On the Effects of Online Black-Box Attacks against Mobile Network Capacity Forecasters under Temporal Constraints

Abstract

Machine Learning (ML) methods are increasingly used to drive resource allocation in mobile networks; however, they are vulnerable to adversarial manipulation. In this work, we investigate attacks on network capacity forecasting that respect temporal constraints inherent to online time series attacks: input sequences overlap, the past is immutable, and future traffic values are unknown. Under these constraints, we propose an attack where an adversary estimates future traffic load, queries the target model, and applies a perturbation that maximizes an evaluation function representing the attacker’s objective. We use real-world wireless traffic datasets to validate our method and target well-known ML models for capacity forecasting. Our experiments show that attack effectiveness increases with the adversary’s accuracy in forecasting future traffic values and that perturbations impact a sequence of predictions, producing an approximately 2.38-fold relative error increase on targeted forecasts under high forecasting accuracy, with values of the employed capacity forecasting function increasing from 0.13 to 0.31 for perturbations equivalent to 10% of the traffic peak.