Deep learning-based defense against phase-injection attacks in pilot-multiplexed continuous-variable quantum key...

Area of Science:

• Quantum Information Science
• Cybersecurity
• Machine Learning

Background:

• Accurate phase estimation is critical for continuous-variable quantum key distribution (CV-QKD), especially in local local oscillator systems using pilot tones.
• Pilot-tone-based phase recovery is vulnerable to phase-injection attacks, which can compromise system security by subtly biasing phase estimation.
• Adversarial manipulation can occur without significant changes in observable channel parameters, posing a stealthy threat.

Purpose of the Study:

• To demonstrate a novel phase-injection attack targeting pilot-multiplexed CV-QKD systems.
• To propose and evaluate a robust phase estimation scheme to mitigate such attacks.
• To enhance the security and performance of CV-QKD systems against adversarial manipulation.

Main Methods:

• A phase-injection attack is demonstrated by introducing non-Gaussian perturbations into pilot tones.
• A Gated-Robust Long Short-Term Memory (GR-LSTM) network is proposed for phase estimation.
• The GR-LSTM is trained with legitimate phase data and synthetic attack-induced perturbations to learn robust temporal representations.

Main Results:

• The proposed GR-LSTM scheme effectively restores robust phase estimation against the demonstrated attack.
• Numerical simulations and a 20 km fiber experiment validate the effectiveness of the GR-LSTM.
• The average secret key rate improved by 38.9%, from 1.143 Mbps (standard LSTM) to 1.588 Mbps (GR-LSTM).

Conclusions:

• The GR-LSTM-based phase estimation offers a practical security enhancement for pilot-multiplexed CV-QKD systems.
• The developed method successfully mitigates vulnerabilities associated with phase-injection attacks.
• This work contributes to the development of more secure and reliable quantum communication protocols.