Fully Quantum Latent Diffusion for Synthetic Malware Records: Effective Denoising, Limited Class Information, and Costly Post-Selection

Abstract

Malware benchmarks are difficult to share, because the samples are harmful and a corpus can reveal the incidents of its holder, while synthetic substitutes can leak training records. Quantum diffusion models, however, have been evaluated only on images or quantum states, leaving the utility, privacy, and sampling cost of their samples unknown. In this research, we introduce a fully quantum latent diffusion model for malware records, in which a parameterized circuit with one mid-circuit measurement denoises amplitude-encoded latent vectors of opcode sequences without classical neural layers. On 17 malware families with ten seeds, we compared it with random states, a perfect generator, a parameter-matched classical control, and a classical latent diffusion model. The circuit lowers its denoising infidelity by 63\%, yet classifiers trained on its records reach a macro-F1 score of 0.11 on real records, against 0.08 for random states, 0.61 for a perfect generator, and 0.44 for a classical diffusion model of comparable size, and adding its records to real training data lowers the score from 0.635 to 0.593. Post-selected sampling accepts $3.0\times10^{-7}$ of runs, and a training penalty that raises the acceptance to 0.14 removes most of the denoising. No synthetic record copies a training record, and smoothing the decoded lengths lowers the area under the curve of a length-based membership test from 0.80 to 0.51, which a differentially private length histogram matches with a formal guarantee. These results indicate that quantum generators of shareable security data require a joint evaluation of utility, privacy, and sampling cost.