CV Detector-Control Attack
The threat actor (Eve) injects bright light into the signal port of a CV-QKD receiver (Bob) to saturate the detector electronics, masking other attacks and gaining complete knowledge of the key.
The injected light is not cancelled out, and the reason is how the receiver has to be built. Its detector balances two photodiodes against the strong reference beam it measures with, and balancing against that beam means it cannot also be balanced against the port the signal arrives on [Qin2018]. Light sent in on that port therefore shows up as a large offset, pushes the detector past the range in which its reading still tracks the input, and the readings are clipped.
Clipped readings make the link look quieter than it really is, so the safety check the transmitter and the receiver run comes out clean while the threat actor is reading every bit. They finish the protocol and accept a key she already holds in full [Qin2018].
The same route has been run against a working CV-QKD system and judged the cheaper of the two ways into the receiver's detector, and the light-injection step has since been performed against a complete CV-QKD receiver rather than a bench detector, with no transmitter present and the channel transmission assumed [Kumar2021,Pereira2026].
| [Gerhardt2011] | I. Gerhardt, Q. Liu, A. Lamas-Linares, J. Skaar, C. Kurtsiefer, and V. Makarov. "Full-field implementation of a perfect eavesdropper on a quantum cryptography system" In: Nature Communications 2, 349. (2011) 10.1038/ncomms1348. |
| [ISOIEC23837P2] | ISO/IEC JTC 1/SC 27. "ISO/IEC DIS 23837-2:2022 Information security - Security requirements, test and evaluation methods for quantum key distribution - Part 2: Evaluation and testing methods". (2022) |
| [Kumar2021] | Rupesh Kumar, Francesco Mazzoncini, Hao Qin, and Romain Alleaum. "Experimental Vulnerability Analysis of QKD Based on Attack Ratings" In: Scientific Reports 11, 9564. (2021) 10.1038/s41598-021-87574-4. |
| [Mao2020a] | Y. Mao, W. Huang, H. Zhong, Y. Wang, H. Qin, Y. Guo, and D. Huang. "Detecting quantum attacks: a machine learning based defense strategy for practical continuous-variable quantum key distribution" In: New J. Phys. 22, 083073. (2020) 10.1088/1367-2630/aba8d4. |
| [Pereira2026] | Daniel Pereira, Vana Pezelj, Florian Prawits, Hannes Hübel. "Experimental demonstration of a coherent detector blinding attack on a real CV-QKD system". (2026) arXiv:2605.03572. https://arxiv.org/abs/2605.03572. |
| [Qin2016] | H. Qin, R. Kumar, and R. Allaume. "Quantum hacking: Saturation attack on practical continuous-variable quantum key distribution" In: Phys. Rev. A 94, 012325. (2016) 10.1103/PhysRevA.94.012325. |
| [Qin2018] | H. Qin, R. Kumar, V. Makarov, and R. Alleaume. "Homodyne-detector-blinding attack in continuous-variable quantum key distribution" In: Phys. Rev. A 98, 012312. (2018) 10.1103/PhysRevA.98.012312. |
| [Xu2022] | S. Xu, Y. Li, Y. Mao, and Y. Guo. "Counteracting a Saturation Attack in Continuous-Variable Quantum Key Distribution Using an Adjustable Optical Filter Embedded in Homodyne Detector" In: Entropy 24, 383. (2022) 10.3390/e24030383. |
b9a7831b-d5a2-4117-b101-41c5a7f58ffd
- Quantum Communication
Technique properties
Operational properties
| Description | Countermeasure |
|---|---|
|
An adjustable attenuation element in the receiver's signal path, driven by feedback from the detection response, keeps the arriving quadratures inside the linear range so the clipping does not distort the estimate. The published proposal embeds an adjustable element in the homodyne detector that compensates for the potential saturation caused by strong received optical power in real time, based on the feedback of the response of detection, and it is analysed against this same clipped-response model, with the saturation induced by displacing the quadratures rather than by injected light [Xu2022]. The loop is block-by-block: the attenuation derived from one data block is applied to the next, and a block that needs no attenuation is the one used to estimate the excess noise [Xu2022]. What the control delivers is a truthful estimate rather than a blocked attack, which is how its authors frame the result: with the attenuation applied the estimated excess noises fall back inside the finite linearity domain, the secret key rate goes negative, and the legal participants can detect the threat actor performing the saturation attack [Xu2022]. The attack still arrives; what stops is its ability to hide. It performs no spectral, spatial or polarisation selection, which is what makes it a variable optical attenuator rather than a filter. The two controls are complementary here: a filter rejects light that is out of band, an attenuator keeps the detector honest about light that is in band. |
CM-0099 Variable Optical Attenuator |
|
Monitoring the receiver's own detector output and electrical parameters is the control the primary assesses as the alternative to a watchdog. The primary's published observable: instead of measuring the difference between the two photocurrents, the receiver may monitor one, or the sum of the two, from its two photodiodes [Qin2018]. The primary's own verdict on the photocurrent route is sceptical, and the reasons belong here. It may be challenging in practice, because it will disturb the detector output and it can be difficult to set a proper discrimination level to detect a blinding attack correctly, and the discrete-variable analogue was later proven insufficient in many particular cases [Qin2018]. The quantitative reason is this attack's small footprint: the threat actor only needs to increase the overall energy by about 12 percent, so local-oscillator intensity fluctuations from the transmitter's laser and the channel environment may exceed the external light's energy and cause false alarms [Qin2018]. Adding electronics also costs performance, since a high-gain, high-bandwidth, high-efficiency, low-noise homodyne detector is already technically challenging and extra components increase electronic noise [Qin2018]. |
