Quantum Trojan-Horse Attack (THA)

System

In a Quantum Trojan-Horse Attack (THA), a threat actor (Eve) learns the settings of a QKD system by reading light that leaves the apparatus through its own legitimate optical port, rather than by measuring the transmitted quantum states. Because she never touches those states, she induces none of the transmission errors that would disclose her, which is what makes the approach stealthy [Vakhitov2001].

The threat actor supplies that light herself, which is why the technique is also called the Large Pulse Attack. She shines bright probe pulses into the transmitter (Alice), or into the receiver (Bob) depending on the protocol, and analyses what the internal optical components reflect back; the transmitter is the certain target because that is where the key is encoded [Gisin2006]. The component that matters most is the modulator, since its setting is the key, and a system built without an internal optical modulator is immune for exactly that reason [Vakhitov2001].

A second mechanism reaches the same port without any injected light at all: the receiver's own single-photon detectors emit stray photons each time they fire, and an eavesdropper can tell from them which detector clicked [Pinheiro2018]. Vakhitov2001 places that mechanism in the same conventional-optical-eavesdropping family as this technique, alongside high-power destruction of optical components, but treats it as a separate method needing study in its own right. SQOUT carries it as its own technique, A-0080 Backflash Attack [Su2025a].

What the threat actor gains ranges from learning which bases were used, through narrowing the raw key to a few candidates, to reading the raw key outright [Vakhitov2001]. The wider lesson is that securing the quantum signals is not sufficient: the physical and optical components of the communication system, and the port they share with the channel, must be secured as well.

Literature 9
[Bondar2025] Klim D. Bondar, Ivan S. Sushchev, Daniil D. Bulavkin, Kirill E. Bugai, Anna S. Sidelnikova, Dmitriy M. Melkonian, Veronika M. Vakhrusheva, Dmitriy A. Dvoretskiy. "Broadband optical time-domain reflectometry for quantum key distribution security analysis" (2025) 10.1103/368v-t34p. arXiv:2507.00159.
[Gisin2006] N. Gisin, S. Fasel, B. Kraus, H. Zbinden, and G. Ribordy. "Trojan-horse attacks on quantum-key-distribution systems" In: Phys. Rev. A, vol. 73, no. 2, p. 022320. (2006) 10.1103/PhysRevA.73.022320.
[Jain2014] N. Jain, E. Anisimova, I. Khan, V. Makarov, C. Marquardt, and G. Leuchs. "Trojan-horse attacks threaten the security of practical quantum cryptography" In: New J. Phys., vol. 16, no. 12, p. 123030. (2014) 10.1088/1367-2630/16/12/123030.
[Jain2015] N. Jain, B. Stiller, I. Khan, V. Makarov, C. Marquardt, and G. Leuchs. "Risk Analysis of Trojan-Horse Attacks on Practical Quantum Key Distribution Systems" In: IEEE Journal of Selected Topics in Quantum Electronics, vol. 21, no. 3, pp. 168–177. (2015) 10.1109/JSTQE.2014.2365585.
[Navarrete2026] Á. Navarrete, M. Pereira, G. Currás-Lorenzo, A. Mizutani, M. Curty, K. Tamaki. "Phase-error estimation for quantum key distribution with leaky receivers". (2026) arXiv:2608.09674.
[Pinheiro2018] P. Pinheiro, P. Chaiwongkhot, S. Sajeed, R. Horn, J. Bourgoin, T. Jennewein, N. Ltkenhaus, and V. Makarov. "Eavesdropping and countermeasures for backflash side channel in quantum cryptography" In: Opt. Express 26, 21020-21032. (2018) 10.1364/OE.26.021020.
[Su2025a] Jialei Su, Junxuan Liu, Zihao Chen, Mingyang Zhong, Qingquan Peng, Jiangfang Ding, Yijun Wang, Anqi Huang, Ying Guo. "Quantifying the Upper Limit of Backflash Attack in Quantum Key Distribution" (2026) 10.1103/xn8h-961f. arXiv:2501.09984. https://arxiv.org/abs/2501.09984.
[Tan2025] Hao Tan, Mikhail Petrov, Weiyang Zhang, Liying Han, Sheng-Kai Liao, Vadim Makarov, Feihu Xu, Jian-Wei Pan. "Wide-Spectrum Security of Quantum Key Distribution" (2025) 10.1103/smdh-dcpm.
[Vakhitov2001] A. Vakhitov, V. Makarov, and D. R. Hjelme. "Large pulse attack as a method of conventional optical eavesdropping in quantum cryptography" In: Journal of Modern Optics, vol. 48, no. 13, pp. 2023–2038. (2001) 10.1080/09500340108240904.
QID: A-0049
UUID: 3cb972c2-9cb7-407f-9fd0-4add9f1b7c85
Tier: T0
Type: Quantum
Domain(s):
  • Quantum Communication
Tactic: Execution
Created: 2024-01-30
Updated: 2026-11-12
Maturity: Experimental
Technique properties

