, Anuska Roy, Soma Biswas OpenReview: https://openreview.net/forum?id=1tYymCWmG5

Relevance

LLM score: 1/3 — The paper focuses on data-efficient adaptation (few-shot learning for deepfake detection), which is tangential to the energy-efficiency and hardware-aware training priorities of the Sutro Group. Keyword hits: data-efficient

TLDR

(none provided)

Abstract

While deepfake detection methods have seen significant progress, current approaches focus on detecting fully synthetic or partially manipulated images separately, and often rely on large amounts of labeled training data. However, in real world, deepfakes can originate from any paradigm. In this work, we propose a generalized deepfake detection method, DeFake (Data-Efficient Adaptation for Generalized Deepfake Detection) which can detect both fully synthetic and partially manipulated images simultaneously. We reframe the generalization problem as a data-efficient adaptation of a base synthetic image detector to the task of partial manipulation detection using limited training samples, without degrading the original synthetic image detection task. We introduce three novel modules: (a) Noise-aware Patch Enhancement (NPE) which captures local manipulation artifacts present in partially manipulated images, (b) Adaptive Score Aggregation (ASA) which modulates the influence of the global image-level semantics and the local patch-level artifacts, and (c) Multi-scale alignment which enhances discriminative learning at both image and patch-level. The proposed modules are generalizable and can be integrated into various base models. Extensive experiments on 14 datasets across both paradigms demonstrate the effectiveness of our proposed DeFake, outperforming state-of-the-art approaches in both settings.

Keywords

Deepfake detection, Data-efficient learning, Vision-language models