Trusting a Fast Spoofable Sensor Safely: Optimistic Attenuation Control Against Detector Blinding

Continuous-variable quantum key distribution (CV-QKD) measures the light field at the receiver with a homodyne detector, a device that is only linear over a finite range of input power. An eavesdropper can inject bright light to push it into saturation. Saturation clips the measurements, so the legitimate users underestimate the channel noise and keep producing key that the attacker can read. One published defense places an adjustable optical attenuator in front of the detector and sets its value from statistics of each received data block. This works only after the fact. The attenuation for block $t+1$ comes from post-processing block $t$, so an attacker who changes power faster than that loop runs can slip saturated pulses through during the delay.

The obvious fix is to add a fast classical photodiode that taps off a little incoming light and feeds its reading forward to the attenuator. The catch is that the tap is itself an optical component the attacker can manipulate. Fused-fiber beam splitters change their splitting ratio sharply with wavelength, so Eve can choose a wavelength that hides power from the tap. We propose treating the tap reading as a hint in optimistic online learning. In that family of algorithms, the learner takes a look-ahead step using a prediction of the next gradient, then corrects using the true one. Its regret, a measure of cumulative extra loss, grows with the prediction error rather than with time, and stays bounded at the worst-case rate when predictions are wrong. That is the guarantee one wants from a sensor that is fast but cannot be trusted. The same prediction error, logged every round, also serves as a label-free alarm.

Core Claim

Optimistic Online Mirror Descent (OMD) is the right way to combine a slow but trusted saturation estimate with a fast but spoofable power tap. The trusted estimate comes from the block statistics that already enter the security analysis, and it supplies the gradient $g_t$. The tap supplies the hint $M_t$. The attenuation setting $a_t$, in decibels, is chosen before block $t$ arrives by the two-step rule of Online Learning with Predictable Sequences. That paper introduced optimistic OMD and showed regret bounded by

$$\text{Regret}_T \le \frac{R^2}{\eta} + \frac{\eta}{2}\sum_{t=1}^T \|g_t - M_t\|_*^2,$$

where $R$ bounds the attenuation range, $\eta$ is the step size, and $\|\cdot\|_*$ is the dual norm.

The loss $f_t(a)$ is designed to be convex in $a$. It has two parts. A hinge term penalizes how far the attenuated signal power sits above the detector's linear limit. A linear term charges for the signal-to-noise ratio given up by attenuating more than necessary.

The claim is a dichotomy that the experiments are built to test. In the first case, the attacker varies power in a way the tap reports honestly. Then the hint is accurate, the controller tracks the attack within a fraction of a block, the detector stays linear, and the noise estimate stays correct. The existing finite-size key-rate check therefore remains valid. In the second case, the attacker spoofs the tap. Then the residual $\|g_t - M_t\|$ jumps, the controller falls back to reactive behaviour with bounded regret rather than following the attacker's fake readings, and a sequential test on the residual raises an alarm. In neither case can Eve get the controller to hold the detector in saturation while Bob believes it is linear for longer than a quantity governed by the hint error.

Relation to Prior Work

The method comes from optimistic online learning. Online Learning with Predictable Sequences supplies the update and the bound above. Online Optimization for Learning to Communicate Over Time-Correlated Channels is the closest bridge to physical systems. It tuned classical decoders and chose codebooks with optimistic OMD, used past channel states as the hint, and outperformed Online Gradient Descent and least squares. Its hint is a benign time correlation and its channel has no adversary. Here the hint comes from a separate sensor that an adversary can partly control, and that changes what the analysis has to deliver.

The research file lists one open problem that motivates this directly: optimism's guarantees are conditional, and become fragile when the hint is built from corrupted feedback. We turn that weakness into a design principle. Corruption shows up in the residual, and the bound degrades to the worst-case rate rather than diverging.

Two papers motivate a test against an attacker who also adapts. Faster Rates for Convex-Concave Games and Acceleration through Optimistic No-Regret Dynamics show that optimistic updates damp the cycling that plain gradient players fall into in zero-sum games. We use this only as a hypothesis. A purely reactive attenuator facing an adaptive Eve may oscillate, and an optimistic one may not. Their acceleration rates depend on strong convexity and smoothness that our clipped saturation loss lacks, so we do not import them.

The detector side supplies the threat model. Defending Against the Homodyne Detector-Blinding Attack on Continuous-Variable Quantum Key Distribution Using an Adjustable Optical Attenuator gives the saturation model, the attenuator architecture, and the block-feedback loop that we make predictive. It assumes the attenuator responds instantly and the estimate arrives without delay, so its latency loophole is exactly what we target.

Attacking a practical quantum-key-distribution system with wavelength-dependent beam-splitter and multiwavelength sources is what makes the tap spoofable. It measured coupling ratios of 0.986 at 1470 nm and 0.003 at 1290 nm against roughly 50:50 at 1550 nm. It also argued that passive filters can be overcome with more power. Underwater Wavelength Attack on Discrete Modulated Continuous-Variable Quantum Key Distribution shows that monitoring the local oscillator, the strong reference beam the homodyne detector mixes the signal with, can itself be defeated by combined wavelength and calibration attacks. That is a further reason not to trust any single monitor outright.