Attack type: Key extraction, Information obtaining
Attack category: Trojan-horse attack
Target party: Receiver, Transmitter
Target function: Quantum layer
Quantum channel: Fibre optic, Free space
Security layers: L1 — Physical Layer (Quantum & Optical) L5 — Post-Processing & Cryptographic
Maturity: Experimental
Operational properties

Technique → Countermeasures 7
Description Countermeasure

Wavelength filtering on the optical port prevents the active route by removing the attacker's main degree of freedom. Anti-reflection coatings are specified for particular wavelengths, so an off-band probe returns a much larger reflection [Vakhitov2001], and she is otherwise free to range across the whole channel passband exploiting spectral transparency windows [Tan2025]. Gisin2006 lists filters first among the required design measures.

CM-0013 Employment of Optical Filters

Detector monitoring can catch a bright-pulse probe at the receiver, because a probe strong enough to read an internal setting also disturbs the detectors it passes. Jain2014 mounted this attack against the receiver of a commercial Clavis2 and found that the probe pulses caused a high level of after-pulsing in the single-photon detectors, resulting in a large quantum bit error rate that effectively protected that system from the attack [Jain2014]. The control is nonetheless weak and unreliable, because the same paper engineers around the observable rather than abandoning the attack. That first result holds only for a naive attempt to read every slot, because probing at the receiver's full 5 MHz modulation rate is what would cause a tremendous amount of afterpulsing [Jain2014]. Jain2014 then devises a selective strategy in which the threat actor controls when the receiver clicks, groups her probes into bursts, and deliberately imposes a dead time from the legitimate photons to mask the afterpulsing a burst causes; under it, neither is the expected detection rate of Bob severely affected, nor does the QBER cross the abort threshold [Jain2014]. The positive-leakage result remains a numerical simulation for a Clavis2-like system equipped with detectors with less-noisy but realistic characteristics rather than a completed theft [Jain2014]. What a defender should take from this is that the error rate catches a careless probe and not a careful one. And the threat actor need not probe brightly at all, since she can interrogate Bob's site with signals of very low intensity, in which regime there is no excursion for the monitor to see [Vakhitov2001]. The control the active-route sources actually require at the receiver is a dedicated detector monitoring the light entering the port, which is CM-0011 [Gisin2006].

