Quantum Certification Fails As 5% Classical Admixture Defeats Detection, Achieving 0.50 ROC AUC

Quantum key distribution promises secure communication, but its security rests on verifying that observed quantum correlations genuinely arise from entanglement and are not the result of an eavesdropper’s manipulation. Davut Emre Taşar, working as an independent researcher in Madrid, Spain, and colleagues demonstrate a critical vulnerability in current certification methods, revealing how an adversary can effectively evade detection. The team’s research establishes fundamental limits to quantum certification, showing that an eavesdropper requires only a small amount of classical data, just five per cent, mixed with quantum data to completely fool all tested detection methods. This work identifies a significant flaw in how detection performance is typically evaluated, revealing that a common calibration practice artificially inflates security estimates by as much as 44 percentage points, and highlights that adversaries employing highly realistic strategies can even outperform current quantum hardware on standard certification benchmarks. These findings necessitate a revised approach to security validation, demanding rigorous adversarial testing and the adoption of corrected evaluation methodologies to ensure the true security of quantum communication systems.

Detection systems are investigated, establishing fundamental adversarial limits for quantum certification using Eve-GAN, a generative adversarial network trained to produce classical correlations indistinguishable from quantum statistics. The central finding reveals that when Eve interpolates her classical correlations with quantum data at a mixing parameter α ≥ 0. 95, all tested detection methods achieve a ROC AUC of 0. 50, equivalent to random guessing. This demonstrates an eavesdropper requires only 5% classical admixture to completely evade detection.

Eve-GAN Mimics Quantum Correlations for Deception

This research details how a generative adversarial network, called Eve-GAN, can successfully mimic quantum correlations and deceive quantum-based detection schemes. The team designed Eve-GAN to generate data closely resembling genuine quantum correlations, consisting of a generator creating simulated quantum data and a discriminator attempting to distinguish between genuine and generated data. Experiments demonstrate Eve-GAN’s ability to generate correlations statistically indistinguishable from those produced by quantum systems. Detailed analysis of the α sweep, representing the mixing ratio between genuine quantum data and GAN-generated data, revealed a significant decrease in detection rates as the proportion of generated data increased. The research also identified a clear phase transition, where detection becomes unreliable as the CHSH value approaches the classical limit. Data from IBM quantum hardware showed that Eve-GAN can achieve a CHSH value comparable to, and even exceeding, that of the actual quantum device.

Eve-GAN Evades Quantum Key Distribution Defenses

The work presents a critical analysis of security vulnerabilities in quantum key distribution (QKD) systems, demonstrating that a classical adversary, employing a generative adversarial network (Eve-GAN), can effectively evade detection with remarkably high success rates. Researchers discovered a significant threshold where, with only 5% classical data mixed with genuine quantum data, all tested detection methods fail, achieving performance equivalent to random guessing. The team developed Eve-GAN, a network trained to generate classical correlations indistinguishable from quantum correlations, and rigorously tested its ability to mimic quantum data. Experiments demonstrate that Eve-GAN can achieve a CHSH value of 2.

736 on IBM Quantum hardware, exceeding the performance of the actual quantum device which measured 2. 691 on the same metric. This “Eve advantage paradox” highlights the potential for classical adversaries to outperform noisy quantum systems on standard certification benchmarks. Further investigation revealed a systematic flaw in existing certification methodologies, stemming from the use of same-distribution calibration, which inflates apparent detection accuracy by as much as 44 percentage points. Corrected methodology employing cross-distribution calibration is therefore recommended for accurate security assessment.

Eavesdropping Threat Undermines Quantum Key Distribution

This research establishes fundamental limits to quantum certification, revealing vulnerabilities in current security assessments for quantum key distribution. Through systematic analysis using a generative adversarial network, Eve-GAN, the team demonstrates that an eavesdropper requires only a 5% admixture of classical correlations to evade all tested detection methods, a remarkably small margin that significantly challenges existing security assumptions. The team developed Eve-GAN, a network trained to generate classical correlations indistinguishable from quantum correlations, and rigorously tested its ability to mimic quantum data. The research identified a sharp phase transition at a CHSH value of 2.

05, below which distinguishing between classical and quantum correlations becomes statistically impossible. Notably, the research demonstrates a counterintuitive “Eve advantage”, where the classical adversary outperforms noisy quantum hardware on standard certification measures. The authors recommend mandatory adversarial testing using GAN-based attacks, the implementation of cross-distribution calibration protocols, and the development of multi-feature detection methods combining CHSH with other relevant parameters.