EVENT DOSSIER
On September 7, 2026, the TreeFI paper was published on arXiv cs.AI. This method aims to enhance the security of deep neural networks through statistical fault injection techniques. Unlike traditional methods, TreeFI introduces a “value perception” mechanism that can dynamically adjust the strategy and intensity of fault injection based on the importance of different neurons or layers in the model to the final output results (i.e., their contribution values). This approach attempts to more accurately identify potential vulnerabilities that significantly affect model decisions while maintaining test efficiency, thereby enabling more comprehensive and effective security assessment of deep learning systems.
TreeFI 提出了一种针对 DNN 激活值和权重的 FP32 单比特故障注入方法。该方法利用回归树将层值分布划分为具有相似位翻转行为的区间，并依据其对故障率估计的相关性分配注入量。在 CIFAR-10 和 ImageNet 数据集上的 CNN 与 Transformer 模型验证中，TreeFI 在相同实验设置下比现有统计基准提供了更准确的估计结果。测试表明，TreeFI 将所需的注入预算减少了最高 72.1 倍，其中激活故障的平均减少率为 44.9 倍，已执行的权重故障平均减少率为 11.2 倍。