CM-0017 Detector Monitoring |
|
A spectral filter at the receiver's optical input attacks this route's defining property, because being out of band is how the injected light stays separate from the local oscillator in the first place. The primary has the threat actor send the external light in a different mode from the local-oscillator pulses, in practice at a different wavelength, and notes that the two photodiodes are classical detectors with large wavelength ranges, typically 800 to 1700 nm, so any light in that range produces photocurrent and makes it impossible for the parties to distinguish the source of the light by measuring the detector output alone [Qin2018]. A filter ahead of the detector removes that freedom. The 2026 experiment endorses the control by name with the condition that matters, that optical filters and watchdog detectors remain valid options provided their optical bandwidth is wide enough to cover the relevant frequency range, and its own source sat far outside the signal band but well inside the photodiodes' response [Pereira2026]. The spectral subtype is what applies. The spatial child CM-0045 does not reach this attack, because the injected light arrives through the same single-mode fibre as the signal and so shares its spatial mode. |
CM-0013 Employment of Optical Filters |
|
A watchdog photodiode at the receiver's optical input raises an alarm on light the link should not be receiving, and it is the first countermeasure the primary proposes against this attack: a straightforward countermeasure is to monitor the light intensity that is going into the signal port of the homodyne detector, which the receiver can implement using a sensitive p-i-n photodiode in order to detect any strong light impinging on the signal port, and which can also serve the energy test required by some security proofs [Qin2018]. Three limitations are published and all three bear on this attack. It is challenging to build a detection system that can in practice detect light in any optical mode the threat actor may try to use [Qin2018], a point the predecessor puts more sharply, that it is in general not possible to predict in which mode an attacker will try to send energy in order to saturate the detector, so a watchdog capable of detecting any such attempt is very difficult to design [Qin2016]. A practical watchdog may not always be able to raise the alarm [Qin2018]. And the threat actor may use a high-power laser to damage the photodiode watchdog and so bypass the security alarm of the protocol [Qin2018]. The 2026 experiment endorses the control by name with the condition that decides whether it works here: watchdog detectors remain valid options provided their optical bandwidth is wide enough to cover the relevant frequency range [Pereira2026]. That condition bites on this route specifically, because the injected light sits at a wavelength away from the local oscillator and the detector's own photodiodes respond across roughly 800 to 1700 nm [Qin2018], so a watchdog scoped to the signal band alone can miss light the detector still sees. |
CM-0011 Watchdog Detector |
|
Acquiring quadrature measurements at randomly chosen optical attenuations tests the linearity the security proof assumes, and clipping destroys that linearity. The primary describes the control as the parties testing the linearity between the noise and the signal measurement using an active attenuation device on the receiver's side, that is an amplitude modulator, and states why it reaches this attack: the randomisation of the signal port's attenuation prevents the threat actor from properly setting the intensity of the blinding pulses [Qin2018]. The predecessor reaches the same verdict independently and calls it precisely the countermeasure that constitutes an effective parade against the saturation attack [Qin2016]. One configuration detail decides whether it works at all. More than two attenuation values are required, and the linearity of the total noise against attenuation has to actually be checked, because the usual two-point scheme shuts the signal off to measure shot noise and blocks the saturated signal along with it, leaving that measure unaltered so a simple excess-noise difference can be defeated [Qin2016]. The base weakness V-0102 carries the published acceptance figure for the test, a squared correlation coefficient of 0.99 against 0.386 observed under saturation. |
CM-0087 Real-time Quadrature Measurements with Random Optical Attenuations |
|
A neural network trained on pulse-level features detects this attack without being told to look for it, and the control names this attack in its own detected set rather than being adapted to it: the security of practical Gaussian-modulated coherent-state CV-QKD can be broken by practical attack strategies such as Trojan-horse attacks, wavelength attacks, calibration attacks, local-oscillator intensity attacks, saturation attacks and homodyne-detector-blinding attacks [Mao2020a]. The scheme derives a feature vector from several features of the pulses and feeds it to an artificial neural network, and reports detecting most known attacks at the cost of a small part of the secret key and transmission distance, establishing a universal attack detection model by monitoring a few pulse features without knowing the attack type in advance [Mao2020a]. Its motivation is also a fair criticism of the per-attack monitors this row's other edges point at: existing countermeasures exploit different real-time monitoring modules for different attacks, which depend significantly on the accuracy of the estimated excess noise and lack a