Simple Algorithm
Quantum Computing · Rydberg State · 28 min

Device-Independent QKD

Security from Bell violations, trusting no hardware.

Every BB84 proof assumes that the devices do what the model says: this photon is a qubit, that detector measures ZZ. Hackers have repeatedly broken commercial systems by exploiting the gap. Device-independent QKD closes it from the other side: treat the devices as black boxes, possibly built by Eve, and certify the key from the observed CHSH violation alone. In 2022 it was demonstrated for the first time.

What you’ll learn

  • Formulate the device-independent setting and list the assumptions that remain.
  • State and sketch the proof of the Acín et al. key-rate bound r≥1−h(Q)−h(1+(S/2)2−12)r\ge1-h(Q)-h\big(\tfrac{1+\sqrt{(S/2)^2-1}}{2}\big) for collective attacks.
  • Explain how the entropy accumulation theorem extends security to general attacks with finite statistics.
  • Compare DIQKD's noise and detection-efficiency requirements with BB84, and summarize the 2022 experiments and open problems.

Why trust nothing?

A security proof is a theorem about a model. Real QKD systems have been broken without contradicting any theorem, by violating the model: bright light can blind single-photon detectors and turn them into classical detectors that Eve controls (Lydersen et al., 2010), detection-time mismatches can leak basis information, and probe light injected into Alice's device can read out her settings. Each was patched, but each patch is another model assumption.

ApproachTrustedUntrusted
Prepare-and-measure (BB84 with decoys)Characterized source and detectorsThe channel
Measurement-device-independent (Lo, Curty & Qi, 2012)Alice's and Bob's sourcesAll detectors: an untrusted relay performs Bell measurements
One-sided device-independentOne party's measurement deviceThe other party's device (security from steering)
Device-independentQuantum theory, isolated labs, local randomness, classical processingEvery quantum device, including the source

From CHSH to secrecy

In the standard protocol (Acín, Brunner, Gisin, Massar, Pironio & Scarani, 2007), one party has an extra setting used only for key. Say Alice chooses x∈{0,1}x\in\{0,1\} and Bob y∈{0,1,2}y\in\{0,1,2\}. Rounds with x,y∈{0,1}x,y\in\{0,1\} estimate the CHSH value

S=⟨A0B0⟩+⟨A0B1⟩+⟨A1B0⟩−⟨A1B1⟩,S=\expval{A_0B_0}+\expval{A_0B_1}+\expval{A_1B_0}-\expval{A_1B_1},

and rounds (x,y)=(0,2)(x,y)=(0,2) produce the raw key, with error rate Q=Pr⁡[a≠b∣x=0,y=2]Q=\Pr[a\neq b\mid x=0,y=2]. Honest devices share ∣Φ+⟩\ket{\Phi^+} and measure A0=ZA_0=Z, A1=XA_1=X, B0,1=(Z±X)/2B_{0,1}=(Z\pm X)/\sqrt2, B2=ZB_2=Z, giving S=22S=2\sqrt2 and Q=0Q=0.

DerivationProof sketch
  1. Reduce to qubits. Each box's binary measurements can be taken projective (enlarge the space, by Naimark). By Jordan's lemma, two binary projective observables on any Hilbert space decompose simultaneously into blocks of dimension at most 2: Ax=⨁αAxαA_x=\bigoplus_\alpha A_x^\alpha, and similarly for Bob. Giving Eve the block labels α,β\alpha,\beta only helps her, so the problem becomes a mixture of qubit-pair problems.

  2. Symmetrize. Within a block, local rotations put the observables in the XX–ZZ plane, and a symmetrization step (random correlated bit flips by Alice and Bob, which leave SS and QQ unchanged and cannot help Eve) makes the state Bell-diagonal with eigenvalues λΦ±,λΨ±\lambda_{\Phi^\pm},\lambda_{\Psi^\pm}.

  3. S ≤ 22 (λΦ+−λΨ−)2+(λΦ−−λΨ+)2S\ \le\ 2\sqrt2\,\sqrt{(\lambda_{\Phi^+}-\lambda_{\Psi^-})^2+(\lambda_{\Phi^-}-\lambda_{\Psi^+})^2}

    The Horodecki criterion: the best CHSH value of a two-qubit state is 2t12+t222\sqrt{t_1^2+t_2^2}, with t1,t2t_1,t_2 the two largest singular values of its correlation matrix; here they are TxxT_{xx} and TzzT_{zz}.

  4. χ(A0:E)=H(λ⃗)−h(λΦ++λΦ−)\chi(A_0{:}E)=H(\vec\lambda)-h\big(\lambda_{\Phi^+}+\lambda_{\Phi^-}\big)

    Eve holds the purification. Her total state has entropy H(λ⃗)H(\vec\lambda), and conditioned on Alice's ZZ outcome it has entropy h(λΦ++λΦ−)h(\lambda_{\Phi^+}+\lambda_{\Phi^-}).

