Hybrid Implementation for Untrusted-node-based Quantum Key Distribution Network

A conceptual schematic of the experimental setup is provided in Fig. 2a. On the source side, two narrow-linewidth semiconductor continuous-wave lasers (Rio PLANEX) with a central wavelength of 1549.32 nm and a linewidth of approximately 2 kHz are frequency-locked to each other using a heterodyne optical phase-locked loop (OPLL). Each laser is capable of maintaining optical frequency stability within 5 MHz over six hours. Alice’s continuous light is divided by a 50:50 beam splitter (BS), while Bob’s continuous light, after passing through a phase modulator (PM), undergoes a similar division by a 50:50 BS. One beam from each side is used for quantum-state encoding, while another beam from Alice is transmitted through the servo channel (0.18 dB/km) and coupled with the corresponding beam from Bob to generate a beat note. A polarization controller (PC) on Bob side aligns the polarization of the beams involved in the beat frequency, maximizing the magnitude of the beat signal. An acousto-optic modulator (AOM) is positioned after the BS to introduce a precise optical frequency shift of 300.7 MHz, which also serves as the reference frequency. By locking the beat note to this signal, the OPLL and the PM0 on Bob side achieve precise optical phase locking. The evaluation of OPLL is discussed in detail in sec. III.

The locked coding beams at both Alice’s and Bob’s sides are processed through two cascaded intensity modulators (IMs), which chop the light into pulse sequences with a 50 MHz repetition rate and a temporal width of 1.5 ns, achieving an extinction ratio exceeding 30 dB. These pulses are directed into Faraday-Michelson interferometers (FMIs)  (?, ?), which split each incident pulse into two time-bin pulses separated by a 10 ns interval: the earlier time-bin pulse $\lvert e\rangle$ and the later time-bin pulse $\lvert l\rangle$. A circulator placed before each FMI filters out any laser light reflected backward. For operations in MDI-QKD mode, IM3 and IM4 handle basis selection and decoy-state modulation. Phase coding in the X basis is performed by adding a relative phase {$0$, $\pi$} to the $\lvert l\rangle$ pulse via PM1, while time-bin coding in the Z basis is achieved by chopping either the earlier or later time-bin pulse, corresponding to a bit value of 0 or 1, respectively. These IMs effectively suppress ambient noise, ensuring a high extinction ratio for both the vacuum state and the eigenstates of the Z basis, resulting in an inherent error rate of just 0.1% for Z basis. Finally, PM2 is used to randomize the phase of both time-bin pulses, mitigating potential attacks on the light sources.

For operations in TF-QKD mode, IM3 and IM4 modulate and encode the later time-bin pulse $\lvert l\rangle$ into the quantum pulse $\lvert q\rangle$ with random intensities while retaining the earlier time-bin pulse $\lvert e\rangle$ as the reference pulse $\lvert r\rangle$ for phase calibration. The majority of duty periods in TF-QKD mode are allocated for transmission, with a smaller portion reserved for compensation. During transmission, phase encoding {$0$, $\pi$}, representing 0 or 1 bit, respectively, is applied to $\lvert q\rangle$ by PM1, along with a random phase slice of $2\pi i/16,\ i\in\{1,\ldots,16\}$. Simultaneously, PM2 alternately scans phases $0$ and $\pi/2$ solely on $\lvert r\rangle$, referred to as two-phase scanning, to acquire global phase drifts across the quantum channel. A detailed description of the phase estimation method is provided in Sec. III. During compensation, the two-phase scanning procedure is applied simultaneously to both $\lvert q\rangle$ and $\lvert r\rangle$, with PM1 ceasing its encoding functions, while IM3 and IM4 modulate both pulses into the reference intensity. The same procedure is used for phase calibration in MDI-QKD mode, yielding a interference visibility of approximately 99.5% in the X basis. An optical delay (OD) is inserted on Alice side after the IMs to fine-tune the photon arrival time with picosecond precision. Finally, an electrically tunable optical attenuator (EVOA) attenuates the pulses to the single-photon level before they are transmitted to Charlie through ultralow-loss fibers (0.17 dB/km).

