The widespread misuse of large language models (LLMs) has made machine-generated text detection (MGTD) a critical task. Existing detection methods are often vulnerable to stylistic noise, which limits their generalization in real-world scenarios and robustness on short texts. Furthermore, many methods generate multiple auxiliary texts through LLM-based rewriting or similar strategies, substantially increasing detection latency and limiting their practicality. To address these challenges, we propose the **DisGAR** detection framework, which strips stylistic noise from the input text to mitigate its impact on detection performance, thereby improving cross-domain and cross-model robustness. Specifically, we first fine-tune the attention layers of the proxy model to capture generation-specific patterns. Then we further introduce a global artifact aggregation mechanism and a contrastive learning strategy based on rewritten texts, enabling the separation of stylistic artifacts from the underlying generation patterns. Experimental results demonstrate that DisGAR outperforms existing state-of-the-art methods in unseen-model and unseen-domain detection. Moreover, DisGAR achieves detection efficiency comparable to Fast-DetectGPT, successfully balancing detection performance and efficiency. Source code is available at https://github.com/m-Clown/DisGAR.git

Distributed LLM inference avoids sending raw inputs by transmitting intermediate hidden states, a practice widely assumed to preserve privacy. We challenge this assumption and demonstrate that intermediate representations alone are sufficient to leak sensitive user attributes. This setting poses a fundamental obstacle for existing attribute inference attacks, which typically rely on auxiliary embedding-attribute pairs. To characterize this previously underexplored privacy risk, we reformulate attribute inference as zero-shot matching over candidate attributes directly in the intermediate representation space, and introduce a purely intermediate-representation-based attribute inference attack, termed IR-AIA. To address two structural challenges that hinder attribute inference from intermediate representations, we propose SG-APCR to address layer-dependent anisotropy in intermediate embeddings and a sliding-window similarity matching strategy to handle subword-level semantic fragmentation. Experiments across three LLMs and three real-world datasets show that sensitive attributes can be reliably inferred using only intermediate representations, achieving Top-1 accuracy of up to 0.997 on CMS, 0.980 on Skytrax, and 0.986 on ECHR. These results reveal that intermediate states commonly considered safe to share can expose sensitive personal attributes on their own.