  5. Optimize and average. Maximizing χ\chi over Bell-diagonal states with a given SS gives F(S)F(S). Since FF is concave and decreasing, mixing blocks with different SαS_\alpha gives Eve at most F(Sˉ)F(\bar S) (Jensen). Finally Devetak–Winter gives r≥H(A0∣E)−H(A0∣B2)≥1−F(S)−h(Q)r\ge H(A_0|E)-H(A_0|B_2)\ge1-F(S)-h(Q). ■\blacksquare

Predict before you look

A DIQKD run observes S=2.7S=2.7 and Q=2%Q=2\%. Estimate the asymptotic key rate per key round, r=1−h(Q)−F(S)r=1-h(Q)-F(S).

Show the answer

0.59

(S/2)2−1=0.8225(S/2)^2-1=0.8225, so 1+0.82252≈0.953\tfrac{1+\sqrt{0.8225}}{2}\approx0.953 and F(S)=h(0.953)≈0.272F(S)=h(0.953)\approx0.272. With h(0.02)≈0.141h(0.02)\approx0.141, r≈1−0.141−0.272≈0.59r\approx1-0.141-0.272\approx0.59. For a depolarized Φ+\Phi^+ (so S=22(1−2Q)S=2\sqrt2(1-2Q)) the rate vanishes at Q≈7.1%Q\approx7.1\%, against 11% for BB84 with trusted devices: distrust has a price.

General attacks: entropy accumulation

Collective attacks assume the boxes behave identically and independently in every round, which is exactly what untrusted, memory-equipped devices need not do, and de Finetti-type symmetrization does not apply to them. Vazirani & Vidick (2014) gave the first proof against fully general attacks. The entropy accumulation theorem (EAT) then yielded essentially optimal rates (Arnon-Friedman, Dupuis, Fawzi, Renner & Vidick, 2018).

The first-order term is the von Neumann rate of the i.i.d. analysis, so coherent attacks cost nothing asymptotically. That matters, because the naive alternative, bounding each round's guessing probability, is much weaker. Pironio et al. (2010) showed Pguess(A∣E)≤12+122−S2/4P_{\rm guess}(A|E)\le\tfrac12+\tfrac12\sqrt{2-S^2/4}, i.e. Hmin⁡≥1−log⁡2(1+2−S2/4)H_{\min}\ge1-\log_2\big(1+\sqrt{2-S^2/4}\big): about 0.490.49 bits per round at S=2.7S=2.7, against 1−F(2.7)≈0.731-F(2.7)\approx0.73.

Check your understanding

Why is the entropy accumulation theorem needed for DIQKD, rather than the de Finetti or post-selection arguments used for BB84?

  1. Because DI devices measure in more than two bases
  2. Because untrusted devices with memory need not act identically and independently across rounds, and the protocol has no permutation symmetry to exploit; EAT handles sequential, memory-dependent rounds directly
  3. Because the CHSH value cannot be estimated from finitely many rounds
  4. Because DI security is only defined asymptotically
Show the answer

B. Because untrusted devices with memory need not act identically and independently across rounds, and the protocol has no permutation symmetry to exploit; EAT handles sequential, memory-dependent rounds directly

de Finetti-type reductions require permutation symmetry on a known, finite-dimensional system; black boxes have unknown dimension and memory. EAT needs only a sequential structure and a single-round entropy bound, and its first-order term matches the collective-attack rate.

The 2022 experiments, and the frontier

DIQKD needs a loophole-free Bell violation, with high fidelity, high detection efficiency and many rounds. Two experiments published together in 2022 crossed the line:

  • Trapped ions (Nadlinger et al., Oxford): two strontium ions in separate traps about two meters apart, entangled via photons, produced a finite-size, composably secure key of order 10510^5 bits over roughly eight hours, the first complete DIQKD protocol.
  • Neutral atoms (Zhang et al., Munich): two rubidium atoms in labs 400 m apart reached a CHSH violation and error rate sufficient for a positive key rate in the asymptotic limit.

A photonic experiment (Liu et al., 2022) demonstrated positive asymptotic rates with high-efficiency detectors over short fiber. In all cases rates are many orders of magnitude below trusted-device QKD.

