Mitigating Label-Flipping Threats in Deep Algorithmic Trading Systems via InfoNCE Contrastive Regularizatio

Adversarial data corruption poses severe systemic financial risks to deep learning-based automated trading architectures operating within high-frequency limit order books (LOB). In this paper, we expose the inherent limitations of deep long short-term memory (LSTM) networks against targeted label-flipping database injections, demonstrating that traditional multi-layer architectures remain highly susceptible to strategically crafted dataset poisoning. To neutralize this threat, we present a robust adversarial deflection framework powered by an InfoNCE contrastive protective shield. Our approach maps clean and poisoned multi-dimensional LOB temporal states onto an invariant geometric hypersphere projection space. Empirical validation reveals that while a highly optimized baseline model undergoes significant degradation under a 40% label corruption strategy, our contrastive pipeline preserves execution integrity, achieving an optimal 96.87% validation accuracy. Furthermore, timing benchmarks reveal a mean inference latency of 0.4959 ms, safely outperforming the strict 1.0 ms enterprise operational constraint. These results demonstrate the feasibility of real-time geometric regularizers for secure quantitative execution systems.