Every public-key system deployed today rests on a conjecture. The security of RSA reduces to factoring. Elliptic-curve Diffie-Hellman reduces to discrete logarithms. The lattice schemes now replacing both rest on problems like Module-LWE. None of these hardness assumptions is proven. Decades of failed cryptanalysis count as evidence, but they are not a proof. Shor's algorithm has already shown that a capable enough adversary can collapse an assumption that once looked permanent.
Quantum key distribution builds on a different foundation. It does not limit the adversary's computational power. It limits what any adversary can physically do. If quantum mechanics is correct, measuring an unknown state disturbs it, and an unknown state cannot be copied. An eavesdropper on the channel therefore leaves statistical fingerprints. The legitimate parties can measure how much she could have learned. They can then distill a key about which she knows essentially nothing.
The phrase unconditional security carries more weight than it first appears to. QKD security proofs are still conditional. They depend on the correctness of physics, on an authenticated classical channel, and above all on the devices behaving as the model says. This article covers three layers of that story: the information-theoretic core of BB84, the device-independent effort to remove trust in hardware entirely, and the implementation gap where most real attacks have actually succeeded.
BB84: Eavesdropping Has a Price
Bennett and Brassard's 1984 protocol is strikingly simple. Alice sends single photons. Each is prepared in one of four states drawn from two conjugate bases, rectilinear and diagonal. Bob measures each photon in a randomly chosen basis. Afterward they publicly compare which bases they used, but not the bit values. They discard the roughly half of rounds where their bases differed. What remains is the sifted key: two bit strings that would be identical if there were no noise and no eavesdropper.
The security intuition rests on the no-cloning theorem of Wootters, Zurek, and Dieks. No physical operation can map an arbitrary unknown state |ψ⟩ to |ψ⟩|ψ⟩. So Eve cannot copy Alice's photon, measure the copy, and forward the original untouched. The precise statement is an information-disturbance tradeoff. Any interaction that gives Eve information about a bit in one basis must introduce errors that show up in the conjugate basis. Eavesdropping is not forbidden, but it is costly. Eve pays in the quantum bit error rate, or QBER.
Turning that intuition into a proof took more than a decade. Mayers produced the first general proof, and Lo and Chau gave another that assumed quantum computers. In 2000, Shor and Preskill showed that BB84 is equivalent to an entanglement-distillation protocol built on CSS codes. That yields the asymptotic key rate r = 1 − 2h(Q), where h is the binary entropy function. The rate falls to zero near Q ≈ 11%. Renner later made the security composable. In his framework, the final key is ε-close in trace distance to a perfectly uniform key that is uncorrelated with Eve.
Information reconciliation is where theory turns into careful accounting. The sifted strings differ because of channel noise and possible tampering. Alice and Bob correct the differences with an error-correction protocol run over a public channel. Cascade was the historical choice; modern systems use LDPC or polar codes. Every syndrome bit they disclose is leakage that must be counted against them, as if Eve learned it. Practical codes fall short of the Shannon limit by an efficiency factor f > 1. The realistic rate therefore becomes 1 − f·h(Q) − h(Q), so inefficient reconciliation directly eats into the key.
Privacy amplification then compresses the reconciled string using two-universal hashing. The quantum leftover hash lemma guarantees that the output is close to uniform, even given Eve's quantum side information. The condition is that the output must be shorter than her smooth min-entropy. One subtlety is often overlooked. The classical channel must be authenticated, usually with Wegman-Carter MACs keyed from a pre-shared secret. QKD therefore expands an existing shared key rather than creating one from nothing.
TakeawayBB84 does not prevent eavesdropping; it puts a price on it. Its security comes from counting every possible leak, whether quantum disturbance or public disclosure, as a measured loss of secrecy that hashing can then remove.
Device-Independent QKD: Trusting Correlations Instead of Hardware
Standard BB84 proofs make two quiet assumptions. Alice's source emits exactly the specified states, and Bob's detectors measure exactly the specified bases. Both assumptions are fragile, which means the device model is effectively a hidden security parameter. Device-independent QKD asks a stronger question. Can two parties prove they share a secret key while treating their quantum devices as black boxes? Those boxes might even be built by the adversary. They simply accept classical inputs and return classical outputs.
The answer comes from Bell nonlocality, an idea already present in Ekert's 1991 entanglement-based protocol. Alice and Bob feed random settings into their boxes and estimate the CHSH value S. Any local hidden-variable model satisfies S ≤ 2. Such a model covers any strategy where outcomes are fixed in advance by shared classical data. Quantum mechanics allows values up to Tsirelson's bound of 2√2. A violation therefore proves the outcomes were not predetermined. Nonlocal correlations are also monogamous: strong correlation between Alice and Bob limits how strongly Eve can be correlated with either.
