无需领域知识，用向量几何自动识别模型错误并融合检测结果。
Adversarially Robust Abductive Fusion of Pre-trained Transformer-based Perception Models
- 利用模型自身训练嵌入的向量空间几何生成错误检测规则。
- 在90%标签翻转攻击下仍保持0.42平均F1，优于多数投票方法22%。
- 适合需要高鲁棒性、无先验知识的视觉检测融合场景。
将预训练感知模型部署于新环境时，分布偏移会导致精度下降，仅组合模型无法恢复性能：多数投票类组合器以召回率为代价换取精度，且对协同故障敏感。现有元认知方法虽能学习逻辑规则标记错误，但依赖人工设计的领域知识线索（如物体尺寸先验、分割掩码），难以迁移至真正新颖场景。本文提出无需任何领域知识的元认知层，通过各模型自身训练嵌入构建的标签向量池（LVP），利用检测结果相对于训练原型的几何关系推导出错误检测规则，在测试集上与基于领域知识的规则性能相当，差异不超过0.002 F1。该方法保持神经符号特性，逻辑规则共享统一框架，可在有领域知识时继续补充。我们将多个不完美ViT检测器的融合建模为一致性驱动的溯因问题，采用精确整数规划（IP）和多项式时间启发式求解。在包含15个气象变化测试集与6个ViT检测器的航拍图像基准上，该方法在干净数据上与最强多数投票变体性能相当（差距≤0.005 F1），且在协同标签翻转攻击下表现更优：当翻转率达90%时，平均F1为0.42，远超MV-Plurality的0.35（相对提升22%），并在所有测试集上当翻转率超过0.4后达到最高F1。
原文摘要 · Abstract (English)
Deploying pre-trained perception models in novel environments degrades their accuracy under distributional shift, and assembling them alone does not recover it: combiners such as majority voting trade recall for precision and are brittle to coordinated failures. Prior metacognitive methods learn logical rules that flag a model's errors, but rely on hand-authored domain-knowledge cues (object-size priors, segmentation masks) that do not transfer to genuinely novel scenes. We show that this metacognitive layer can be learned without any domain knowledge by exploiting vector-space geometry: per-model Label Vector Pools (LVP), built from each model's own training embeddings, yield error-detection rules from the geometry of detections relative to training-determined prototypes, reaching parity with domain-knowledge rules to within $0.002$ every F1 on test set. Because the approach remains neurosymbolic, these geometric rules share a single logical framework and can still be complemented by domain knowledge when available. We frame the fusion of multiple imperfect ViT-based detectors as a consistency-based abduction problem solved at test time by an exact Integer Program (IP) and a polynomial-time heuristic. On an aerial-imagery benchmark of 15 weather-shifted test sets and six ViT detectors, our domain-knowledge-free layer matches the strongest majority-vote variant on clean data (within $0.005$ F1) and, unlike every majority-vote baseline, retains its performance under a coordinated label-flipping attack: at a $90\%$ flip rate it averages $0.42$ F1 versus $0.35$ for MV-Plurality (a $22\%$ relative gain) and attains the highest F1 on \emph{every} test set once the flip rate exceeds $0.4$
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。