Introduction
Deep Neural Networks (DNNs) are increasingly utilized in the financial sector to replace computationally intensive "oracles," such as PDE solvers, for pricing complex derivatives. While the vulnerability of classification models to "backdoor" attacks is well-documented, regression models—where outputs are continuous—present a different set of challenges and risks. This paper explores how an adversary can plant a stealthy backdoor in a deep regression model and proposes a novel defense mechanism based on local error maximizers.
The Localized Poisoning Attack
A standard backdoor attack involves adding mislabeled samples to the training set. However, in regression, simply adding outliers can lead to "smoothing" by the optimizer, which either degrades global performance or fails to create a sharp trigger.
To overcome this, the authors introduce Localized Poisoning. In this approach, the attacker surrounds poisoned samples with "clean" samples that follow the ground-truth distribution. This "buffer" forces the model to learn a sharp, surgical deviation in a specific region of the input space without affecting the model's behavior elsewhere. As shown in the experimental results, increasing the number of localizing clean samples ($N_{Clean}$) helps maintain low error on legitimate data while ensuring high attack success.
Gradient-Based Defense
The proposed defense mechanism identifies "suspicious" regions by finding local error maximizers. The core logic rests on the Proximity Paradox: while high error is expected in sparse regions of the input space (extrapolation), high error in densely sampled regions (interpolation) is a strong indicator of a backdoor.
The defender uses a gradient-ascent method to locate points where the model's error relative to the ground-truth oracle is maximized. For each local maximizer, a "proximity score" is calculated based on the number of training samples in its vicinity. Points that exhibit both high error and high proximity are flagged as malicious.
Key Results
The study utilized a synthetic financial dataset based on "down-and-out" put options. Key findings include:
- High Success and Stealth: With poisoning levels as low as 5-8% (including localizing samples), attack success rates exceeded 95%. Crucially, the Mean Squared Error (MSE) on clean test data often remained stable or even improved, making the attack invisible to standard validation metrics.
- Effective Detection: Local error maximizers within the backdoor region were found to be in the 99.9th percentile for proximity to training data. This clear separation allows defenders to isolate malicious samples with high confidence.
- Mitigation: By removing flagged samples and retraining the model, the authors reduced the "Attack MSE" by approximately 50%, significantly neutralizing the adversary's impact.
Conclusion
Localized poisoning represents a significant threat to deep regression models in finance, where standard performance metrics fail to detect surgical manipulations. However, the same density of data required to make an attack effective also provides a signal for detection via gradient-based auditing. Future work will investigate these dynamics in higher-dimensional "exotic" option pricing surfaces where backdoors may be even harder to hide.
Summary updated