Abhishek K. Mishra | Searcharxiv

arXiv2026

StateFi: Effectively Identifying Wi-Fi Devices through State Transitions

Randomized MAC addresses aim to prevent passive device tracking, yet Wi-Fi management frames still leak structured behavioral patterns. Prior work has relied primarily on syntactic probe-request features such as Information Elements (IEs), sequence numbers (SEQ), or RSSI correlations, which degrade in dense environments and fail under aggressive randomization. We introduce StateFi, a fingerprinting framework that models device behavior as finite-state machines (FSMs), capturing both structural transition patterns and temporal execution logic. These FSMs are embedded into compact feature vectors that support efficient similarity computation and supervised classification. Across five heterogeneous campus environments, StateFi achieves 94-97% accuracy for in-network fingerprinting using full management-frame FSMs. With probe-only FSMs, it re-identifies devices under MAC randomization with up to 97% accuracy across large public datasets comprising more than a million frames. When looking at the discrimination accuracy of the model, StateFi reaches 98%, outperforming the strongest prior signature by up to 17 percentage points. These results demonstrate that FSM-level behavioral dynamics form a powerful and largely unmitigated side channel, stable enough to defeat randomization and expressive enough for robust, scalable device identification.

cs.CR↗

arXiv2025

The Model's Language Matters: A Comparative Privacy Analysis of LLMs

Large Language Models (LLMs) are increasingly deployed across multilingual applications that handle sensitive data, yet their scale and linguistic variability introduce major privacy risks. Mostly evaluated for English, this paper investigates how language structure affects privacy leakage in LLMs trained on English, Spanish, French, and Italian medical corpora. We quantify six linguistic indicators and evaluate three attack vectors: extraction, counterfactual memorization, and membership inference. Results show that privacy vulnerability scales with linguistic redundancy and tokenization granularity: Italian exhibits the strongest leakage, while English shows higher membership separability. In contrast, French and Spanish display greater resilience due to higher morphological complexity. Overall, our findings provide the first quantitative evidence that language matters in privacy leakage, underscoring the need for language-aware privacy-preserving mechanisms in LLM deployments.

cs.CL↗