Scalable and Transparent Attribution for Human-AI Collaborative Code

As LLMs become embedded in software development, code increasingly emerges from human--AI collaboration rather than purely human effort. This shift raises key questions for deep learning applied to code: how can AI assistance be detected at scale, which structural signals distinguish human from machine authorship, and how do generation signatures manifest in real codebases? Existing detectors depend on opaque 100M$+$ parameter transformers, making them computationally expensive and difficult to interpret. We present NanoDroid, a backbone-free detector with only 5.1M parameters that runs $\mathbf{20\times}$ faster than the 125M-parameter UniXcoder (SOTA) while remaining fully interpretable. NanoDroid provides per-stream attribution scores that reveal how statistical and structural signals, token distributions, transition patterns, and surface structure drive authorship predictions, enabling transparent analysis of human versus AI coding signatures. On the challenging 4-class DroidCollection benchmark (Human / Machine / Refined / Adversarial), NanoDroid achieves 90.77% weighted F1, within 2.6% of SOTA, despite being over $\mathbf{21\times}$ smaller. Beyond efficient detection, we position NanoDroid as a scalable framework for studying code authorship through a deep learning lens, providing an interpretable window into how human and AI coding styles diverge, adapt, and increasingly blend in collaborative software development.