Abstract: To address the quadratic computational complexity of self-attention architectures and the neglect of the "audio-driven vision" generative causal logic under symmetric equal-weight interaction in existing audio-visual temporal forgery localization, a localization method centered on a Cross-Modal Selective State Space Model (Cross-Mamba) was proposed. Firstly, an asymmetric coupling mechanism was designed at the state-dynamics level, where auditory priors dynamically modulated the time-varying step size of visual features, and boundary-aware gating achieved cross-modal alignment and mismatched-state reset. Secondly, a Dual-Stage Relative Boundary Refinement (DS-RBR) head was constructed, transforming traditional coarse-grained anchor regression into continuous sub-interval offset refinement to handle the boundary uncertainty of extremely short tampered segments. At the strictest temporal Intersection over Union (tIoU) threshold of 0.95, the proposed method achieved an Average Precision (AP) of 49.2% on the LAV-DF benchmark, outperforming the previous best model AuViRe (46.5%) by 2.7 percentage points, and 14.3% on the million-scale open benchmark AV-Deepfake1M, exceeding AuViRe (11.7%) by 2.6 percentage points. Experimental results show that the method overcomes the computational bottleneck of long-video modeling with linear complexity, mitigates the quantization error of extremely short tampered segments, and maintains stable boundary localization performance in highly degraded and open scenarios.

Key words: deepfake detection, Temporal Forgery Localization (TFL), State Space Model (SSM), multimodal fusion, boundary regression refinement

摘要： 针对现有视听时序伪造定位方法依赖自注意力架构导致二次计算复杂度、且平权对称交互忽略“音频驱动视觉”因果逻辑的问题，提出了一种以跨模态选择性状态空间模型(Cross-Modal Selective State Space Model，Cross-Mamba)为核心的时序伪造定位方法。首先，在状态动力学层面设计非对称耦合机制，利用听觉先验动态调制视觉特征的时变步长，并结合边界感知门控实现跨模态对齐与失配状态重置；其次，构建双阶段相对边界精化(Dual-Stage Relative Boundary Refinement，DS-RBR)头，将传统粗粒度锚点回归转化为亚区间连续时序偏移的精化过程，以应对极短篡改片段的边界不确定性。在LAV-DF基准数据集上，本文方法在时序交并比阈值为0.95的最严格对齐条件下取得49.2%的平均精度(Average Precision，AP)，较现有最优模型AuViRe(46.5%)提升2.7个百分点；在百万级开放基准AV-Deepfake1M上的同等阈值条件下取得14.3%的AP，较AuViRe(11.7%)提升2.6个百分点。实验结果表明，本文方法以线性复杂度克服了长视频序列建模的计算瓶颈，并依托非对称动力学调制与连续边界精化机制克服了极短篡改片段的量化误差，在高度退化与开放场景下仍保持了稳定的边界定位性能。

关键词: 深度伪造检测, 时序伪造定位, 状态空间模型, 多模态融合, 边界回归精化