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

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

TL;DR

Researchers have developed a hybrid defense strategy that protects deep learning models for wireless traffic forecasting against sophisticated data poisoning attacks. By decoupling detection and denoising from the forecasting task, their "M3" model retains over 92% of its original accuracy on clean data while outperforming standard adversarial training by more than 2.5x under attack.

What Problem Does This Paper Address?

As we move toward autonomous 5G and 6G networks, deep learning models are increasingly used to predict wireless traffic demand. These forecasts are critical for resource allocation, energy management, and proactive maintenance. However, these models are vulnerable to data poisoning attacks, where an adversary subtly manipulates historical data to trigger network failures or financial losses.

The industry faces a classic "security vs. performance" trade-off. Standard defense methods, like adversarial training, often make models "too pessimistic." While they become harder to fool, their accuracy on normal, clean data typically drops significantly. This paper aims to break that boundary: creating a system that is robust against "worst-case" attacks without degrading its day-to-day forecasting utility.

Methods

The authors propose a modular, hybrid defense framework. Instead of trying to bake robustness directly into the forecaster's weights, they treat security as a pre-processing pipeline.

The Attack: Bi-level Masking PGD

The researchers first designed a more realistic "white-box" attack using Projected Gradient Descent (PGD). Unlike previous studies that perturbed entire data sequences, this paper introduces bi-level masking. This allows an attacker to manipulate specific, individual time steps within a sequence, making the attack stealthier and harder to detect.

The Defense: The M3 Model

The standout contribution is the M3 architecture, which consists of three components trained independently:

  1. The Classifier ($C$): An Inception-based neural network that flags whether an incoming data sequence looks "poisoned."
  2. The Denoiser ($D$): A basic auto-encoder trained to "clean" perturbed samples by removing adversarial noise.
  3. The Forecaster ($F_1$): A standard LSTM-based model trained only on clean data to ensure maximum accuracy.

When data arrives, the classifier decides its path. If flagged as suspicious, the data is sent through the denoiser before reaching the forecaster. If it looks clean, it goes straight to the forecaster.

Datasets

The study utilizes the Telecom Italia dataset, specifically focusing on call volumes in Milan’s urban environment.

  • Scope: Data from 100 base stations over an 8-week period.
  • Structure: Univariate time series divided into sub-sequences of length 3 (representing lags of 1, 2, and 24 hours).
  • Challenge: The dataset has a high standard deviation (0.26 after normalization), making it difficult to distinguish between natural traffic spikes and malicious perturbations.

Key Insights and Ideas

  • Decoupling is Key: The most significant "aha moment" is that separating detection from forecasting works better than end-to-end adversarial training. By keeping the forecaster "clean," the model avoids the performance degradation typical of robust models.
  • Parallel Efficiency: Because the classifier, denoiser, and forecaster are independent modules, they can be trained in parallel. This offers a significant computational advantage over complex bi-level optimization training schemes.
  • The "Shift" Blind Spot: The researchers found that if an attacker shifts an entire sequence ($k=3$), the classifier struggles because the relative patterns remain the same. However, for more common "point" perturbations, the classifier is highly effective.
  • Breaking the Trade-off: The results suggest that the "robustness-accuracy" trade-off isn't an unbreakable law; rather, it’s a limitation of current training paradigms that can be bypassed with smart architectural choices.

Key Results

  • Superior Accuracy: The M3 model retained 92.02% of the original model's performance on clean data.
  • Massive Robustness Gains: Compared to standard adversarial training (the $M_2$ model), the M3 model achieved a Mean Squared Error (MSE) that was 2.71x lower on clean data and 2.51x lower on perturbed data.
  • Detection Capability: The classifier maintained high accuracy (often 80%+) across various perturbation levels ($\ell_\infty$-norm from 0.1 to 0.4), effectively acting as a gatekeeper for the system.

Limitations and Future Work

Despite the success, the authors acknowledge a few hurdles:

  • False Positives: The classifier has a relatively high false-positive rate (~39%) on normal data due to the high natural variance of telecom traffic. While the denoiser mitigates the impact of these "false alarms," improving classifier precision is a priority.
  • The $k=3$ Scenario: As mentioned, when every step in a sequence is perturbed identically, the current classifier's performance dips. Future work will look into more sophisticated anomaly detection that considers broader temporal contexts.
  • Generalization: The current study focuses on univariate data. Extending this hybrid approach to multivariate forecasting (where multiple base stations interact) remains an open challenge.

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× and 2.51× 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},
}

Loading related content...

Figure