Abstract
Deep learning models for multivariate time-series forecasting (MTSF) are increasingly deployed in critical infrastructure, from energy grid management to financial market prediction. However, their vulnerability to adversarial manipulation remains under-explored, particularly in the multivariate context. This paper identifies a novel threat model: the Indirect Sparse Attack, where an adversary disrupts a target forecast by perturbing a small subset of related time series. To address this, we introduce a differentiable "sparse layer" for generating sophisticated attacks and propose two defense mechanisms: a Mini-max adversarial training framework and the first application of Randomized Smoothing to provide robust certificates for multivariate forecasting.
Introduction
While adversarial robustness has been extensively studied in computer vision, research in time-series forecasting has largely focused on univariate scenarios where an attacker modifies the same sequence they intend to disrupt. In real-world multivariate systems, such direct tampering is often easily detected by anomaly detection systems. This research exposes a more insidious vulnerability: the interconnectedness of multivariate models allows for "indirect" attacks. By making imperceptible changes to correlated "neighbor" series, an attacker can exploit the model's learned dependencies to cause significant errors in a target forecast.
Methodology
1. Indirect Sparse Attacks
The authors formalize the Indirect Sparse Attack, where the goal is to maximize the forecasting error of a target series by perturbing at most $k$ other series within a budget $\epsilon$. Two primary attack vectors are developed:
- Deterministic Attack: Utilizes Projected Gradient Descent (PGD) to optimize perturbations, followed by an analytical selection of the most impactful series to target.
- Probabilistic Attack (The Sparse Layer): To overcome the non-differentiable nature of selecting discrete series to attack, the authors developed a novel "sparse layer." By combining a Normal distribution with a Dirac density (a spike at zero), the attack selection becomes a continuous optimization problem. This allows the adversary to "learn" the optimal sparsity pattern through gradient descent, resulting in more effective attacks at low sparsity levels.
2. Defense Mechanisms
To counter these threats, the paper proposes two distinct strategies:
- Mini-max Defense: This approach employs adversarial training where the model is trained to minimize the maximum error produced by the probabilistic attacker. This "game-theoretic" training forces the model to become resilient to the worst-case perturbations it might encounter.
- Randomized Smoothing (RS): Adapted from image classification, RS involves adding random noise to the input during inference and averaging the resulting forecasts. The authors provide a formal proof (Theorem 4.1) that establishes a "robust certificate," offering a mathematical guarantee that the forecast will remain within a specific bound given a defined noise radius.
Experimental Evaluation
Datasets and Metrics
The framework was evaluated on four diverse multivariate datasets: Traffic (SF car lanes), Electricity (customer consumption), Taxi (NYC trip data), and Wiki (Wikipedia page views). Performance was measured using weighted Quantile Loss (wQL), where lower values indicate higher accuracy.
Key Results
- Attack Potency: Indirect attacks proved highly effective. On the Electricity dataset, perturbing just 7 out of 9 available series increased the wQL from a baseline of 0.29 to 0.99. The study also demonstrated that univariate-specific attacks are roughly 3x less effective than multivariate-aware attacks, highlighting the need for specialized research in this domain.
- Defense Efficacy: The Mini-max defense significantly outperformed standard data augmentation. In high-sparsity attack scenarios on the Traffic dataset, the Mini-max defense maintained a wQL of 0.19, whereas the standard model's error escalated to 0.46.
- Stealth and Sparsity: The results indicate that perturbing even a single neighboring series can increase target forecast error by 20%, demonstrating that significant disruption can be achieved with minimal, hard-to-detect changes.
Conclusion and Future Work
This work establishes a foundation for understanding and defending against adversarial threats in multivariate time-series forecasting. By introducing the concept of indirect sparse attacks and providing both empirical defenses (Mini-max) and theoretical guarantees (Randomized Smoothing), the authors provide a comprehensive toolkit for building more resilient forecasting systems.
Future research directions include evaluating these attacks against a broader range of architectures, such as Transformer-based or State-Space models, and optimizing the computational efficiency of the Mini-max defense to make it more viable for real-time, large-scale applications.
Summary updated