PINSIGHT: A Comprehensive Threat Exploration of Domain-Adaptive Wi-Fi based PIN Code Inference

TL;DR

This paper presents PINSIGHT, a methodology for assessing Wi-Fi-based PIN inference attacks, revealing their limited generalization across different typing conditions and challenging previous high-performance claims.

Contribution

Introducing PINSIGHT, a novel threat assessment framework that isolates environmental and typing effects, providing a rigorous evaluation of attack generalization capabilities.

Findings

Attacks reliably generalize across environmental changes.

Performance degrades when channel encoding of typing shifts.

Current state-of-the-art attacks overestimate real-world threat.

Abstract

Wi-Fi signals can be exploited by adversaries as a sensing side channel to eavesdrop on physical information. By monitoring propagation effects of radio waves within the victim's environment, attackers can remotely infer sensitive information. One particularly concerning example is PIN code inference, where the attacker faces the challenge of mapping Wi-Fi physical-layer channel estimations back into typed digits. While effective in their training environment, such attacks typically fail as soon as they are deployed in unseen environments. The current state-of-the-art attack, WiKI-Eve, attempts to overcome this problem using a deep-learning approach, reporting high PIN code inference accuracy independent of environments, devices, and users. While this suggests a significant real-world threat, it is not well understood how far the attack actually reaches, nor what its underlying generalization performance is based on. In this work, we close this gap by presenting PINSIGHT, a novel methodology that separates the effects of environmental variation and PIN code typing. This enables the first rigorous threat assessment of such attacks, evaluating their generalization capabilities and limitations. Our approach leverages a robotic typing platform that produces highly repeatable keystroke events across systematically varied environment changes [...]. This dataset constitutes the first benchmark for environment generalization in Wi-Fi PIN code inference attacks. Evaluating several state-of-the-art methods, we find that attacks generalize reliably across changes in the surrounding environment but degrade substantially when the channel's encoding of typing itself shifts - precisely the condition that defines a realistic attack scenario. We conclude that the reported performance of current state-of-the-art Wi-Fi PIN inference attacks is not representative of the actual real-world threat.