DDNet: A Dual-Stream Graph Learning and Disentanglement Framework for Temporal Forgery Localization

TL;DR

DDNet addresses temporal forgery localization by combining local and global reasoning through a Dual-Stream Graph Learning framework, augmented by Cross-Level Feature Embedding and Trace Disentanglement and Adaptation. CLFE fuses semantic priors from CLIP with texture cues from ResNet; DSGL uses a Local-distance stream and a Dynamic Semantic stream to capture short- and long-range cues, while TDA isolates generic forgery fingerprints for better cross-domain generalization. The approach achieves state-of-the-art results on ForgeryNet and TVIL, notably about a 9% improvement in AP@0.95 and enhanced cross-domain robustness, validating both the local-global integration and the domain-robust disentanglement strategy. This work advances video forensics by providing precise, interpretable localization of forged segments with strong generalization to unseen domains.

Abstract

The rapid evolution of AIGC technology enables misleading viewers by tampering mere small segments within a video, rendering video-level detection inaccurate and unpersuasive. Consequently, temporal forgery localization (TFL), which aims to precisely pinpoint tampered segments, becomes critical. However, existing methods are often constrained by \emph{local view}, failing to capture global anomalies. To address this, we propose a \underline{d}ual-stream graph learning and \underline{d}isentanglement framework for temporal forgery localization (DDNet). By coordinating a \emph{Temporal Distance Stream} for local artifacts and a \emph{Semantic Content Stream} for long-range connections, DDNet prevents global cues from being drowned out by local smoothness. Furthermore, we introduce Trace Disentanglement and Adaptation (TDA) to isolate generic forgery fingerprints, alongside Cross-Level Feature Embedding (CLFE) to construct a robust feature foundation via deep fusion of hierarchical features. Experiments on ForgeryNet and TVIL benchmarks demonstrate that our method outperforms state-of-the-art approaches by approximately 9\% in AP@0.95, with significant improvements in cross-domain robustness.

Figures (6)

• Figure 1: Illustration of research background and motivation. (a) Comparison between forgery detection and temporal forgery localization (TFL). Forgery detection outputs a single video-level label (Real/Fake), often overlooking brief forgery segments. In contrast, TFL aims to precisely determine the location of forged segments. (b) Comparison between existing TFL methods and proposed DDNet. Existing TFL methods are often constrained by local view, failing to capture long-range dependencies between distant nodes. Our proposed DDNet overcomes this by introducing Dual-Stream Graph Learning (DSGL). By coordinating the Temporal Distance Stream and Semantic Content Stream, we integrate local inconsistency with global cues, thereby achieving more precise localization.
• Figure 2: Overview of the proposed DDNet framework. The framework comprises three main components: (a) Cross-Level Feature Embedding (CLFE): This module extracts visual features using frozen CLIP and ResNet encoders, and subsequently fuses semantic priors with textural details via a bidirectional cross-attention mechanism to bridge the feature gap. (b) Dual-Stream Graph Learning (DSGL): The core module designed to overcome the local view limitation. It integrates a Temporal Distance Stream to capture local temporal inconsistency and a Semantic Content Stream to perform global semantic reasoning across disjoint manipulated segments. (c) Trace Disentanglement and Adaptation (TDA): An auxiliary training module that imposes adversarial and orthogonal constraints to isolate generic forgery fingerprints from domain-specific noise. Finally, the learned representations are projected to generate frame-level manipulation probabilities.
• Figure 3: Illustration of the proposed Dual-Stream Graph Learning (DSGL) module.
• Figure 4: Qualitative visualization. The Ground Truth (green) shows a continuous forgery. The ablated variants suffer from severe prediction fragmentation. Our full model (red) can maintain consistency, avoiding the gaps seen in ablated variants.
• Figure 5: Distribution of Manipulation Methods in the Training Set of the Standard Subset. The distribution is highly imbalanced, with Method 6 accounting for a significant portion of the data, while other methods are underrepresented.