Breaking Boundaries: Balancing Performance and Robustness in Deep Wireless Traffic Forecasting

Romain Ilbert, Thai V. Hoang, Zonghua Zhang, Themis Palpanas

Introduction

As telecommunications infrastructure transitions toward autonomous 5G and 6G architectures, deep learning (DL) models have become indispensable for wireless traffic forecasting. These models enable proactive resource allocation, energy management, and network maintenance by predicting fluctuations in demand. However, the integration of DL into critical infrastructure introduces significant security risks, most notably data poisoning attacks. In these scenarios, adversaries manipulate historical traffic data to induce forecasting errors, potentially leading to service outages or inefficient resource utilization.

Historically, defending these models has involved a "security vs. performance" trade-off. Robustness techniques like adversarial training often degrade a model's accuracy on clean, non-malicious data—a phenomenon known as the "robustness price." This paper introduces the M3 framework, a hybrid defense strategy designed to secure wireless traffic forecasting without sacrificing operational performance.

The Vulnerability: Bi-level Masking PGD

To evaluate the robustness of forecasting models, the researchers developed a sophisticated "white-box" attack mechanism: Bi-level Masking Projected Gradient Descent (PGD). Unlike traditional adversarial attacks that apply perturbations across an entire data sequence, this method allows an attacker to target specific, individual time steps. This selective manipulation makes the attack significantly stealthier, as the overall statistical properties of the sequence may remain within plausible bounds while still triggering incorrect forecasts.

The M3 Framework: A Modular Defense

The core contribution of this work is the M3 architecture, which decouples the security functions from the primary forecasting task. Rather than training a single robust model, the framework utilizes a three-component pipeline:

  1. The Classifier ($C$): An Inception-based neural network trained to distinguish between clean and poisoned data sequences.
  2. The Denoiser ($D$): A basic auto-encoder designed to reconstruct clean data from perturbed inputs, effectively stripping away adversarial noise.
  3. The Forecaster ($F_1$): A standard Long Short-Term Memory (LSTM) model trained exclusively on clean data to ensure peak predictive performance.

In this hybrid workflow, incoming data is first processed by the classifier. If the data is flagged as suspicious, it is routed through the denoiser before reaching the forecaster. If the data is deemed clean, it bypasses the denoiser entirely. This modularity allows the forecaster to remain optimized for standard traffic patterns, avoiding the "pessimism" inherent in models trained on adversarial examples.

Experimental Evaluation

The framework was tested using the Telecom Italia dataset, focusing on call volumes from 100 base stations in Milan over an eight-week period. The data was structured as univariate time series with sub-sequences representing specific temporal lags (1, 2, and 24 hours). This dataset presents a significant challenge due to its high natural variance (standard deviation of 0.26 after normalization), which can often mask malicious perturbations.

Performance and Robustness

The M3 model demonstrated a remarkable ability to break the traditional trade-off:

  • Accuracy Retention: The system retained 92.02% of the original model's performance on clean data, significantly outperforming traditional defense methods.
  • Robustness Gains: When compared to standard adversarial training (referred to as the $M_2$ model), the M3 framework achieved a Mean Squared Error (MSE) that was 2.71x lower on clean data and 2.51x lower under attack.
  • Detection Efficacy: The Inception-based classifier maintained high accuracy (typically above 80%) across various perturbation intensities ($\ell_\infty$-norm ranging from 0.1 to 0.4).

Key Insights

The success of the M3 framework highlights several critical takeaways for secure network AI:

  • Decoupling as a Paradigm: Separating detection and denoising from forecasting prevents the "robustness price" from affecting day-to-day operations.
  • Computational Efficiency: Because the three modules are independent, they can be trained in parallel, offering a faster and more scalable alternative to complex bi-level optimization schemes.
  • The "Shift" Blind Spot: The researchers identified a specific vulnerability where attackers perturb an entire sequence ($k=3$). Because the relative patterns remain consistent, the classifier struggles to detect the shift, indicating a need for broader temporal context in future iterations.

Conclusion and Future Directions

The M3 framework proves that high-performance wireless traffic forecasting can be made robust against sophisticated data poisoning. While the current iteration faces challenges with false positives due to the high variance of telecom data, the modular approach provides a clear path forward. Future research will focus on improving classifier precision, addressing the $k=3$ sequence-shift vulnerability, and extending the framework to multivariate forecasting scenarios where interactions between multiple base stations are considered.

Summary updated

Abstract

Balancing the trade-off between accuracy and robustness is a long-standing challenge in time series forecasting. While most of existing robust algorithms have achieved certain suboptimal performance on clean data, sustaining the same performance level in the presence of data perturbations remains extremely hard. In this paper, we study a wide array of perturbation scenarios and propose novel defense mechanisms against adversarial attacks using real-world telecom data. We compare our strategy against two existing adversarial training algorithms under a range of maximal allowed perturbations, defined using $\ell_{\infty}$-norm, $\in [0.1,0.4]$. Our findings reveal that our hybrid strategy, which is composed of a classifier to detect adversarial examples, a denoiser to eliminate noise from the perturbed data samples, and a standard forecaster, achieves the best performance on both clean and perturbed data. Our optimal model can retain up to $92.02\%$ the performance of the original forecasting model in terms of Mean Squared Error (MSE) on clean data, while being more robust than the standard adversarially trained models on perturbed data. Its MSE is 2.71$\times$ and 2.51$\times$ lower than those of comparing methods on normal and perturbed data, respectively. In addition, the components of our models can be trained in parallel, resulting in better computational efficiency. Our results indicate that we can optimally balance the trade-off between the performance and robustness of forecasting models by improving the classifier and denoiser, even in the presence of sophisticated and destructive poisoning attacks.
BibTeX
@inproceedings{ilbert2023breaking,
  author = {Romain Ilbert and Thai V. Hoang and Zonghua Zhang and Themis Palpanas},
  title = {{Breaking Boundaries: Balancing Performance and Robustness in Deep Wireless Traffic Forecasting}},
  booktitle = {ARTMAN@CCS},
  year = {2023},
  url = {https://openreview.net/forum?id=fVTvTV78NP},
}
Figure