The SaTML '24 CNN Interpretability Competition: New Innovations for Concept-Level Interpretability

可解释性技术对于帮助人类理解和监督人工智能系统非常有价值。SaTML 2024 CNN可解释性竞赛征集了研究ImageNet规模下卷积神经网络（CNN）的新方法。比赛的目标是帮助人类众包工作者识别CNN中的木马。本报告展示了四个参赛作品的方法和结果。通过可解释性工具帮助人类可靠地诊断木马仍然具有挑战性。然而，这次比赛的参赛作品提供了新的技术，并在Casper等人于2023年发布的基准测试中创造了新记录。

SaTML 2024 CNN Interpretability Competition: Studying CNNs for Trojan Detection

The competition aims to develop novel interpretability techniques for CNNs to help human crowd-workers identify trojans in large-scale image classification tasks. The competition entries have contributed new techniques and achieved a new record on the benchmark from Casper et al., 2023.

The competition showcases four featured entries that propose different interpretability techniques for CNNs, including activation-based methods, gradient-based methods, and adversarial training. The entries are evaluated on a large-scale image classification dataset and the performance is measured by accuracy and AUC. The competition results demonstrate the potential of interpretability techniques for trojan detection in CNNs.

Related work includes previous research on interpretability techniques for CNNs, such as saliency maps, occlusion analysis, and feature visualization. There are also studies on trojan attacks and defenses in deep learning, such as BadNets, STRIP, and Fine-Pruning.