[Paper Review] Explanation of Unintended Radiated Emission Classification via LIME

Tom Grimes, E. Church+2 more|arXiv (Cornell University)|Sep 4, 2020

Integrated Circuits and Semiconductor Failure Analysis7 references4 citations

TL;DR

This paper proposes a LIME-based explanation framework to interpret deep learning models for classifying unintended radiated emissions (URE) from consumer electronics. By applying LIME to a ResNet-18 classifier trained on short-time Fourier transforms of voltage signatures from the Flaming Moes dataset, the authors identify key frequency bands used for device classification, achieving near 100% accuracy and enhancing model trustworthiness and transferability through robust feature extraction via ensemble learning.

ABSTRACT

Unintended radiated emissions arise during the use of electronic devices. Identifying and mitigating the effects of these emissions is a key element of modern power engineering and associated control systems. Signal processing of the electrical system can identify the sources of these emissions. A dataset known as Flaming Moes includes captured unintended radiated emissions from consumer electronics. This dataset was analyzed to construct next-generation methods for device identification. To this end, a neural network based on applying the ResNet-18 image classification architecture to the short time Fourier transforms of short segments of voltage signatures was constructed. Using this classifier, the 18 device classes and background class were identified with close to 100 percent accuracy. By applying LIME to this classifier and aggregating the results over many classifications for the same device, it was possible to determine the frequency bands used by the classifier to make decisions. Using ensembles of classifiers trained on very similar datasets from the same parent data distribution, it was possible to recover robust sets of features of device output useful for identification. The additional understanding provided by the application of LIME enhances the trainability, trustability, and transferability of URE analysis networks.

Motivation & Objective

• To improve interpretability of deep learning models in unintended radiated emission (URE) classification for power engineering applications.
• To identify the specific frequency bands that a neural network uses to classify different consumer electronic devices.
• To enhance model trustworthiness, trainability, and transferability through explainable AI techniques like LIME.
• To develop robust feature sets for device identification using ensembles of classifiers trained on similar data from the same distribution.

Proposed method

• A ResNet-18 architecture is adapted for image classification tasks using short-time Fourier transforms (STFTs) of voltage signal segments from the Flaming Moes dataset.
• The model is trained to classify 18 device classes and a background class with high accuracy using STFT spectrograms as input.
• LIME (Local Interpretable Model-agnostic Explanations) is applied to individual predictions to identify salient frequency bands influencing classification decisions.
• LIME explanations are aggregated across multiple predictions for the same device to extract consistent, robust frequency features.
• Ensemble learning is employed using multiple classifiers trained on slightly varied subsets of the same data distribution to improve feature stability and reliability.
• The resulting frequency band features are analyzed to determine their relevance and consistency across models and devices.

Experimental results

Research questions

• RQ1 Which frequency bands are most influential in the deep learning model’s classification of unintended radiated emissions?
• RQ2 How can LIME be effectively applied to interpret complex neural network decisions in URE classification?
• RQ3 Can ensemble methods improve the robustness and consistency of interpretable features extracted from URE data?
• RQ4 To what extent does explainability via LIME enhance the trustworthiness and transferability of URE classification models?

Key findings

• The ResNet-18 model achieved near 100% accuracy in classifying 18 device classes and a background class using STFT representations of voltage signals.
• LIME analysis successfully identified specific frequency bands that the model relied on for classification decisions, providing interpretability.
• Aggregating LIME explanations across multiple predictions for the same device revealed consistent and robust frequency features for device identification.
• Ensemble training on similar data subsets improved the stability and reliability of the extracted frequency features.
• The integration of LIME enhanced model trustworthiness by making decision-making processes transparent and interpretable.
• The approach demonstrated strong potential for improving transferability of URE analysis models across different device types and signal conditions.