Optimal Attack against Autoregressive Models by Manipulating the Environment

Yiding Chen, Xiaojin ZhuAAAI 2020 · DOI

Abstract

This paper investigates the vulnerability of autoregressive (AR) time-series forecasting models to sequential adversarial attacks. Unlike traditional batch attacks on static data, the authors frame the problem as an optimal control task where an adversary subtly manipulates an environment's state over time. By leveraging Linear Quadratic Regulator (LQR) and Model Predictive Control (MPC) frameworks, the study demonstrates how an attacker can steer a victim's forecasts toward a specific target trajectory while minimizing the cost of intervention. Experimental results on synthetic and real-world data reveal that optimal control strategies significantly outperform simple greedy attacks, which often lead to system divergence.

Introduction

Most research in adversarial machine learning focuses on "batch" attacks, such as adding a single perturbation to an image to deceive a classifier. However, many critical real-world systems—including financial markets, weather sensors, and industrial controllers—operate sequentially. These systems often rely on autoregressive (AR) models, which predict future values based on a history of past observations.

The authors identify a significant security gap in how we understand vulnerabilities in these recursive systems. In a "state attack" scenario, the attacker does not merely change a static data point; they perturb the underlying dynamical system (the environment) at each time step. Because AR models are recursive, a small change today ripples through all future forecasts. The challenge for an adversary is to find a sequence of perturbations that achieves a long-term deceptive goal without being so large that the attack becomes obvious or prohibitively expensive.

System Architecture and Problem Formulation

The attack is modeled as a multi-layered interaction between an adversary, the environment, and a forecaster. The environment evolves according to a state $x_t$, which the forecaster observes to produce predictions $y_{t+k|t}$. The adversary introduces a control input $u_t$ to the environment's transition function to influence these future observations.

System architecture showing the interaction between the adversary, environment, and forecaster.

As shown in the architecture above, the adversary's goal is to minimize a cost function that balances two competing objectives: the accuracy of the deception (how close the victim's forecast is to the adversary's target $y^\dagger$) and the "expenditure" or magnitude of the perturbations $u_t$.

Methodology

The authors propose three primary attack strategies based on the adversary's level of knowledge:

  1. White-Box Linear Attack (LQR): When both the environment and the forecasting model are linear and known, the attack is formulated as a Linear Quadratic Regulator (LQR) problem. This provides a closed-form mathematical solution for the most efficient sequence of perturbations.
  2. White-Box Nonlinear Attack (MPC + iLQR): For nonlinear environments, the authors employ Model Predictive Control (MPC). This involves planning several steps into the future using an iterative LQR (iLQR) algorithm, executing the immediate next step, and then re-planning as new observations arrive.
  3. Black-Box Attack (System Identification): If the environment dynamics and the victim's model are unknown, the attacker observes the system during an initial "buffer" period. They use this data to perform System Identification (estimating models via Least Squares or Maximum Likelihood) before proceeding with the MPC/LQR strategy.

Comparative Analysis: LQR vs. Greedy Attacks

A critical insight of the paper is the "myopic failure" of greedy attacks. A simple greedy attacker only attempts to fool the next immediate forecast ($y_{t+1|t}$). While this seems intuitive, it fails to account for the long-term dynamics of the AR model.

Comparison of forecasting behavior under no attack, optimal LQR attack, and unstable greedy attack.

In experiments on an oscillating AR(3) system (visualized above), the LQR attack successfully drives the forecasts toward the target with a mean cost of 11.5. In contrast, the greedy approach leads to massive instability and divergence, resulting in a catastrophic cost of 1492. This demonstrates that optimal control is superior because it understands that an action taken now can "set the stage" for easier manipulation in the future.

Key Results and Real-World Application

The researchers validated their framework using both synthetic AR processes and real-world data from the US Real GNP (1947–1991).

  • GNP Manipulation: Using the nonlinear MPC+iLQR approach, the attacker successfully manipulated forecasts of US economic growth toward a target of 0.01, significantly outperforming greedy baselines in both precision and cost-efficiency.
  • Black-Box Feasibility: The study found that an attacker does not need prior knowledge of the victim's model. By observing the environment's natural evolution for a short period, they can build a "good enough" proxy model to launch effective attacks that eventually match the performance of an "oracle" attacker with full knowledge.

Limitations and Future Work

While the framework is robust, several challenges remain. The current black-box approach assumes the attacker can correctly guess the "order" ($p$) of the autoregressive model; an incorrect assumption may lead to failure. Furthermore, the paper focuses on minimizing attack magnitude but does not explicitly model a "defender" or anomaly detection system that might look for statistical signatures of manipulation. Future research will look toward high-dimensional multivariate time series, such as those found in global supply chains, where the state space is significantly more complex.

Summary updated

BibTeX
@inproceedings{chen2020optimal,
  author = {Yiding Chen and Xiaojin Zhu},
  title = {{Optimal Attack against Autoregressive Models by Manipulating the Environment}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i04.5760},
}
Figure