Sub-Millisecond Contrastive Embedding Shields for Safeguarding Automated Pricing Engines Against High-Frequency Order Book Spoofing

Abstract—Modern electronic exchanges rely heavily on deep sequence learning networks to predict microsecond price velocity and manage localized inventory distribution. However, these unprotected deep-learning engines remain highly vulnerable to adversarial data-poisoning attacks, such as localized high- frequency limit order book (LOB) spoofing. While malicious actors use rapid-fire ghost orders to artificially distort volume vectors and trick pricing systems, platforms cannot simply implement blanket bans on high-speed bots without destroying crucial market liquidity and widening bid-ask spreads. To resolve this paradox, this paper introduces a real-time preprocessing security framework named the Behavioral Shield Encoder. Combining a 1D Temporal Convolutional Network (Conv1D) with Gated Recurrent Units (GRU), our architecture leverages Supervised Contrastive Learning to project volatile microsecond inputs into an invariant 32-dimensional behavioral embedding space. Empirical validation indicates that our system completely neutralizes targeted Fast Gradient Sign Method (FGSM) perturbations. Specifically, it achieves an absolute Adversarial Deflection Rate (ADR) of 100.00% while preserving efficient-market classification baselines. Crucially, the system clocks a mean processing overhead of just 0.4601 ms, comfortably clearing the sub-millisecond production thresholds required by institutional high-frequency execution pipelines.