RequirementBB84 (trusted devices)CHSH-based DIQKD
Tolerable QBER, depolarizing noise11.0% (one-way)≈ 7.1% (Acín et al. bound)
Detection efficiencyLosses only reduce the rate (with decoys)No-clicks must be kept as outcomes; even a CHSH violation needs η>2/(1+2)≈82.8%\eta>2/(1+\sqrt2)\approx82.8\% for a maximally entangled state (as low as 2/32/3 with partially entangled states, Eberhard); positive key needs more
What certifies securityDevice model + error ratesThe CHSH value alone

Key takeaways

  • DIQKD treats devices as black boxes and certifies secrecy from the Bell statistics, assuming only quantum theory, sealed labs, free settings and an authenticated channel.
  • Against collective attacks, r≥1−h(Q)−h(1+(S/2)2−12)r\ge1-h(Q)-h\big(\tfrac{1+\sqrt{(S/2)^2-1}}{2}\big) (Acín et al. 2007), proved via Jordan's lemma, Bell-diagonal symmetrization and Devetak–Winter.
  • Entropy accumulation extends the same first-order rate to general attacks with O(n)\order{\sqrt n} finite-size corrections.
  • DIQKD is far more demanding than BB84 (≈7.1% noise tolerance, high detection efficiency); the first complete demonstration (trapped ions) and an atom-based system 400 m apart appeared in 2022.

Check yourself

In the interactive lesson these are scored with your confidence, and what you miss comes back for review.

Question 1

Which of these is not an assumption of device-independent QKD?

  1. Alice's and Bob's labs do not leak their boxes' inputs or outputs
  2. The measurement settings are chosen with randomness independent of the devices
  3. The devices measure qubits in the intended bases
  4. The classical channel is authenticated
Show the answer

C. The devices measure qubits in the intended bases

Not modeling the devices is the whole point: security follows from the observed CHSH value, whatever the boxes contain. Sealed labs, free settings, authentication and quantum theory itself remain assumptions.

Question 2

A run gives S=2.00S=2.00 and Q=0Q=0. What does the device-independent bound certify?

  1. One secret bit per key round, since there are no errors
  2. Nothing: F(2)=h(12)=1F(2)=h(\tfrac12)=1, so r≥0r\ge0 and Eve may know the entire key
  3. About half a bit per round
  4. A key secure against individual but not collective attacks
Show the answer

B. Nothing: F(2)=h(12)=1F(2)=h(\tfrac12)=1, so r≥0r\ge0 and Eve may know the entire key

At S=2S=2 the statistics can be produced by a local deterministic strategy that Eve fixed in advance, so perfectly correlated outputs may be completely known to her. Error-free correlations alone certify nothing without a Bell violation.

Question 3

What role does Jordan's lemma play in the Acín et al. proof?

  1. It shows that the devices must contain exactly one qubit each
  2. It bounds the CHSH value by 222\sqrt2
  3. It decomposes any two binary observables into 2×22\times2 blocks, reducing arbitrary-dimensional devices to a mixture of qubit problems
  4. It converts collective attacks into coherent ones
Show the answer

C. It decomposes any two binary observables into 2×22\times2 blocks, reducing arbitrary-dimensional devices to a mixture of qubit problems

Two projections on any Hilbert space can be simultaneously block-diagonalized into blocks of size at most 2. Eve may be handed the block label, which reduces the analysis to qubits.

Question 4

Why can't a DIQKD protocol simply discard the rounds in which a detector fails to click, as BB84 does?

  1. Because Eve's boxes could choose when to "fail", keeping only rounds that fake a Bell violation (the detection loophole); no-clicks must be assigned outcomes
  2. Because discarded rounds reduce the key rate too much
  3. Because the CHSH value is defined only for photons
  4. Because no-clicks reveal the measurement setting to Eve
Show the answer

A. Because Eve's boxes could choose when to "fail", keeping only rounds that fake a Bell violation (the detection loophole); no-clicks must be assigned outcomes

With post-selection, a local model can reproduce S=22S=2\sqrt2 by refusing to answer when its preprogrammed outcome would lose. Assigning a fixed outcome to no-clicks keeps the test honest, which is why high detection efficiency is essential.

Further reading

  • A. Acín, N. Brunner, N. Gisin, S. Massar, S. Pironio & V. Scarani, "Device-independent security of quantum cryptography against collective attacks", Physical Review Letters 98, 230501 (2007).
  • U. Vazirani & T. Vidick, "Fully device-independent quantum key distribution", Physical Review Letters 113, 140501 (2014).
  • R. Arnon-Friedman, F. Dupuis, O. Fawzi, R. Renner & T. Vidick, "Practical device-independent quantum cryptography via entropy accumulation", Nature Communications 9, 459 (2018).
  • D. P. Nadlinger et al., "Experimental quantum key distribution certified by Bell's theorem", Nature 607, 682 (2022).
  • W. Zhang et al., "A device-independent quantum key distribution system for distant users", Nature 607, 687 (2022).