On the measurement side, the incident pulses from Alice and Bob are first routed through a polarization-maintaining module comprising an electronic polarization controller (EPC) and a polarization beam splitter (PBS) to ensure consistent polarization of the outgoing pulse pairs. These pairs are then coupled at a BS for interference and subsequently detected by superconducting nanowire single-photon detectors (SNSPDs) from PHOTEC (model P-SPDNS). Both detectors, D0 and D1, feature a detection efficiency of 68.5% and a dark count rate of 10 Hz. After accounting for the insertion loss of all components, the overall detection efficiency is 64%, and the dark count rate per pulse is approximately $1.5\times 10^{-8}$ when gated with a 1.7 ns time window. In TF-QKD mode, single-photon detection events are recorded by a time-to-digital converter (TDC), with distinct time windows assigned for quantum and reference pulses to facilitate post-processing. In MDI-QKD mode, detection events from D0 and D1 are recorded to calculate coincidence events, enabling the Bell state measurement. By switching the encoding modes of Alice and Bob, the hybrid system smoothly transitions between TF-QKD and MDI-QKD.

We follow the SNS-QKD protocol with AOPP method  (?) and the MDI-QKD protocol with double-scanning method  (?) in this experiment. For TF-QKD mode, the X basis is encoded in phase of WCPs, and Z basis is encoded in intensity {$\mu$, $\nu$, $\omega$, $o$}, where $\mu$ and $o$ denote the signal and vacuum intensity, respectively, and $\nu$ and $\omega$ are the decoy intensities. Alice (Bob) randomly decides whether the ith time window is a decoy window or a signal window. For a signal window, Alice (Bob) randomly chooses $\mu$ ($o$) with probabilities $\epsilon$ ($1-\epsilon$), indicating a sending (not-sending) choice in SNS-QKD protocol. Alice (Bob) actually prepares a phase-randomized WCP with intensity $\mu$. For a decoy window, Alice (Bob) randomly chooses sources from $\nu$, $\omega$ and $o$, and prepares a phase-randomized WCP. The prepared pulse pairs are send to Charlie to perform interferometric measurements. For MDI-QKD mode, the X basis is encoded in the relative phase of earlier and latter pulse, and Z basis is encoded in its time-bins. Signal intensity $\nu$ is only prepared in Z basis while decoy intensities $\nu$ and $\omega$ are in X basis. Both Alice and Bob do not choose any bases for vacuum states $o$. The prepared quantum states are similarly send to Charlie to perform the $\lvert\psi^{-}\rangle$ projection measurement. All intensities and probabilities are optimized by the differential evolutionary algorithm. As a proof-of-principle demonstration, we encode the WCPs into 16 phase slcies between $0$ and $2\pi$, and generate an 2000-bit pseudo-random pattern, which is preloaded into an arbitrary waveform generator (AWG) to drive the modulators repeately. Tab. 1 lists the system parameters adopted in this experiment.

The SNS-QKD protocol and MDI-QKD protocol are performed over a range of different channel distance. The secure key rates calculated in the finite-size regime from the experiment and theoretical simulations shown in Fig. 3, together with the absolute PLOB bound  (?) which is calculated as $-log_{2}(1-\eta)$ with 100% detection efficiency. The asymptotic case refers to the analysis of QKD when the total number of pulses N tends to infinity. The asymptotic lines present a performance potential of the hybrid system in both protocols. In practical implementations, the finite-size analysis provides the security against statistical fluctuations for the final secure keys, bounded by the Chernoff bound  (?, ?).

We therefore conduct experiments at total distances from 150 km to 431 km between Alice and Bob in a symmetric fiber setup. Positive secure keys are generated for all these distances in finite-size regime, i.e., $N=10^{11}$ and $N=10^{12}$ ($5\times 10^{7}$ and $5\times 10^{8}$ pattern repetition), and the detailed results can be found in Supplementary text. The experimental secure key rates of SNS-QKD scales as $O(\sqrt{\eta})$ and are in agreement with the theory. From 310 km to 431 km, the secure key rates break the absolute PLOB bound in both pulse sizes. The experimental secure key rates of MDI-QKD are measured at 150 km and 241 km with either pulse sizes, which hold agreements with the theoretical simulations. At 241 km fiber distance, the hybrid system sends $N=10^{9}$ repeated patterns continuously. For the first $N=5\times 10^{8}$, the system runs the SNS-QKD protocol and encodes the WCPs, while for the remaining repeated patterns, the system runs the MDI-QKD protocol and encodes the four kinds quantum states. In both protocols, Charlie simultaneously records the single-photon clicks of each detector window. And the global phase recovery and two-photon coincidence events are calculated by post-processing before sifting. We adopt the AOPP method to improve the key rates of SNS-QKD, which is a kind of two-way classical communication.

