Abstract
This work introduces a lightweight framework for privacy-preserving neuralnetwork inference based on keyed chaotic masking a deterministic, user-specificobfuscation method derived from cryptographically seeded chaotic dynamicalsystems. The approach applies masks to input and output tensors usingkey-conditioned graph dynamics, enabling authenticated inference, userattribution, and soft output watermarking without modifying modelarchitectures. While the underlying chaotic system used to generate each maskis not analytically invertible, the masking operation itself is algebraicallyreversible by authorized key holders, offering functional privacy withoutformal cryptographic guarantees. Unlike traditional encryption or securemulti-party computation, this method operates in continuous space and imposesminimal computational overhead. We describe the construction of the maskingsystem, including graph sampling, dynamical rule selection, and chaosdiagnostics. Applications include privacy-preserving inference, secure datacontribution, and per-user watermarking in shared model pipelines. Thisframework offers a practical and modular building block for user-controlledprivacy in modern AI systems.