The current system of detecting attacks based on high Quantum Bit Error Rate (QBER) during the Quantum Key Distribution (QKD) is easy to bypass by modern tools. Practical scenarios suggests, smart attackers now blend them with environmental noise by intercepting very small portion of signal. In this paper, we introduce 'Noise-Camouflaged Attacks' (NCA). These threats hide within the thermal drift of the fiber, making them undetectable from environmental noise. We demonstrate that standard Deep Learning defenses (e.g. LSTMs) struggle to resolve these sub-Poissonian statistical patterns, suffering from high computational latency and poor interpretability. Instead, we propose a physics-informed approach. We show that these attacks produce a unique fingerprint by 'clamping' the inherent variance of the photon stream, forcing the statistics into a sub-Poissonian regime (F<1) a state highly improbable under standard Markovian thermal drift. By explicitly training a lightweight XGBoost classifier on these Fano Factor artifacts, we achieve superior detection accuracy (AUC 0.85) compared to heavy deep learning baselines (AUC ≈0.77). Furthermore, our featurebased method demonstrates a wire-speed inference latency of approximately 20 ns, making it a suitable solution for real-time FPGA security integration.