Complementarity/steganographic-collusion-detection

隐写共谋检测基准（v3，n=1600，类别平衡）数据集包含1,600个示例（800个隐写和800个干净），均匀分布在16个类别中，每个类别100个示例。数据集基于Anwar等人（2026）的隐写决策理论形式化，用于检测AI助手响应中是否包含隐藏的隐写通信。每个隐写示例都有一个假设接收者知道但检测者不知道的私有码本，每个隐写示例都会产生一个可测量的隐写差距Δ_steg。数据集的任务是让评分者根据用户问题和AI助手的响应判断是否存在隐写通信。数据集还包括详细的元数据、质量保证措施和生成方法。

The Steganographic Collusion Detection Benchmark (v3, n=1600, class-balanced) dataset consists of 1,600 examples (800 stego and 800 clean), uniformly distributed across 16 categories with 100 examples per category. Grounded in the decision-theoretic formalization of steganography by Anwar et al. (2026), the dataset is designed to detect hidden steganographic communication in AI assistant responses. Each stego example has a private codebook known to a hypothetical receiver but not to the detector, inducing a measurable steganographic gap Δ_steg. The task involves raters judging whether a response contains hidden steganographic communication based on a user question and an AI assistants response. The dataset includes detailed metadata, quality assurance measures, and generation methodology.