CM-0017 Detector Monitoring

Optical isolation on the transmitter attenuates the reverse path the back-reflected probe must travel, raising the probe power the attacker needs. Gisin2006 is dismissive of it as a standalone answer, calling the use of attenuators or isolators a first naive idea for circumventing the need for an auxiliary detector, on the ground that the threat actor is not limited by technology and could merely send in more intense light; an optical fuse to cap that intensity was considered and rejected, because no such fuse operates for ultrashort pulses [Gisin2006]. The quantitative case comes from Vakhitov2001 instead: an isolator contributes about 50 dB, which together with a doubly-traversed 60 dB output attenuator reaches roughly 170 dB in total, and the probe power needed to beat that exceeds the damage threshold of the fibre itself [Vakhitov2001]. Two scope limits decide how much that is worth here. The protection holds for a one-way attenuated-source transmitter, it is not applicable with a true single-photon source, and round-trip systems like plug and play do not allow the use of non-reciprocal devices such as optical isolators at all [Vakhitov2001]. The control is therefore architecture-dependent rather than uniformly weak: on a one-way attenuated-source transmitter it is close to total, because the probe power needed exceeds what the fibre survives, and on a round-trip system it cannot be installed at all. A third limit is spectral: isolators are specified for a band, and the threat actor can choose her probe wavelength anywhere in the channel passband [Tan2025].

CM-0008 Employment of Optical Isolation

Switching to a protocol whose transmitter carries no internal optical modulator eliminates the transmitter route architecturally. Vakhitov2001 states that Koashi-Imoto and EPR-based systems are intrinsically immune to the large pulse attack, because signals reflected from them cannot carry any information on the transmitted quantum states [Vakhitov2001].

CM-0006 Employing Another Quantum Protocol

An auxiliary watchdog detector monitoring incoming intensity detects this technique and is the non-optional control against it. Gisin2006 requires a per-qubit test and states that without such monitoring the QKD system is insecure irrespective of the quality of the source [Gisin2006]. The requirement reaches both endpoints rather than the transmitter alone: every sensitive apparatus, Alice for sure and Bob depending on the protocol, must have an active control on the intensity of the incoming light [Gisin2006]. Its limit at the receiver is set by its own noise floor, because the threat actor can interrogate Bob's site with signals of very low intensity rather than having to send a bright pulse [Vakhitov2001].

CM-0011 Watchdog Detector

Privacy amplification neutralises the leak by removing the attacker's information, provided the legitimate parties first hold a bound on it. Gisin2006 derives that bound from the monitoring detector's limit, whose accuracy determines how much amplification is needed: at a 0.1-photon cap the threat actor gains 0.135 bits per qubit without phase randomisation and 0.095 bits with it, and that is the fraction the amplification has to remove [Gisin2006]. Because the leak adds no transmission errors in the low-intensity regime, nothing in the link statistics indicates how much amplification is required, so the bound has to come from the monitoring and characterisation controls rather than from the observed error rate [Vakhitov2001]. The bound is also wavelength-dependent, since the attacker can probe where the components leak most [Jain2015].

CM-0001 Privacy amplification

The proof-based line neutralises the leak analytically, and it is the only countermeasure Gisin2006 derives rather than recommends: he computes the optimal information gain, shows global phase randomisation cuts it by about 1.44, and reduces the two-way plug-and-play case to the standard one-way security proof given phase randomisation, per-pulse intensity monitoring and sufficient attenuation [Gisin2006]. Navarrete2026 extends the same analytic approach to the receiver side, covering leakage through back-reflected Trojan-horse light under a finite-key proof that needs only a bound on the distinguishability of the side-channel states [Navarrete2026]. What the proof consumes is a measured bound on the light that can leave the apparatus, so its input is architecture-dependent: Vakhitov2001 computes roughly 170 dB of round-trip attenuation for a one-way attenuated-source transmitter, and states that round-trip systems like plug and play do not allow the use of non-reciprocal devices such as optical isolators at all, so the same proof is fed a far weaker bound there [Vakhitov2001].

CM-0014 Full Device Imperfections Incorporation in Scientific Security Proof
Technique → Security Layers 2
Description Category Impact Type

Physical Layer (Quantum & Optical) is the primary target: the technique exploits the optical response of internal components, chiefly the modulator whose setting carries the key, by injecting bright probe light and reading what the components reflect back. Gisin2006 states the basis as a general property of this layer, that every optical apparatus can be examined from the outside by shining into it well controlled light and analysing the backscattered light [Gisin2006], and Vakhitov2001 is the founding demonstration of reading a modulator setting that way [Vakhitov2001]. The layer's spectral extent is part of the exposure: component behaviour deviates far outside the signal band, and the threat actor may choose any wavelength the channel passes [Tan2025].

