Ning Zhang

Research profile

Research summary

Ning Zhang is a researcher at the University of Kentucky whose reviewed work to date spans two tracks of machine learning: security and privacy centered on gradient inversion attacks, and multimodal machine learning focused on audio-visual understanding. The Evolve catalog holds two 2026 arXiv papers, one in each field — an even 0.5/0.5 split that drives the lowest focus score — and lists Zhang as a middle author on both, without corresponding status. The first paper reconstructs private trajectories in embodied reinforcement learning through temporal gradient inversion, reporting 18.8 dB PSNR and 100% action accuracy; Evolve judges it a credible and thorough study bounded by a narrow threat model, with reproducibility readiness rated low. The second, AVTrace, is a carefully built seven-task diagnostic benchmark for audio-visual temporal reasoning in omni models, covering event localization and synchronization; it is credited for candid limitation reporting, though silver model-assisted references and unequal inputs temper its conclusions, and its readiness is medium. Collaboration data records 11 distinct coauthors, including recurring partners such as Sudip Bhujel, Shanghao Shi, Ruiquan Huang, Yang Xiao, and Longyin Zhang. The reproducibility mean across the two evaluated papers is 38/100. Output, impact, and activity remain unscored because the underlying bibliometric data is not yet in this snapshot.

Research keywords

Highlights

• Gradient inversion recovers embodied RL trajectories

  In the 2026 temporal gradient inversion paper, Evolve reports credible recovery of private embodied-RL trajectories at 18.8 dB PSNR with 100% action accuracy, calling it a thorough study bounded by a narrow threat model; reproducibility readiness is low.

• AVTrace: a 7-task audio-visual temporal benchmark

  The 2026 AVTrace paper diagnoses temporal reasoning in omni models over seven tasks, spanning event localization and audio-visual synchronization; Evolve credits candid limitation reporting while noting silver model-assisted references and unequal inputs temper conclusions. Reproducibility readiness: medium.

• Dual-track profile: attack side and evaluation side

  Across the two reviewed papers, Ning Zhang contributes on both the privacy-attack side (gradient inversion) and the evaluation side (audio-visual benchmarking), each field at 0.5 of the observed distribution — an unusually even start for an early-stage catalog entry.

Timeline

1. 2026Temporal gradient inversion for private trajectory reconstruction in embodied reinforcement learning appears on arXiv; Evolve later rates its reproducibility readiness low.
2. 2026AVTrace, a seven-task diagnostic benchmark for audio-visual temporal reasoning in omni models, appears on arXiv, with reproducibility readiness rated medium.
3. 2026Both papers enter the Evolve catalog, splitting the observed field distribution evenly: 0.5 machine learning security and privacy, 0.5 multimodal machine learning.

Papers in the Evolve catalog

• 2026-09-24

  Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning

• 2026-09-17

  AVTrace: Diagnosing Audio-Visual Temporal Reasoning in Omni Models

Frequent co-authors