Inspiration
I’m a cybersecurity professional with a software engineering background (offensive security, SIEM and incident-response automation). I came to this hackathon with two questions: how will quantum technology change cryptography and key security, and how reliable are quantum algorithms on noisy, imperfect hardware? The three challenges cover both.
What I built
Three Qiskit notebooks, all running on simulators:
E91 quantum key distribution (security). Bell-pair preparation, random-basis measurement, key sifting, CHSH evaluation, and detection of an intercept-resend eavesdropper. Treating Eve as an attacker, I measured how her presence shows up in the statistics. Theory predicts S = 2√2 ≈ 2.83 without Eve and S = √2 ≈ 1.41 under full interception, which is below the classical bound of 2. The key error rate rises from 0 to 5/24 ≈ 21%. My runs gave S = [measured, no Eve] versus [measured, p = 1], matching the analytic curve S(p) = 2√2 − √2·p within statistical error.
Grover search under noise (attack and defense). Grover searches for a hidden key, compared against a query-matched classical baseline. I measured how depolarizing and bit-flip noise change the success probability and the optimal iteration count: [your finding, e.g. “at n = 4 the best r dropped from 3 to 2 at p = 0.01 and to 1 at p = 0.05”].
VQE for LiH (software engineering under constraints). A Jordan-Wigner qubit Hamiltonian built from precomputed integrals and tapered from 6 to 4 qubits, with a [your ansatz] ansatz and a noisy, shot-limited optimizer. Final energy error at [bond lengths]: [numbers, in mHa].
How I built it
Python, Qiskit 2.6.1 and Qiskit Aer, with NumPy/SciPy for optimization and Matplotlib for figures.
Challenges
Environment and verification. The biggest practical challenge was that I couldn’t run the Qiskit notebooks end to end while writing them. I checked the maths separately in plain NumPy instead: the E91 prediction formula against a brute-force density-matrix calculation, the Grover oracle and diffuser against a statevector mock, and the Jordan-Wigner Hamiltonian and tapering against a dense construction.
Getting the E91 numbers right. My first estimate of the key-pair correlations was wrong. Under full interception the two key pairs have correlations 1/2 and 2/3, giving a QBER of 5/24, not the value I first wrote. Rechecking the formula numerically caught it, and the written answers were corrected.
Eavesdropper vs. noise. Depolarizing noise scales every correlation equally, while interception changes them pair by pair, so the classifier fits all nine setting pairs. The hard limit is an Eve who picks her direction uniformly from the whole x-z plane. She is mathematically identical to depolarizing noise, and only a measurement outside that plane (such as y) separates them.
Grover’s cost per iteration. Without ancillas, the multi-controlled gate decomposes into many two-qubit gates, so each iteration is expensive. Under noise the best iteration count drops below the ideal one quickly, especially as n grows.
VQE under a hard budget. The limits were 40 two-qubit gates, 600 energy evaluations and 20M shots. An ansatz with too many gates is dominated by noise, and one with too few can’t represent the ground state. Shot noise also makes a gradient-based optimizer unreliable, which pushed me toward a noise-tolerant optimizer.
What I learned
A Bell test certifies entanglement, not security. Each extra interception rate shows up linearly in S, but a random-direction Eve looks identical to depolarizing noise unless you measure outside the x-z plane.
Grover’s quadratic speed-up erodes quickly on noisy hardware, because each iteration costs many two-qubit gates and the useful iteration count shrinks.
Notes
All code was written by me. AI tools were used only to understand concepts, in line with the hackathon rules.