By utilizing the filter-enhanced two-phase scan, phase compensation is carried out for 200 $\mu$s every 10 seconds in both protocols at each distance, achieving an duty cycle of near 99% after accounting clock synchronizing and polarization compensation. The obtained secure key rates of TF-QKD in key bit per second (per pulse) are 2105.59 bps, 463.52 bps, 108.91 bps, 26.16 bps and 22.60 bps at 241 km, 310 km, 351 km, 400 km and 431 km fiber distance, respectively. And the obtained secure key rates of MDI-QKD in key bit per second are 107.61 bps and 0.72 bps at 150 km and 241 km, respectively. The experimental results demonstrate the feasibility of the hybrid implementation for QKD networks with untrusted-nodes. With a larger pulse number and a lower dark count rate, the current transmission distance can further achieve a distance similar to the Ref.  (?) and  (?).

The schematic of the OPLL is depicted in the inset of Fig. 2a. Inside the OPLL, a local oscillator (LO) clock at 37.5875 MHz is generated by the high-precision onboard clock of the field-programmable gate array (Xilinx ZCU111 development board). This signal is then frequency-multiplied by 8 using an electronic phase-locked loop (PLL) and amplified by a radio-frequency (RF) amplifier, producing the 300.7 MHz RF signal that drives the AOM and serves as the reference clock for the mixer in the OPLL system. The beat note, converted through a 12 GHz bandwidth photodetector (Optilab PR-12-B-M), along with the reference clock, is fed into a loop filter, providing feedback for the proportional-integral-derivative (PID) algorithm ($g_{p}=-6$ dB, $f_{I}=16$ kHz, $f_{D}=250$ kHz). The PID algorithm calculates the necessary adjustments to the control signal, performing fast wavelength modulation of Bob’s laser with the direct current signal in the range of $\pm 4$ V, ensuring a small and stable frequency difference between the two lasers. The residual phase noise is then processed by the PM0 on Bob side, which is controlled by a low-pass filter. The measured relative Allan deviation of the beat note in the time domain, ranging from $10^{-11}$ to $10^{-13}$, demonstrates excellent frequency stability. The measurement results as well as the beat note spectra are given in supplementary text. In the current setup, the residual mean-square phase error associated with the OPLL is about $5.6\times 10^{-3}\ \mathrm{rad^{2}}$.

In this section, we showcase the efficient post phase-estimation method by using the FMIs. As illustrated in Fig. 2b, after passing through the FMI, the optical pulses are split into two time-bin pulses: the earlier bin pulse $\lvert e\rangle$ and the latter bin pulse $\lvert l\rangle$. The earlier pulses are modulated into reference pulses $\lvert r\rangle$ using IMs, while the latter pulses are modulated into quantum pulses $\lvert q\rangle$. Then, the global phase of reference pulses $\lvert r\rangle$ when arriving at the beam splitter (BS) on Charlie side can be written as

$\displaystyle\phi_{ref}=\phi_{0}+\varphi_{short}+\phi_{channel},$

(1)

where $\phi_{0}$ denotes the initial phase of lasers plus the optical path before input into the interferometers, $\varphi_{short}$ denotes the phase applied by the short arm of interferometers, and $\phi_{channel}$ denotes the phase applied by the quantum channel link after outgoing the interferometers. Similarly, the global phase of quantum pulses $\lvert q\rangle$ when arriving at the BS on Charlie side is

$\displaystyle\phi_{q}=\phi_{0}+\varphi_{long}+\phi_{channel},$

(2)

where $\varphi_{long}$ denotes the phase applied by the long arm of interferometers. In our scheme, since the reference pulses $\lvert r\rangle$ and quantum pulses $\lvert q\rangle$ are emitted from the same laser on either the Alice or Bob side, and travel through nearly identical channels with consistent polarization over a period of transmission time, the initial laser phase and the phase of quantum channel can be considered nearly equal  (?, ?). The path difference between the long and short arms of FMI allows us to adjust the phase applied by the PM inserted in either arm, thereby equalizing $\phi_{ref}$ and $\phi_{q}$ within a certain period.