In 2007, Acín and collaborators quantified this against collective attacks. They bounded Eve's Holevo information as a function of S alone. A full proof against coherent attacks needed one more ingredient: the entropy accumulation theorem of Arnon-Friedman, Dupuis, Fawzi, Renner, and Vidick. The theorem shows that smooth min-entropy grows roughly linearly with the number of rounds, which makes finite-key analysis practical. The remaining assumptions are short: quantum mechanics holds, the measurement settings are chosen with private randomness, the labs are isolated, and classical processing is trusted.
Experiments arrived in 2022 with three demonstrations. Trapped ions in Oxford spanned about two meters. Neutral atoms in Munich spanned roughly 400 meters. A photonic system at USTC ran over 220 meters of fiber. These were landmarks, and also sobering. Key rates were tiny. Closing the detection loophole requires very high end-to-end detection efficiency. Without it, an untrusted box could cheat by selectively failing to respond, and that failure would look like ordinary photon loss. Over longer distances, loss grows quickly and the requirement becomes brutal.
The black-box model also has remaining weaknesses. Barrett, Colbeck, and Kent identified memory attacks. A device reused across sessions could store earlier outputs and leak them through later public communication. The isolation assumptions are themselves physical claims that must hold in practice. Device independence does not eliminate trust. It moves trust away from detailed hardware specifications and onto a short list of structural conditions. That shift makes the security argument much easier to audit.
TakeawayA Bell violation certifies that the measurement outcomes could not have been fixed in advance. Device independence turns security from a claim about hardware into a claim about observed statistics, but those statistics are only as trustworthy as the isolation that produced them.
Where Physics Meets Engineering
Most successful attacks on QKD never broke a security proof. They exploited the gap between the abstract model and the real hardware. An early example is the photon-number-splitting attack. Practical systems use attenuated lasers instead of true single-photon sources, so some pulses carry two or more photons. Eve can split off one photon, store it, and measure it after the bases are announced, without introducing any errors. Decoy-state methods closed this gap. Hwang proposed them, and Lo, Ma, and Chen, along with Wang, developed them further. Alice randomly varies the pulse intensity, which lets the parties bound how much key comes from multi-photon pulses.
Detectors turned out to be an even bigger problem. In 2010, Lydersen and colleagues in Makarov's group attacked commercial systems with bright illumination. The light blinded the avalanche photodiodes and forced them into a linear mode. Eve could then control which detector clicked by sending tailored pulses. She obtained the full key without raising the QBER. Other attacks followed. Time-shift attacks exploited small timing differences in detector efficiency. Trojan-horse attacks shined light into Alice's modulators and read the reflections. Each attack was patched, and each exposed an assumption nobody had written down.
Measurement-device-independent QKD, proposed by Lo, Curty, and Qi in 2012, removed the detector problem by design. Alice and Bob both send states to an untrusted relay. The relay performs a Bell-state measurement and announces the results. The parties never rely on how the detectors behave, so every detector side channel is closed by construction. Twin-field QKD extended the idea. Its key rate scales with √η, where η is channel transmittance. That lets it beat the PLOB bound, the limit for point-to-point QKD without quantum repeaters. The sources, however, must still be trusted and carefully characterized.
Then there are problems at the system level. At realistic block sizes, finite-key effects dominate, and errors in estimating channel parameters feed directly into the security parameter ε. Long-distance networks such as the Beijing-Shanghai backbone rely on trusted relay nodes. Keys exist in the clear at each node, which gives up the end-to-end guarantee QKD is meant to provide. National agencies including the NSA, NCSC, ANSSI, and BSI have publicly favored post-quantum cryptography for most uses. They cite QKD's dependence on special hardware, its cost, its need for separate authentication, and the difficulty of validating it.
None of this invalidates QKD. It clarifies what a proof actually delivers. A security proof is a statement about a model. The overall security claim is only as strong as the weakest match between that model and the real device. Quantum hacking has grown into a discipline that systematically tests those matches. Certification efforts such as ETSI's protection profiles try to formalize them.
TakeawayAn unconditional proof is only as strong as the match between its model and the real hardware. In physical cryptography, the attack surface is the set of assumptions nobody wrote down.
QKD offers something genuinely unusual: secrecy whose guarantee does not expire when algorithms improve. When the key is used for one-time-pad encryption, an adversary who records traffic today gains nothing from computational advances tomorrow. There is no hard problem left for her to solve later.
Yet the history of the field shows that physics-based security does not escape assumptions. It changes their kind. Computational cryptography assumes certain problems are hard. QKD assumes faithful devices, authenticated channels, and isolated laboratories. Device-independent and measurement-device-independent protocols are disciplined efforts to make that list shorter and explicit. They accept lower rates and shorter distances in exchange for fewer trusted components.
The principle worth carrying forward applies well beyond quantum channels. Security is not a property of a protocol alone. It belongs to the full chain from mathematical model to physical device. Often the most valuable design work is not strengthening the proof. It is reducing the number of things the proof needs you to believe.