While the majority of adversarial machine learning research focuses on classification—where an attacker’s success is defined by a discrete "label flip"—a growing body of work reveals that models predicting continuous values are equally, if not more, vulnerable. In regression tasks, the absence of discrete decision boundaries changes the nature of both attacks and defenses. Instead of pushing an input across a boundary, attackers must systematically shift a numerical output along a continuous spectrum, often with catastrophic consequences in safety-critical domains like autonomous driving and energy forecasting.
The Leverage of Continuous Outputs
In classification, an attacker is constrained by the set of available categories. However, Liu et al. demonstrate that continuous response variables provide significantly more leverage. By optimizing the response variable alongside feature values, attackers can increase the effectiveness of poisoning attacks by nearly sevenfold compared to standard classification baselines.
This vulnerability stems from the fine-grained control an attacker can exert over the model's behavior. Because there is no "margin" or "boundary" to cross, even small, imperceptible perturbations can be optimized to maximize the mean square error or steer a prediction toward a specific malicious target. As Chernikova et al. show in the context of autonomous steering, these perturbations can increase prediction errors by over 60 times, potentially causing a vehicle to veer off-course without any human-detectable change in the input image.
Defining Success in a Boundaryless Space
A central challenge in regression security is defining what constitutes a "successful" attack. In classification, the metric is simple: did the label change? In regression, researchers must define specific thresholds of deviation. Scholz et al. propose a taxonomy for these objectives:
- Untargeted Error Maximization: Simply making the prediction as inaccurate as possible.
- Semi-targeted Range Enforcement: Forcing the output into a specific, dangerous range.
- Fully Targeted Output Shaping: Forcing a time-series forecast to follow a specific malicious pattern or constant value.
This last category is particularly dangerous for sequential tasks. For example, Lin et al. introduced BACKTIME, a framework that forces multivariate time-series models to predict specific attacker-defined sequences. Similarly, in image quality assessment, Zhang et al. found that attackers can push continuous scores toward extreme opposite boundaries, effectively turning high-quality inputs into low-quality outputs and vice versa.
Poisoning and Scalable Backdoors
Regression models are susceptible to both inference-time evasion and training-time poisoning. In poisoning attacks, an adversary injects malicious data into the training set to skew the resulting model. Liu et al. highlight that the continuous nature of regression makes these attacks harder to detect using standard statistical tools, as the malicious points may not appear as obvious outliers.
Backdoor attacks in regression also require a different approach than in classification. Traditional backdoors force a model into a single "wrong" category when a trigger is present. In contrast, Yu et al. developed BAIQA, which uses a "scalable trigger." By adjusting the intensity of a frequency-domain trigger, an attacker can precisely manipulate a predicted quality score to any desired value. This treats the backdoor as a mathematical offset rather than a fixed label override, allowing for fine-grained control over the model's continuous output.
Why Classification Defenses Fail
Many standard defenses, such as adversarial training or trigger reverse-engineering, do not translate directly to regression. Lingyu Du argues that existing backdoor defenses fail because they cannot exhaustively check an infinite, continuous output space for malicious mappings. Furthermore, regression models often lack the discrete decision neurons that classification defenses target; instead, malicious triggers may influence the entire feature vector globally.
Specialized defenses for regression often focus on robust statistics or feature consistency:
- TRIM: Liu et al. proposed an iterative "trimming" defense that identifies and removes high-residual points during training to recover a clean model.
- Randomized Ablation (RA): Wu et al. developed a certified defense for crowd counting that provides mathematical guarantees by training on sampled versions of images, mitigating the impact of local adversarial patches.
- GD+: Meyer Scetbon showed that while standard early-stopping in gradient descent protects against some attacks, more complex non-Euclidean perturbations require feature-dependent learning rates to achieve near-optimal robustness.
Vulnerability in Dense and Sequential Tasks
The threat extends beyond simple scalar regression to dense pixel-wise tasks and sequential forecasting. Wong et al. demonstrated that monocular depth prediction networks can be tricked into "seeing" objects disappear or shifting the entire scene's depth by 30%. In these cases, the model's reliance on non-local context becomes a liability, as an attacker can manipulate one part of an image to affect the regression output in another.
Similarly, in optical flow estimation, Schmalfuss et al. found a troubling trade-off: the most accurate modern models are often the most vulnerable to tiny, adversarial changes in input. This suggests that as regression models become more expressive and precise, they may inadvertently become more sensitive to the subtle, high-dimensional perturbations that characterize adversarial attacks.
Key Foundational Work
Manipulating Machine Learning: Poisoning Attacks and Countermeasures for Regression Learning (2018) — contributed 3 concept elements to this direction.
Perturbation Analysis of Learning Algorithms: A Unifying Perspective on Generation of Adversarial Examples (2018) — contributed 4 concept elements to this direction.
Manipulating Machine Learning: Poisoning Attacks and Countermeasures for Regression Learning (2018) — contributed 2 concept elements to this direction.