On the detection side, the D0 and D1 detectors operate with two separate time-bin windows to distinguish between reference and quantum pulses. The interference results of the reference pulses from Alice and Bob are projected onto the earlier time-bin windows of the D0 and D1 detectors, enabling the estimation of global phase drifts across the quantum channel. Over extended periods, however, the phase differences between the reference and quantum pulses are primarily attributed to fluctuations in the optical paths of the two arms of the interferometers. Theoretically, this problem can be solved by adopting phase calibrations similar to phase-encoding BB84 QKD or MDI-QKD utilizing asymmetric interferometers.

In our experiment, to compensate for the phase difference between pulse $\lvert r\rangle$ and $\lvert q\rangle$, Alice and Bob first modulate their both pulses into reference intensity. Either Alice or Bob performs the filter-enhanced two-phase scan using a PM located after the FMI, covering the arrival times of both $\lvert r\rangle$ and $\lvert q\rangle$. During the scan, the PM inside the FMI sweeps through $I=16$ phase slices, recording the corresponding scanning statistics for both $\lvert r\rangle$ and $\lvert q\rangle$. Subsequently, it selects and applies the phase slice that minimizes the counting difference between $\lvert r\rangle$ and $\lvert q\rangle$, calibrating the phase drifts of the two arms of the interferometer. During transmission, Alice and Bob modulate $\lvert q\rangle$ into signal, decoy, or vacuum intensities, applying phase randomization and encoding solely to $\lvert q\rangle$. Meanwhile, the PM positioned after the FMI continues to load the scanning signal and carries out the filter-enhanced two-phase scan exclusively on the time sequence of the $\lvert r\rangle$ pulses. All perturbations in the fiber channel affect $\lvert r\rangle$ and $\lvert q\rangle$ equally, enabling us to estimate $\phi_{q}$ with $\phi_{ref}$ and minimizing interruptions of quantum pulse transmission.

Drifts in the long and short arms of the FMI can disrupt this equivalence; thus, it remains necessary to periodically repeat the above steps to compensate for phase shifts between pulse $\lvert r\rangle$ and $\lvert q\rangle$. This process ensures minimal QBER and reduces residual estimation error, which is defined as

$\displaystyle\varDelta\phi_{error}=|\phi_{q}-\phi_{ref}|.$

(3)

We test this method over a range of fiber distance with an execution time of 40 $\mu$s for the two-phase scan. Fig. 4(a) shows the monitoring results of $\varDelta\phi_{error}$ with and without compensation at 300 km. It demonstrate that $\varDelta\phi_{error}$ can consistently maintain a stable phase slice for at least 10 seconds across long fiber distances. Fig. 4(b) shows the phase slice variations during the first 5 ms of global phase estimation from both the reference pulses $\lvert r\rangle$ and quantum pulses $\lvert q\rangle$, modulated with the same reference intensity. The overlap of these two stair curves highlights the effectiveness of this method. For a 10-second compensation period, the histogram of $\varDelta\phi_{error}$ exhibits a standard deviation of 3.98°, as shown in Fig. 4(c). In this configuration, the hybrid system can run continuously for several seconds in TF-QKD mode, requiring only 200 $\mu$s of five round scanning to compensate for the arm length fluctuations in the FMI, significantly enhancing the system’s duty cycle.

In previous TF-QKD experiments  (?, ?, ?, ?, ?, ?, ?, ?), a time gap of several hundred nanoseconds was typically introduced between strong reference frames and weak quantum pulses to allow SNSPDs to recover. In our scheme, the close temporal proximity of quantum and reference pulses negatively affects the signal-to-noise ratio (SNR), leading to increased misalignment errors. To address this, we integrate a least mean square (LMS) adaptive filter with the two-phase scan, enabling accurate estimation even at reduced reference intensity and effectively mitigating the impact of double Rayleigh scattering noise on misalignment errors. The noisy count acquired in two-phase scan can be expressed as $N(M)=\hat{N}(\hat{M})+\tilde{\epsilon}+\tilde{\delta}$, where $\hat{N}(\hat{M})$ is the expected count, $\tilde{\epsilon}$ is Gaussian-distributed additive noise, and $\tilde{\delta}$ accounts for systematic errors caused by shot-noise and overshoot amplitudes. After sweeping through $I$ phase slices while conducting $I$ times two-phase scan, we obtain two sequences of input $\textbf{x}_{r}(n)=[N_{0}^{r},N_{1}^{r},M_{0}^{r},M_{1}^{r}]^{T}$ and $\textbf{x}_{q}(n)=[N_{0}^{q},N_{1}^{q},M_{0}^{q},M_{1}^{q}]^{T}$, $n\in 1,...I$ where $N_{0}^{r(l)}$ and $N_{1}^{r(l)}$ denote the recording detector D0’s count of pulses $\lvert r\rangle$ ( $\lvert q\rangle$ ) when loading phase $0$ and $\pi/2$ within the first 20 $\mu$s and the second 20 $\mu$s, respectively, and $M_{0}^{r(l)}$ and $M_{1}^{r(l)}$ denote detector D1’s count of pulses $\lvert r\rangle$ ( $\lvert q\rangle$ ) when loading phase $0$ and $\pi/2$ within the first 20 $\mu$s and the second 20 $\mu$s, respectively. The filter is controlled by a set of 10 weights, denoted as $\textbf{w}=[w_{1},...,w_{10}]^{T}$. By convolving the input signal with w, we get the expected output sequences of the filter

