TRAIL: A Three-Step Robust Adversarial Indoor Localization Framework

Indoor localization utilizing received signal strength (RSS) fingerprint has garnered significant attention over the past decade because it is readily captured from the MAC layer of ubiquitous hardware devices. However, the localization accuracy of RSS fingerprint-based methods is notably influenced by two primary factors: 1) disparities between offline and online data distributions induced by dynamic environmental changes and device heterogeneity and 2) inconsistencies among hetero-measure samples (different RSS samples collected at the same reference point (RP) during the online stage) stemming from unknown noise and interference. To address these issues, we propose a three-step robust adversarial indoor localization (TRAIL) framework. The model is pretrained in the first step (Step A), and an adversarial game is played between a regressor and a feature extractor within the model in the second step (Step B) and third step (Step C). Specifically, Step B trains the regressor to discover more 'hard' samples, i.e., hetero-measure samples with notable positioning differences, and Step C trains the feature extractor to learn a suitable transformation that eliminates the disparities between offline and online data distributions and the 'hard' samples. To harmonize the contributions of the two factors in model training, we integrate the multiple gradient descent algorithm (MGDA). Experimental results on both actual and simulated datasets demonstrate that TRAIL outperforms state-of-the-art methods and exhibits robustness in low signal-to-noise ratio (SNR) environments.