DeExp: Revealing Model Vulnerabilities for Spatio-Temporal Mobile Traffic Forecasting With Explainable AI

The ability to perform mobile traffic forecasting effectively with Deep Neural Networks (DNN) is instrumental to optimize resource management in 5G and beyond generation mobile networks. However, despite their capabilities, these Deep Neural Networks (DNN)s often act as complex opaque-boxes with decisions that are difficult to interpret. Even worse, they have proven vulnerable to adversarial attacks which undermine their applicability in production networks. Unfortunately, although existing state-of-the-art EXplainable Artificial Intelligence (XAI) techniques are often demonstrated in computer vision and Natural Language Processing (NLP), they may not fully address the unique challenges posed by spatio-temporal time-series forecasting models. To address these challenges, we introduce DeExp in this paper, a tool that flexibly builds upon legacy EXplainable Artificial Intelligence (XAI) techniques to synthesize compact explanations by making it possible to understand which Base Stations (BSs) are more influential for forecasting from a spatio-temporal perspective. Armed with such knowledge, we run state-of-the-art Adversarial Machine Learning (AML) techniques on those BSs to measure the accuracy degradation of the predictors under adversarial attacks. Our comprehensive evaluation uses real-world mobile traffic datasets and demonstrates that legacy XAI techniques spot different types of vulnerabilities. While Gradient-weighted Class Activation Mapping (GC) is suitable to spot BSs sensitive to moderate/low traffic injection, LayeR-wise backPropagation (LRP) is suitable to identify BSs sensitive to high traffic injection. Under moderate adversarial attacks, the prediction error of the BSs identified as vulnerable can increase by more than 250%.