	$\displaystyle\textbf{y}_{r}(n)=\textbf{w}^{T}\ast\textbf{x}_{r}(n)=[\hat{N}_{0}^{r},\hat{N}_{1}^{r},\hat{M}_{0}^{r},\hat{M}_{1}^{r}]^{T},$		(4)
	$\displaystyle\textbf{y}_{q}(n)=\textbf{w}^{T}\ast\textbf{x}_{q}(n)=[\hat{N}_{0}^{q},\hat{N}_{1}^{q},\hat{M}_{0}^{q},\hat{M}_{1}^{q}]^{T}.$		(5)

Then, $\textbf{y}_{r}(n)$ and $\textbf{y}_{q}(n)$ sequences are calculated by the two-phase scan method given in the supplementary text. Denote the output of two-phase scan algorithm as $\textbf{s}_{r(q)}(n)=2PS(\textbf{y}_{r(q)}(n))\in 1,...I$, we can define the desired signal $\hat{\textbf{d}}(n)$ as

$\displaystyle\hat{\textbf{d}}(n)=cos[\dfrac{\pi\cdot|\textbf{s}_{r}(n)-\textbf{s}_{q}(n)|}{I}]^{2}.$

(6)

And the reference signal is $\textbf{d}(n)=cos[\pi(|n-\mathit{K}|)/I]^{2}$, in which $\mathit{K}$ is the phase slice order corresponding to the minimum value of $|\textbf{x}_{r}(n)-\textbf{x}_{q}(n)|$. It is a linear constraint on the differences between the phase slices of pulse $\lvert r\rangle$ and $\lvert q\rangle$ considering no noise. Then, the mean square error is given as $J=E[\textbf{e}^{2}(n)]$, where $\textbf{e}(n)=\textbf{d}(n)-\hat{\textbf{d}}(n)$ is the cost function. Utilizing the steepest descent algorithm, the updated weight vector $\textbf{w}_{upd}$ is adjusted along the negative gradient direction during every compensation, given as

$\displaystyle\textbf{w}_{upd}=\textbf{w}+-\lambda\dfrac{\partial\textbf{e}^{2}(n)}{\partial\textbf{w}},$

(7)

where $\lambda=7.5\times 10^{-3}$ is the step-size. Through adaptive feedback cancellation, the LMS filter effectively reduces the effect of statistical fluctuation of fewer counts on the accuracy of estimation, and mitigates modulation deviation of PMs and intensity fluctuations from laser source. And the initialization of LMS filter is conducted with the data pre-acquired from an attenuation-simulated channel by traversing the entire voltage range of the PM to extract the expected output for precise calibration.

We examine the mean absolute error (MAE) between the global phase estimates with and without the filter-enhanced two-phase scan method under varying reference light intensities, as shown in Fig. 5. Considering the trade-off between MAE and Rayleigh scattering noise, the reference intensity is set as about 0.16 pW through a 400 km fiber, corresponding to the total SNSPD counts of 1.0 MHz. The two insets of Fig. 5 show the shape of interference curves acquired at the total counts of 0.5 MHz and 2.5 MHz, respectively. The results demonstrate that the proposed method can effectively reduce reference light intensity while maintaining estimation accuracy.

Supplementary Text
Figures S1 to S4
Tables S1 to S3