The machine-learning detectors surveyed in Secure Continuous-Variable Quantum Key Distribution with Machine Learning are offline supervised classifiers: SVMs, hidden Markov models, and neural networks. They need labeled attack data and give no guarantee when the attack changes. Our alarm needs no labels. Its false-alarm rate is set by benign predictability, which can be calibrated on clean traffic.

Research Program

We will build a block-level CV-QKD simulator using the homodyne saturation model of Wang et al. at fiber distances up to their 40 km. It will add three components. First, the attenuator will have a finite response time and quantized settings, which removes the ideal-device assumption. Second, a tap photodiode will sit behind a fused-fiber coupler whose wavelength response is interpolated through the Li et al. measurements. Third, benign drift in local-oscillator and signal power will be modeled as a first-order Markov process, following the channel models of the time-correlated communication paper.

On this simulator we will implement the optimistic controller. Its step size will adapt to the observed residuals, so no prior bound on hint quality is needed. We will also implement a sequential CUSUM-style test on the normalized residual, and calibrate its threshold on attack-free runs to a target false-abort rate.

The analysis has two parts. The first bounds saturation exposure, meaning the number of pulses that reach the detector above its linear limit, in terms of the residual sum and the attenuator's response delay. This follows from the regret bound plus the hinge structure of the loss. The second characterizes the attacks that leave the residual small. We will argue, and then check empirically, that these are exactly the attacks the controller absorbs, which leaves the standard key-rate check sound.

Baselines

The naive reference point is a fixed attenuator set conservatively for the worst expected power. It cannot be dragged by Eve but pays an SNR penalty at all times. We tune its single setting on clean traffic so that it is allowed the same benign false-abort rate as the other methods.

The strongest existing method on this question is the block-feedback attenuator of Defending Against the Homodyne Detector-Blinding Attack on Continuous-Variable Quantum Key Distribution Using an Adjustable Optical Attenuator. We reimplement it on the same simulator, with the same attenuator latency and quantization, block size, and finite-size key-rate computation. That paper reports that unmitigated strong blinding drives the computed key rate negative in simulations up to 40 km. It does not report saturation exposure or key leaked under time-varying attacks; no reported result was found on Lacuna for those metrics. Its numbers in our tables therefore come from our own reimplementation.

We also run plain Online Gradient Descent on the same loss, and a pure feedforward controller that sets attenuation directly from the tap. The ablation that isolates the key ingredient is optimistic OMD with the hint set to the previous block's gradient, $M_t = g_{t-1}$. This keeps optimism but discards the separate fast sensor. A second ablation keeps the tap but removes the residual alarm.

All controllers receive the same tuning budget, a grid search over step size or gains on attack-free runs, and are scored on identical seeds and attack traces. We count the proposal as beating the block-feedback baseline only if it does two things together. It must cut leaked insecure key under fast honest-tap attacks by a clear margin across seeds. It must also keep a benign key rate within a few percent of the best baseline at matched false-abort rate.

Why not just…?

Why not just feed the tap reading forward to the attenuator? Because Eve chooses the wavelength. A coupler that sends 0.3% of 1290 nm light to one port, as Li et al. measured, lets her blind the detector while the tap reports little power. In that case a feedforward controller follows the false reading exactly. Our spoofed-tap experiment exists to show this failure, and to show that the optimistic controller degrades only to reactive behaviour.

Why not just use a Kalman filter to fuse tap and block estimates? This is the real threat. Under benign Gaussian drift a well-tuned Kalman filter may match or beat optimistic OMD. Its weakness is that it trusts its noise model, and an adversarial tap breaks that model. We include it as a baseline with its noise parameters tuned on clean data. If it matches our method under the spoofed-tap and adaptive-attacker conditions, the contribution shrinks to the residual alarm and the exposure analysis, and we will report that.

Why not just abort whenever the tap shows high power? A threshold monitor gains nothing in benign key rate, and the wavelength-hidden blinding attack defeats it in the same way it defeats feedforward. The underwater study likewise shows that monitoring the local oscillator can be bypassed with combined attacks.

Why not just switch to measurement-device-independent QKD? Measurement-device-independent QKD removes detector side channels by sending the measurement to an untrusted relay. The cost is throughput: Long-Distance Free-Space Measurement-Device-Independent Quantum Key Distribution reports 6.11 bps over 19.2 km and needed adaptive optics and picosecond synchronization. We regard it as the right answer where that rate is acceptable, and target direct CV-QKD links where it is not.

Decisive Experiments

Four tests settle the claim. The first is a fast square-wave blinding attack with an honest tap. Here the proposal should keep saturation exposure well below the block-feedback baseline, because the hint arrives before the block does.

The second is the same attack with the tap spoofed using 1290 nm light. Here the residual alarm should fire within a few blocks, and exposure should be no worse than plain Online Gradient Descent.

The third is a slow ramp attack. Here the residual should stay quiet, so the alarm cannot be what saves the system. Instead, the excess-noise estimate should remain unbiased because the detector stays linear. We will check this directly against the true excess noise in the simulator. If the estimate is biased, the dichotomy claim fails.

The fourth pits the controller against an adaptive Eve who runs her own no-regret learner over pulse power to maximize leaked key. We measure whether reactive controllers oscillate while optimistic ones settle.

The primary metric is leaked insecure key, defined as bits accepted while the true key rate is negative. Secondary metrics are benign secret key rate, false-abort rate, and detection delay. All are reported with confidence intervals over seeds and across a range of attenuator latencies.