L1 — Physical Layer (Quantum & Optical) Primary Target

Post-Processing & Cryptographic layer is a secondary effect: the optical information leak compromises the distilled key, so privacy amplification (CM-0001, already linked) must absorb the leaked fraction. The size of that absorption is this layer's exposure, and Gisin2006 quantifies it: at a 0.1-photon monitoring limit the threat actor gains 0.135 bits per qubit without phase randomisation and 0.095 bits with it [Gisin2006]. The fraction is not a single number, because Jain2015 finds wavelength regimes where the attacker gains considerably more than at 1550 nm, so the bound fed into amplification has to be measured across the band rather than at the signal wavelength [Jain2015]. This layer also cannot derive the bound for itself, because the attack leaves the error rate untouched in the regime where the threat actor interrogates the apparatus with signals of very low intensity, so the figure has to arrive from the monitoring and characterisation controls rather than from post-processing statistics [Vakhitov2001].

L5 — Post-Processing & Cryptographic Secondary Effect
Techniques Relating to This One 4
Relation type Description Technique
Alternative to

Two ways to the same leak from the transmitter. This technique reads setting information that the light source puts into unused properties of its own light: passive generation of information leakage may be produced via mode dependencies in Alice's transmitter [Zapatero2024]. The Trojan-horse attack reaches a similar effect from outside: Eve could actively provoke it, for example by launching a Trojan-horse attack against the transmitter [Zapatero2024]. The two routes need different defences. Duplinskiy2021 notes that the modulator-free carrier of this attack is immune to conventional Trojan-horse attacks while still being insecure because of information leakage beyond the single-mode space [Duplinskiy2021].

A-0083 Side Channel Exploitation in Transmitter Imperfections
Alternative to

A-0049 Quantum Trojan-Horse Attack also reaches information through the legitimate optical port of the QKD equipment. The two rows divide the routes between them: A-0049's own scope section states that it covers the active route only and names this row as the carrier of the passive one. The grouping has a source: Vakhitov2001 treats large-pulse probing and light emission from avalanche photodiodes as two members of one conventional-optical-eavesdropping family, and Navarrete2026 covers receiver leakage through detector backflashes or back-reflected Trojan-horse light in a single clause [Navarrete2026]. What separates them is direction and consequence, and those stay distinct on the two rows: this one is cat04 with channel_action measure, channel_directionality backward alone, leakage_stage detector and outcome_class partial_key, while A-0049 is cat15 with inject, backward and bidirectional, source and full_silent_key. The vulnerability chains are separate too: this technique links V-0030 and V-0065, while A-0049 links V-0002, V-0031 and V-0033.

A-0080 Backflash Attack
Combines with

[Huang2019] section V states the two can be run at the same time: the threat actor (Eve) reads back-reflected light for the transmitter's modulator settings while her injected light modifies the source. They are distinct mechanisms and the same source is careful to separate them, noting that unlike a Trojan-horse attack, which analyses light reflected back out, laser seeding manipulates the functioning of the transmitter's laser diode directly. The concurrency is what makes the pair worth recording, since one attack reads a setting and the other changes the state.

A-0063 Laser Seeding Attack (LSA)
Preceded by

A-0053 is the step before a Trojan-horse attack against a transmitter whose isolation would otherwise defeat it. BSI names the target of this technique as enabling other attacks and points at the Trojan-horse table; Tan2022 quantifies the result, a 236 dB three-isolator transmitter going from leak-free at 281 km to no secure key above 678.55 Gs [Tan2022].

A-0053 External Magnetic Field Attack on Optical Isolators and Circulators
See Also
A-0047 Denial
A-0048 Destruction
A-0053 External Magnetic Field Attack on Optical Isolators and Circulators
A-0054 Phase-Remapping Attack
A-0055 Non-Random Phase Attack