Discriminator-based evaluation of the joint-structure attack

Evaluate the joint-structure adversarial LOB-data attack using the trained discriminator component of LOB-Bench, which measures the empirical separability of real and generated order-book trajectories.

Background

The paper constructs an adversarial Level-2 limit-order-book sequence by preserving prices, timestamps, message features, and Level-1 quantities while rearranging size and order-count information at deeper levels. This perturbation preserves most marginal statistics and the price-impact response used by LOB-Bench, yet disrupts temporal, cross-level, bid–ask, and Level-1/deep-book dependencies. LOB-ID detects the resulting distortion, whereas the distributional-statistics and impact-response components of LOB-Bench do not.

LOB-Bench also includes a trained discriminator that assesses how well real and generated trajectories can be separated. The paper omits this component in the attack experiment because the experiment is intended to isolate the sensitivity of the selected statistics and impact-response measures. Applying the discriminator to the same adversarial data is explicitly left unresolved.