SENTINEL: Securing Indoor Localization against Adversarial Attacks with Capsule Neural Networks

Overview

• This paper presents a novel embedded machine learning framework called SENTINEL to address challenges in indoor localization using Wi-Fi received signal strength (RSS) fingerprinting.
• Key challenges include RSS fluctuations due to dynamic indoor environments, device heterogeneity, and adversarial attacks from malicious or rogue access points (APs).
• SENTINEL uses modified capsule neural networks to improve resilience against these issues, outperforming state-of-the-art approaches.
• The paper also introduces a new dataset called RSSRogueLoc to capture the effects of rogue APs in real-world indoor environments.

Plain English Explanation

Using Wi-Fi signal strength to determine a device's location indoors has become popular as GPS does not work well inside buildings. However, this approach faces several problems. The Wi-Fi signal can fluctuate a lot due to changes in the indoor environment, and different devices may interpret the signal differently. Additionally, attackers could set up fake Wi-Fi access points to deliberately confuse the location tracking system.

To address these challenges, the researchers developed a new machine learning framework called SENTINEL. It uses a modified type of neural network called capsule networks to make the location tracking more robust against signal fluctuations, device differences, and adversarial attacks. Experiments showed SENTINEL can reduce the location error by up to 3.5 times compared to other state-of-the-art approaches, both in simulated attacks and when tested with a new real-world dataset the researchers created.

This research is significant as it helps enable more reliable and secure indoor location-based services, which have many applications, like guiding people in large buildings or tracking assets. The new dataset the researchers built could also be useful for further work in this area.

Technical Explanation

The paper presents SENTINEL, a novel embedded machine learning framework for indoor localization using Wi-Fi RSS fingerprinting. SENTINEL utilizes modified capsule neural networks to enhance resilience against adversarial attacks, device heterogeneity, and dynamic RSS fluctuations.

The researchers identified key challenges in this domain, including:

• RSS fluctuations due to dynamic changes in indoor environments
• Heterogeneity of edge devices leading to diminished localization accuracy
• Emerging threats from adversarial attacks, especially those introduced by malicious or rogue APs

To address these issues, SENTINEL employs a modified capsule network architecture that can better capture the intricate spatial relationships in RSS fingerprints and improve robustness. The framework also incorporates techniques to mitigate the impacts of device heterogeneity and adversarial attacks.

Additionally, the researchers introduced a novel dataset called RSSRogueLoc, which captures the effects of rogue APs in real-world indoor environments. This dataset complements existing benchmarks and can support further research in this area.

Experimental evaluations demonstrate that SENTINEL achieves significant improvements over state-of-the-art frameworks. It shows up to a 3.5x reduction in mean error and 3.4x reduction in worst-case error under simulated adversarial attacks. SENTINEL also delivers up to 2.8x improvements in mean error and 2.7x in worst-case error when evaluated with the real-world RSSRogueLoc dataset.

Critical Analysis

The paper provides a comprehensive approach to address critical challenges in indoor localization using Wi-Fi RSS fingerprinting. The introduction of the RSSRogueLoc dataset is a valuable contribution, as it enables evaluating the resilience of localization systems against adversarial attacks in real-world environments.

However, the paper does not extensively discuss the limitations of the SENTINEL framework. For instance, it would be important to understand how the performance of SENTINEL scales with the number of access points, the density of the training data, and the complexity of the indoor environments. Additionally, the paper does not explore the trade-offs between the increased computational complexity of the capsule network architecture and the improved localization accuracy.

Further research could also investigate the generalizability of SENTINEL to different indoor environments and its ability to adapt to changing conditions over time. Federated learning approaches could be explored to enhance the adaptability and personalization of the localization system.

Conclusion

This paper presents SENTINEL, a novel embedded machine learning framework that leverages modified capsule neural networks to improve the resilience of indoor localization systems against adversarial attacks, device heterogeneity, and dynamic RSS fluctuations. The introduced RSSRogueLoc dataset also contributes to advancing research in this domain.

The experimental results demonstrate the effectiveness of SENTINEL, with significant improvements in localization accuracy compared to state-of-the-art approaches. This research is important for enabling more reliable and secure indoor location-based services, which have numerous applications in areas such as asset tracking, navigation, and emergency response.

While the paper shows promising results, further exploration of the limitations and potential extensions of the SENTINEL framework could lead to even more robust and adaptive indoor localization solutions that can withstand the challenges of dynamic real-world environments.