universal defence [Mao2020a]. That dependency is exactly what this attack corrupts, which is why a control that looks at the pulses rather than at the derived excess noise is worth linking. Reported performance is strong but simulated: the trained network identifies and classifies attacks with precision and recall values above 99 percent, from generated data rather than from a live link, and the result is sensitive to the number of hidden-layer neurons [Mao2020a]. The standing cost is stated plainly: 10 percent of pulses are spent on real-time shot-noise estimation and the amplitude modulator the scheme puts in the receiver's signal path adds a typical 2.7 dB of insertion loss, which costs both key rate and distance [Mao2020a]. Scope also matters for a classifier: the trained set is the calibration attack, the local-oscillator intensity attack, the saturation attack and two hybrids, all against Gaussian-modulated CV-QKD with homodyne detection, which is this row's own configuration [Mao2020a]. |
CM-0088 Pattern Recognition Using Machine Learning |
|
The control the primary itself recommends. The receiver sets security thresholds inside its detector's measured limits, the parties estimate for each block the fraction of measurements recorded outside them, and if too large a fraction falls beyond they know the detector was not working in its linear range for a non-negligible part of the quantum phase and discard the block [Qin2018]. Clipped data therefore never reaches parameter estimation and the bias cannot form. The obvious implementation is the wrong one. A per-point confidence-interval post-selection would give the threat actor the power to influence which data is post-selected just by controlling the displacement, leaving the retained data non-Gaussian with no security guarantee; the safe form discards whole blocks, with the interval typically wider than six standard deviations so the per-event saturation probability is below [Qin2016]. A refined Gaussian post-selection keeps some key under moderate saturation, retaining 15.37 percent of points in simulation [Qin2016]. ISO/IEC 23837-2 clause 8.10 raises this from a published proposal to the pass condition of a drafted international test standard: the receiver module under test must report an exceptional event about homodyne detector saturation and abandon the whole block of data recorded partially or totally outside its stated detection limits [ISOIEC23837P2]. |
CM-0089 Discarding Non-linear Data Blocks |
| Description | Category | Impact Type |
|---|---|---|
|
Electronics and analog components is the layer the attack primarily hits, and the primary is unusually specific about it. What saturates is the detection electronics: a bright light pulse sent on the signal port can lead to a saturation of the detector electronics [Qin2018]. The limits that gate the attack are voltages and currents rather than optical powers, set in the primary's own detector by the digitiser range rather than by the amplifier [Qin2018], and the photodiodes are explicitly ruled out as the saturating element because their limit is a few milliwatts, far above the system's total optical power [Qin2018]. The 2026 experiment identifies the saturating stage on its own receiver as the analog-to-digital converter [Pereira2026]. |
L2 — Electronics & Analog Components | Primary Target |
|
Physical quantum and optical is the layer the attack arrives through. The saturating light is delivered into the receiver's signal port as bright optical pulses whose width, repetition rate and polarisation are matched to the transmitter's signals, at a wavelength chosen both to stay out of the local oscillator's mode and to tune the splitting ratio the receiver's beam splitter gives it [Qin2018]. The optical route is also where the exploited imbalance lives, since the beam-splitter reflection and transmittance coefficients are never perfectly balanced [Qin2018]. Secondary rather than primary because no optical component is itself defective or driven out of specification: the optics only carry the power that drives the electronics out of range. |
L1 — Physical Layer (Quantum & Optical) | Secondary Effect |
|
Post-processing and cryptographic is where the attack's effect is realised. The clipped readings are consumed by parameter estimation, so the excess-noise estimate comes out below the null-key threshold and the channel-transmission estimate comes out biased downwards, and the secret-key-rate decision is then computed from the wrong inputs: at 25 km the estimate falls from 2.5813 to 0.0367 against a threshold of 0.1013 on a change of one parameter [Qin2018]. Secondary because nothing at this layer is itself defective, the arithmetic is correct and only its inputs are corrupted. It matters for defence because the two most effective controls act here rather than optically, by discarding blocks recorded outside the detector's calibrated limits [Qin2018] and by testing the linearity of noise against randomised attenuation [Qin2016]. |
L5 — Post-Processing & Cryptographic | Secondary Effect |
| Relation type | Description | Technique |
|---|---|---|
| Alternative to |
This row holds the coherent-displacement route, in which the threat actor resends displaced coherent states that saturate the receiver's homodyne detector [Qin2016]. A-0102 CV Detector-Control Attack holds the incoherent-injection route, in which a separate laser shines into the receiver's signal port [Kumar2021]. [Kumar2021] treats the two as strategies of one saturation attack, and [Mao2020a] lists them as two separate attacks. [Kumar2021] attributes its own incoherent route to the paper behind A-0102 and shows the choice is economic rather than mechanical, Attack Potential 14 against 26, because an incoherent laser needs no phase-drift compensation [Kumar2021]. |
A-0098 Saturation Attack |
|